Email alert
Latest Articles
Articles in press have been peer-reviewed and accepted, which are not yet assigned to volumes/issues, but are citable by Digital Object Identifier (DOI).
Display Method:
Available online , doi: 10.11999/JEIT260410
Abstract:
Objective In emergency rescue scenarios, on-site unmanned aerial vehicles and other intelligent equipment often lack sufficient computing power to execute complex perception tasks such as disaster identification and personnel detection, while transmitting raw image data back to edge servers suffers from severe bandwidth limitations and channel degradation. Task-oriented semantic communication systems address this challenge by transmitting only task-relevant semantic features, thereby achieving high compression ratios and robust performance under poor channel conditions. However, existing methods require end-to-end training, which becomes computationally prohibitive when integrating large generative models like Stable Diffusion. Moreover, channel noise accumulated during the reverse diffusion process can cause semantic drift, leading to degraded task accuracy. To overcome these limitations, this paper proposes a channel-adaptive conditional diffusion generation method for task-oriented semantic communication, specifically targeting image classification under low-signal-to-noise-ratio (SNR) conditions. The objective is to achieve robust semantic feature transmission and accurate classification by leveraging a pre-trained diffusion model at the receiver while maintaining low training complexity and adapting dynamically to varying channel qualities. Methods The proposed framework comprises three key components: a task-channel joint sensing encoder at the transmitter, a conditional diffusion generator based on Stable Diffusion (SD) with ControlNet at the receiver, and an adaptive condition guidance strategy. At the transmitter, the encoder extracts task-relevant semantic features, taking the input image and the current SNR as inputs. The encoder adaptively adjusts its feature extraction according to the channel condition via an attention-based modulation module. The semantic features are transmitted over an additive white Gaussian noise channel, yielding noisy features. At the receiver, a pre-trained SD model serves as the core generator, and a trainable ControlNet takes the noisy features as input to produce residual corrections that guide the reverse diffusion process. After a fixed number of DDIM sampling steps, the generated features are decoded by a task decoder to output classification results.To avoid the high computational cost of end-to-end backpropagation through the multi-step diffusion chain, a two-stage training framework is proposed. In the first stage, under a noise-free environment, an information bottleneck is used to learn task-relevant features and train the task decoder. The loss function combines cross-entropy classification loss and KL divergence for variational information bottleneck. In the second stage, the SD backbone is frozen, and only the task encoder and ControlNet are trained to adapt to channel noise. The training loss includes a standard diffusion denoising loss and a distillation loss that aligns the transmitted features with the first-stage features. To mitigate semantic drift caused by noise accumulation during reverse diffusion, an adaptive condition guidance strategy is introduced. A learnable weight network takes the current timestep and SNR as inputs and outputs a scaling factor between zero and two. When the scaling factor equals one, the conditional residual is injected as is; when it is less than one, the conditional guidance is weakened; when greater than one, the conditional guidance is enhanced. Results and Discussions Experiments are conducted on the ImageNet-100 dataset, comprising 130,000 training images and 5,000 validation images across 100 classes. All images are resized to 256×256 pixels and compressed to 4×32×32 latent features (48× compression). The Stable Diffusion v1.5 backbone is frozen during the second stage, with only the task encoder and ControlNet trained for 15 epochs at a learning rate of 1e-4. The adaptive weight network is trained during the DDIM sampling process for 5 epochs with a batch size of 96 and a learning rate of 1e-4, using 20 sampling steps.The proposed method is compared with two baselines: T-DeepSC, a single-task semantic communication system without diffusion, and DDSC+VGG16, a diffusion-driven semantic communication method that maps channel noise to forward diffusion steps and cascades VGG16 for classification. Under SNR values from -3 dB to 7 dB, the proposed method achieves the highest classification accuracy across all conditions. Specifically, at -3 dB, the proposed method attains 44.0% accuracy, outperforming T-DeepSC at 39.8% by 4.2 percentage points. At -1 dB, the proposed method reaches 53.2%, surpassing T-DeepSC at 40.0% by 13.2 percentage points. At 7 dB, the proposed method achieves 66.8%, exceeding DDSC+VGG16 at 58.4% by 8.4 percentage points. These results demonstrate that the proposed two-stage training framework with adaptive conditional guidance effectively recovers task-relevant semantic features from noisy transmissions, delivering superior performance even under severe channel degradation.Ablation studies further validate the contribution of each component. Removing the adaptive weight control, denoted as configuration B, reduces accuracy by 3.5 percentage points at -3 dB and 3.6 percentage points at -1 dB compared to the full method, configuration C, which achieves 44.0% and 53.2% respectively. Compared to the DDSC+VGG16 baseline, configuration A, introducing the task encoder and ControlNet, configuration B, yields dramatic improvements, for example, from 58.4% to 65.0% at 7 dB. This confirms that the two-stage framework successfully learns task-relevant semantic features and adapts to channel noise without end-to-end backpropagation through the diffusion chain. Conclusions This paper presents a channel-adaptive conditional diffusion generation method for task-oriented semantic communication, targeting image classification under bandwidth and resource constraints. The proposed framework integrates a task-channel joint sensing encoder, a ControlNet-guided Stable Diffusion generator, and a two-stage training strategy that decouples task feature learning from channel noise adaptation, significantly reducing training complexity compared to end-to-end diffusion-based alternatives. The adaptive condition guidance strategy dynamically adjusts the injection strength of noisy conditional information during the reverse diffusion process, suppressing semantic drift and improving classification robustness. Experimental results on ImageNet-100 demonstrate that the proposed method outperforms both task-oriented semantic communication without diffusion (T-DeepSC) and diffusion-driven reconstruction-then-classification (DDSC+VGG16) across most SNR conditions. In addition, the framework can also generate reconstructed images at the receiver, which may serve as visual auxiliary evidence for human verification in emergency rescue scenarios. Future work will extend the framework to multi-modal semantic fusion, model lightweighting for edge deployment, and more complex tasks such as object detection and visual question answering, facilitating practical adoption in emergency rescue and other mission-critical scenarios.
Available online , doi: 10.11999/JEIT260260
Abstract:
Objective Federated Split Learning (FSL) has emerged as a promising paradigm for Edge Intelligence (EI) scenarios. By partitioning neural networks, FSL enables collaborative training while alleviating resource constraints on terminal devices. However, frequent interactions between devices and edge servers create high communication overhead, and existing FSL methods have limited adaptability to dynamic network environments. These challenges restrict deployment efficiency. In addition, data heterogeneity in practical edge environments can degrade personalized model performance, while existing studies have not jointly optimized personalized test accuracy and communication efficiency. This study addresses these challenges by integrating Reinforcement Learning (RL), multi-strategy communication optimization, and multi-task learning to adapt to network bandwidth variations, reduce communication overhead, and improve personalized test accuracy. Methods This study proposes a communication-efficient adaptive FSL framework that integrates RL-based adaptive model splitting, multi-strategy communication optimization, and multi-task learning-based personalization. First, the Reinforcement Learning-based Communication-Efficient Adaptive Federated Split Learning (RLCEA-FSL) algorithm is proposed to adjust model partition points adaptively in response to network bandwidth variations and reduce communication overhead between end devices and edge servers through multiple communication optimization strategies. Second, the Multi-Task Learning-based Communication-Efficient Personalized Federated Learning (MTLCE-PFL) algorithm is proposed to address the joint optimization of personalized test accuracy and communication efficiency under Non-Independent and Identically Distributed (Non-IID) data distributions. Specifically, batch normalization layers serve as private personalized layers to adapt models to local data distributions, thereby formulating the Non-IID setting as a multi-task learning problem. Additionally, quantization-based compression is applied to intermediate activations transmitted from devices to servers, while dynamic hierarchical gradient compression based on Compressed Sensing (CS) is used to reduce the communication overhead of gradient transmission from servers to devices. Results and Discussions The proposed framework and algorithms are evaluated on the CIFAR-10 and CIFAR-100 datasets using VGG-11 and ResNet-9 models. Four baseline methods are considered: classical Federated Learning (FL), SplitFed, FedAdapt, and EcoFed. With activation reuse enabled, RLCEA-FSL (Algorithm 1) achieves speedups of up to 101.23×, 24.27×, and 117.61× compared with FL, SplitFed, and FedAdapt, respectively. Without activation reuse, it reduces average per-round training time by 37.91% compared with EcoFed (Fig. 3 ). Under bandwidth-constrained conditions, RLCEA-FSL reduces overall communication overhead by 72.75×, 17.33×, and 66.62× compared with FL, SplitFed, and FedAdapt, respectively, and by 8.1% compared with EcoFed (Fig. 6 ). These improvements result from the adaptive model-splitting strategy, which adjusts partition points in response to bandwidth variations, combined with multiple communication optimization strategies. The RL agent trained on VGG-11 can also be applied directly to ResNet-9 without retraining, yielding favorable results in training latency, test accuracy, and communication overhead (Figs. 7 and 8 ). MTLCE-PFL improves personalized test accuracy under Non-IID data distributions while reducing communication overhead and training latency. Compared with the other algorithms, MTLCE-PFL (Algorithm 2) achieves faster convergence and higher personalized test accuracy on both CIFAR-10 and CIFAR-100 with both VGG-11 and ResNet-9 (Fig. 9 ). These improvements result from the multi-task learning framework and private personalized layers, which adapt to local data distributions and enable client-specific optimization. In terms of communication efficiency, MTLCE-PFL reduces bidirectional communication overhead by 11.97×, 2.86×, and 10.33× compared with FL, SplitFed, and FedAdapt, respectively, through quantization-based compression of intermediate activations and dynamic hierarchical CS-based compression of gradients transmitted back to devices (Fig. 10 ). Despite the additional computational complexity introduced by personalized layers, MTLCE-PFL reduces average per-round training time by 11.02% compared with SplitFed and achieves speedups of 4.28× and 18.68× over FedAdapt and FL, respectively (Fig. 11 ). These results demonstrate the effectiveness of the proposed methods in maintaining global model accuracy while improving personalized test accuracy and reducing communication overhead and training latency. Conclusions The proposed communication-efficient adaptive FSL framework addresses high communication overhead, dynamic network conditions, and data heterogeneity in EI scenarios through complementary optimization strategies. RLCEA-FSL combines RL-based adaptive model splitting with communication optimization to respond to bandwidth variations and reduce training latency and communication overhead while maintaining global model accuracy comparable to that of the baseline methods. The direct application of the RL agent trained on VGG-11 to ResNet-9 without retraining suggests potential deployment flexibility; however, further experiments are needed to establish robust cross-model generalization. MTLCE-PFL combines multi-task learning with private personalized layers to improve personalized test accuracy under Non-IID data distributions while reducing bidirectional communication overhead through quantization-based activation compression and dynamic hierarchical CS-based gradient compression.
Available online , doi: 10.11999/JEIT260470
Abstract:
Objective Frequency-Modulated Continuous-Wave Multiple-Input Multiple-Output (FMCW-MIMO) radar provides a non-contact approach to full-field displacement measurement by combining sparse imaging with inter-frame phase-difference estimation. Compared with conventional point-based measurement methods, radar imaging enables displacement information to be obtained simultaneously from multiple spatial targets. However, the accuracy of phase-based displacement estimation depends strongly on the consistency of reconstructed target locations across consecutive frames. Existing sparse imaging methods may produce fluctuations in the reconstructed support set because of sidelobe leakage, target omission, and mainlobe drift, particularly under low Signal-to-Noise Ratio (SNR) conditions. Such support-set instability introduces phase errors and causes discontinuities or fluctuations in the estimated displacement. Therefore, a sparse imaging method that maintains a stable support set while suppressing sidelobe interference is developed for high-precision MIMO radar displacement measurement. Methods A UNet-enhanced Alternating Direction Method of Multipliers (ADMM) algorithm, termed UNetE-ADMM, is proposed by embedding a UNet gain module into the ADMM proximal operator. Spatially adaptive normalized weights are generated from imaging features to distinguish target mainlobes from sidelobes. The mainlobes are preserved while the sidelobes are progressively suppressed, stabilizing the support set during iterative reconstruction. The reconstructed complex scattering coefficients are then used for inter-frame phase-difference estimation and radial displacement measurement. A support-set stability criterion and a displacement-error model are also established to analyze the effects of mainlobe drift, grid mismatch, and target motion during the intervals between transmitter transmissions. Results and Discussions Simulations are conducted using a cascaded AWR2243 MIMO radar configuration with a 77-GHz carrier frequency, a 3.65-GHz bandwidth, and a 50-Hz frame rate (Table 2 ). For distributed point targets at SNRs ranging from −10 to 10 dB, the proposed method produces more stable target locations and lower sidelobe levels than the L1-regularized ADMM (L1-ADMM) and Plug-and-Play UNet (PnP-UNet) methods (Fig. 9 ). At −10 dB, the mainlobe drift probability decreases from 37.5% with L1-ADMM to 6.3%, corresponding to an 83.2% reduction, while the displacement Root Mean Square Error (RMSE) decreases from 44.62 to 1.54 μm (Table 3 ). In the vibration experiment with two corner reflectors, the proposed method achieves displacement RMSEs of 185 and 18 μm for the moving and stationary targets, respectively, compared with 2,247 and 122 μm, respectively, for L1-ADMM (Table 4 ). In the human vital-sign experiment, the reconstructed chest-displacement curve is smoother (Fig. 16 ), and the heartbeat spectral peak is more clearly resolved (Fig. 17 ), while the measured heart rates remain consistent with the reference measurements (Table 5 ). The convergence curves and target-background ratio results further demonstrate effective sidelobe suppression and stable iterative reconstruction (Fig. 18 ). Conclusions The proposed UNetE-ADMM method integrates learned target-structure priors into ADMM-based sparse imaging and improves support-set stability through adaptive mainlobe preservation and sidelobe suppression. Experimental results demonstrate a lower mainlobe drift probability and improved displacement accuracy compared with L1-ADMM and PnP-UNet, particularly under low-SNR conditions. The method is also applicable to both discrete targets and continuous structures. Grid mismatch, the computational cost of dense imaging grids, and generalization to more complex scenes require further investigation.
Available online , doi: 10.11999/JEIT260729
Abstract:
Objective Semantic segmentation assigns a semantic category to each image pixel and supports scene understanding in autonomous driving and robot navigation. Methods based on visible-light images can lose accuracy under low illumination, strong glare, and occlusion. Thermal infrared and depth images provide complementary information, but extracting features with two full backbone networks increases model size and computational cost, limiting deployment on embedded devices with restricted computing resources. Parameter-efficient adaptation reduces the number of weights updated during training, but a large pretrained backbone can still impose substantial inference overhead. Model compression and runtime optimization are therefore also needed. This study develops a multi-modal segmentation framework that combines dynamic visual prompts, structured knowledge distillation, and deployment optimization. The framework aims to improve the use of auxiliary-modality information with a limited number of trainable parameters, transfer the resulting representations to a smaller student network, and increase inference throughput across heterogeneous embedded platforms. Methods A Multi-Modal Feature Fusion Prompt Network (MFFPNet) is constructed using an ImageNet-pretrained Mix Transformer B5 (MiT-B5) backbone and a SegFormer decoder. Most pretrained backbone parameters remain frozen, while the prompt, fusion, normalization, and decoding components are adapted to the segmentation task. The encoder retains four hierarchical stages. A lightweight Feature Fusion Prompt (FFP) module encodes thermal infrared or depth information into dynamic visual prompts, which are added to visible-light features. The prompt generated at each stage is passed to the next stage after spatial and channel adjustment, allowing auxiliary information to contribute at multiple feature scales. Global learnable prompts provide an additional mechanism for adapting pretrained representations to the task (Fig. 1 ). Within the FFP module, spatial and channel attention assign different weights to the two modalities. Complementary feature reweighting is followed by cross-attention to capture cross-modal relationships and generate the visual prompt. The decoder aggregates outputs from all four stages to produce pixelwise predictions. Training combines label-smoothed cross-entropy loss and Dice loss to account for pixel-level classification and regional overlap. The backbone is initialized with pretrained weights, and training uses stochastic gradient descent with momentum and polynomial learning-rate decay. Data augmentation includes multi-scale resizing, random cropping, horizontal flipping, and photometric perturbations. For model compression, the trained MFFPNet serves as a frozen teacher in an offline distillation framework. MFFPNet-S uses MiT-B1 as its main student backbone, and MiT-B2 is also evaluated on NYUDv2 (Fig. 4 ). A feature adapter applies convolution, batch normalization, and a rectified linear unit to reduce differences between student and teacher features (Fig. 5 ). Distillation combines feature content loss, channel relation loss, spatial structure loss, and logit supervision. A confidence mask emphasizes teacher predictions in high-confidence regions, while a global supervision term and ground-truth labels preserve training information outside the selected regions. Combining local and global supervision preserves structural information while reducing the influence of low-confidence local predictions (Fig. 6 ). For deployment, Open Neural Network Exchange (ONNX) provides an intermediate model representation for platform-specific compilation. TensorRT uses half-precision inference on NVIDIA Jetson Orin NX, the Compute Architecture for Neural Networks (CANN) software stack supports Huawei Atlas 200I DK A2, and the Rockchip Neural Network (RKNN) toolchain supports RK3588. Mixed-precision inference on RK3588 retains higher precision for sensitive attention and output layers. Containerized environments and a common inference interface support deployment, while a three-stage asynchronous pipeline overlaps data loading, inference, and network communication. Results and Discussions The evaluation uses the MFNet dataset, which pairs Red-Green-Blue (RGB) images with thermal images, and the NYUDv2 dataset, which pairs RGB images with depth images. On MFNet, MFFPNet achieves a mean Intersection over Union (mIoU) of 63.3% with 9.4 million trainable parameters, approximately 10.4% of its 90.4 million total parameters. Its mIoU exceeds the reported values for CRM-B, DPLNet, and MixPrompt by 1.9, 5.0, and 3.2 percentage points, respectively. MixPrompt uses fewer trainable parameters, indicating a trade-off between segmentation accuracy and adaptation size (Table 2 ). Visual comparisons show improved pedestrian contours under challenging nighttime illumination (Fig. 2 ). On NYUDv2, MFFPNet achieves an mIoU of 60.6% under single-scale testing, compared with 58.3% for DPLNet under the same testing protocol. The reported multi-scale results for DPLNet and MixPrompt use a different evaluation protocol and therefore cannot be directly compared with the single-scale results (Table 3 ). Indoor examples show finer object boundaries and better separation of visually similar regions (Fig. 3 ). Ablation results show that the global prompt and full fusion module improve segmentation performance (Table 4 ). Removing cross-attention reduces the MFNet mIoU from 63.3% to 62.1% (Table 5 ). Applying the fusion module at all four stages yields higher mIoU than applying it only at the first or last stage, supporting the use of features at multiple resolutions (Table 6 ). On MFNet, the MiT-B1 student achieves an mIoU of 56.2% with approximately 18.1 million parameters and 19.6 Giga Floating-Point Operations (GFLOPs), compared with 63.3% mIoU, 90.4 million parameters, and 122.8 GFLOPs for the teacher. The student therefore retains approximately 88.8% of the teacher’s mIoU while using 20.0% of its parameter count and reducing computation by approximately 84%. Its measured inference speed is 23.9 Frames Per Second (FPS), compared with 10.0 FPS for the teacher (Table 8 ). On NYUDv2, distillation increases the MiT-B1 student’s mIoU from 38.6% to 42.9%, a gain of 4.3 percentage points. The MiT-B2 student reaches 48.5%, providing higher accuracy with a larger model (Table 9 ). The MFNet distillation ablation also shows successive performance improvements as additional supervision components are introduced (Table 10 ). In synchronous deployment tests, the student achieves mIoUs of 55.7%, 51.5%, and 49.4% at 48.6, 31.5, and 28.6 FPS on Jetson Orin NX, Atlas 200I DK A2, and RK3588, respectively; the RK3588 result is obtained with mixed-precision inference. The 8-bit Integer (INT8) configuration on RK3588 reaches 31.2 FPS but reduces mIoU to 41.1%, illustrating the accuracy cost of more aggressive quantization (Table 12 ). With a continuous 1 000-frame input, asynchronous execution increases throughput to 68.5, 41.2, and 34.8 FPS on the three platforms, respectively, corresponding to gains of 40.9%, 30.8%, and 21.7% over their serial configurations (Table 13 ). Conclusions The results support dynamic visual prompting as a parameter-efficient approach to semantic segmentation with RGB-Thermal and RGB-Depth inputs. Structured distillation reduces the inference cost of the large teacher while retaining 88.8% of its mIoU on MFNet. Results on NYUDv2 also show a trade-off between model capacity and segmentation accuracy. Platform-specific compilation and asynchronous scheduling further improve throughput, whereas quantization can substantially reduce segmentation accuracy. The framework thus addresses training adaptation, model compression, and runtime execution as complementary aspects of embedded deployment. Future work will investigate modality completion using generative models or cross-modal distillation to improve segmentation when auxiliary sensor information is unavailable.
Available online , doi: 10.11999/JEIT260277
Abstract:
Objective With the continuous expansion of the Internet of Vehicles (IoV), its applications have become increasingly widespread. Users commonly outsource encrypted data to cloud servers, and Dynamic Searchable Symmetric Encryption (DSSE) provides an effective approach for dynamically updating and searching encrypted data. However, existing DSSE schemes still face several challenges in practical applications. On the one hand, DSSE cannot simultaneously support multi-user scenarios and verifiable search results. Multi-user functionality improves collaborative efficiency among groups, whereas search result verification ensures the correctness and reliability of query results. On the other hand, DSSE needs to achieve forward and backward privacy to protect sensitive data. Forward privacy ensures that newly inserted vehicle data cannot be associated with historical search tokens. Backward privacy ensures that deleted vehicle data are not exposed in subsequent queries and that cloud servers cannot recover deleted data using current or future query information. Furthermore, the risks of collusion attacks between malicious users and cloud servers, as well as the robustness of the scheme, should be considered during design. Methods A state chain mechanism based on random numbers is designed to achieve forward privacy by ensuring that newly added data cannot be associated with historical search tokens. Meanwhile, a bitmap index-based construction method is adopted to achieve backward privacy and prevent information leakage from deleted data during subsequent queries. Furthermore, a dual bitmap index structure consisting of a keyword bitmap index and an operation bitmap index is constructed and combined with Boolean circuit technology to design a DSSE framework that simultaneously satisfies forward privacy, backward privacy, and robustness. The framework maintains stable system operation when facing abnormal or unreasonable update requests. In addition, an IDentity-based Key Encapsulation Mechanism (ID-KEM) is adopted to support multi-user scenarios where corrupted users and untrusted cloud servers may collude. This mechanism enables data owners to avoid storing additional information for authorized users, thereby reducing the storage overhead at the data owner’s side. Moreover, the cloud server is required to provide a proof of work based on verification labels when returning search results, enabling users to verify the integrity and correctness of the results. Results and Discussions The security of the RDSSE-CR scheme is demonstrated through a series of games between the real game \begin{document}$ \text{REAL}_{\mathcal{A}}^{\varPi }({1}^{\lambda }) $\end{document} \begin{document}$ \mathrm{IDEAL}_{\mathcal{A}}^{\varPi }({1}^{\lambda }) $\end{document} \begin{document}$ {G}_{0} $\end{document} \begin{document}$ {G}_{7} $\end{document} \begin{document}$ \mathcal{L} $\end{document} Fig. 4 ). Regarding algorithm performance (Fig. 5 ), the RDSSE-CR scheme achieves high verification efficiency through efficient XOR operations while maintaining a constant search token size. Although the introduction of security mechanisms slightly increases the update and search costs, the scheme provides improved functionality and security. In the Internet of Vehicles simulation based on Simulation of Urban MObility (SUMO) and OMNeT++, the proposed scheme effectively resists attack interference and enables vehicle queues to maintain stable spacing and consistent speed under attacks (Fig. 6 ). This prevents vehicle lag and queue collapse observed without the proposed scheme, demonstrating its practical value in improving the safety and stability of the Internet of Vehicles in complex communication environments. Conclusions A verifiable multi-user dynamic searchable symmetric encryption scheme supporting forward and backward privacy is designed to meet the practical requirements of the Internet of Vehicles. To address security threats in multi-user environments where untrusted cloud servers and corrupted users may collude, the proposed scheme implements multi-user authorization based on ID-KEM and effectively resists collusion attacks. The RDSSE-CR scheme ensures system robustness against unreasonable operation requests by constructing a dual bitmap index structure combined with Boolean circuit design. During the search process, the cloud server generates a proof of work based on verification labels, allowing users to verify the correctness and integrity of search results. Future research will focus on efficient user revocation mechanisms to further improve the practicality and security of the scheme.
Available online , doi: 10.11999/JEIT260490
Abstract:
Objective Unmanned Aerial Vehicles (UAVs) provide flexible and rapid image acquisition for disaster assessment and emergency response. Reliable transmission of the acquired images is essential for interpreting damaged buildings, blocked roads, and other critical scene elements. However, disaster-area communication links are often constrained by limited bandwidth, noise, time-varying fading, and imperfect Channel State Information (CSI). Low-contrast boundaries, narrow cracks, and scattered damaged regions are particularly vulnerable to feature compression and channel distortion. Deep Joint Source-Channel Coding (DeepJSCC) enables image compression and transmission to be jointly optimized, but several limitations remain for disaster imagery. Global pixelwise objectives can assign insufficient importance to sparse weak structures, while feature downsampling can attenuate information that later channel adaptation cannot recover. Channel adaptation also needs to account for both channel conditions and the relative importance of image content. Furthermore, reconstruction must balance pixel fidelity, structural consistency, perceptual quality, and computational cost. This study therefore aims to develop a unified transmission framework that strengthens weak-structure representation, coordinates semantic features with channel conditions, and supports efficient reconstruction of UAV disaster images. Methods A Semantic-Driven Adaptive Joint Source-Channel Coding method, termed SDAJSCC, is proposed. Semantic importance is modeled implicitly in the feature domain through multiscale responses, curvature saliency, and channel-spatial attention. The encoder combines a Patch-Aware Residual Block (PARB) with a Semantic Differential Fusion Block (SDFB). PARB aggregates local and regional information at multiple patch scales, while SDFB combines local-global feature interaction with Adaptive Curvature-Guided Attention (ACGA). Curvature information supplements content similarity and relative positional information to strengthen responses to boundaries and structural discontinuities. A Semantic Adaptive Modulation (SAM) module is introduced at multiple encoding stages and at the final latent projection. Channel-dependent affine transformations and channel-spatial attention adjust feature responses using CSI and the Signal-to-Noise Ratio (SNR). This modulation operates before power normalization, allowing the relative responses of latent features to be adjusted while maintaining the average transmit-power constraint. The real-valued latent representation is mapped to complex channel symbols for transmission. At the receiver, regularized zero-forcing equalization limits noise amplification associated with small estimated channel gains. The remaining distortion is analyzed as a combination of signal-dependent mismatch and transformed noise. A ChannelRefiner then combines the equalized latent representation with channel information to predict a gated residual correction and a feature bias. The corrected features are reconstructed by two independently parameterized decoder branches. The fidelity branch emphasizes global structure and pixel consistency, whereas the perceptual branch incorporates a Perceptual Texture Transform Adaptation (PTTA) module. PTTA applies ternary quantization to query and key representations and 6-bit quantization to the value representation, together with a local texture bias. An AdaptiveFusion module predicts the relative contributions of the two reconstructions from their aggregated features and the channel state. All components are jointly trained using a primary reconstruction loss, auxiliary pixelwise supervision for both branches, and a Learned Perceptual Image Patch Similarity (LPIPS) loss. Experiments use xView2 and RescueNet image crops of 256 × 256 pixels. Training covers Additive White Gaussian Noise (AWGN) and Rayleigh fading channels, with SNR uniformly sampled from -5 to 20 dB. AdamW optimization and cosine learning-rate decay are used for 100,000 iterations with a batch size of 16. Evaluation includes Peak Signal-to-Noise Ratio (PSNR), Multi-Scale Structural Similarity (MS-SSIM), LPIPS, critical-region quality, edge preservation, feature similarity, and inference efficiency. Results and Discussions The bandwidth comparison on xView2 under AWGN at 20 dB shows improved structural and perceptual preservation under severe compression. At channel bandwidth ratios of 1/192 and 1/96, SDAJSCC has PSNR values 2.15 and 2.08 dB below DeepJSCC, respectively, but increases MS-SSIM by 0.0599 and 0.0196 and reduces LPIPS by 29.6% and 27.8%. At a bandwidth ratio of 1/24, it improves PSNR by 6.31 dB and MS-SSIM by 0.1749 over JPEG2000+Turbo, with a 48.6% reduction in LPIPS. These results indicate a trade-off between pixel fidelity and structural-perceptual preservation at very low bandwidth, followed by gains across all three metrics as bandwidth increases (Fig. 7 ). At a fixed bandwidth ratio of 1/24 and an SNR of 20 dB, SDAJSCC improves PSNR and MS-SSIM over JSCCFormer-f by 1.44 dB and 0.0149 under AWGN, with a 29.1% reduction in LPIPS. Under Rayleigh fading, the corresponding improvements are 0.708 dB, 0.0156 , and an 11.6% reduction in LPIPS. Diffusion-based methods obtain lower LPIPS in some cases, but exhibit lower PSNR and MS-SSIM, demonstrating the need to assess fidelity and perceptual quality together (Fig. 8 ). Ablation experiments on RescueNet at a bandwidth ratio of 1/24 support the complementary roles of the proposed components. Under AWGN at 20 dB, removing PARB reduces PSNR by 0.300 dB, while removing SDFB increases LPIPS by 0.0138 , indicating their respective contributions to reconstruction fidelity and perceptual preservation (Fig. 9 ). Under Rayleigh fading at 20 dB, removing SAM or ChannelRefiner reduces PSNR by 0.652 or 0.332 dB, respectively, supporting the combination of transmitter-side adaptation and receiver-side correction (Fig. 10 ). Under AWGN at 20 dB, the complete dual-branch decoder reduces LPIPS by 5.5% relative to the fidelity-only branch, with a PSNR decrease of 0.161 dB, demonstrating an effective fidelity-perception compromise (Fig. 11 ). For imperfect CSI, the forward-link SNR is fixed at 10 dB under Rayleigh fading, and the CSI estimation SNR is varied independently. At a CSI estimation SNR of 0 dB, SDAJSCC achieves a PSNR of 24.34 dB, an MS-SSIM of 0.817, and an LPIPS of 0.357. Relative to the baseline, PSNR increases by 3.62 dB and LPIPS decreases by 28.2%. Across the tested CSI estimation conditions, ChannelRefiner provides PSNR gains of 0.52-0.57 dB within the complete framework (Table 2 ). Critical-region evaluation on RescueNet under Rayleigh fading at 10 dB and a bandwidth ratio of 1/24 uses damaged-building and blocked-road annotations to define the evaluation regions. Compared with the Region-of-Interest-Guided method (ROI-JSCC), SDAJSCC improves critical-region PSNR by 0.73 dB and reduces critical-region LPIPS by 14.3%, while increasing edge F1 and high-level feature cosine similarity by 0.035 and 0.032, respectively (Table 3 ). Qualitative comparisons under AWGN at 20 dB also show clearer road structures, building contours, and local textures in the complete model (Fig. 12 ). The model contains 29.34 million parameters and requires 18.94 billion floating-point operations per image. With 256 × 256 inputs and a batch size of one on an NVIDIA RTX 4090 Graphics Processing Unit (GPU), its mean inference time is 0.029 s per image, equivalent to approximately 34.5 frames/s. The measurements exclude data loading and preprocessing. Compared with the tested diffusion-based methods using 50 reverse-denoising steps, inference latency decreases by 97.4%-98.5%, although JSCCFormer-f remains faster and less computationally demanding (Table 4 ). Conclusions SDAJSCC integrates multiscale semantic-structural representation, channel-semantic modulation, post-equalization latent correction, and adaptive dual-branch reconstruction within a jointly optimized framework. The experiments support improved preservation of disaster-relevant weak structures and robustness under the tested fading and imperfect-CSI conditions, together with a favorable balance between reconstruction quality and inference efficiency. The low-bandwidth results also identify the remaining trade-off between pixel fidelity and perceptual preservation. The reported processing speed is measured on a desktop GPU, while CSI robustness is evaluated using independent complex Gaussian estimation errors. Future work will investigate model compression, online adaptation to dynamic channels, and generalization to correlated and non-Gaussian CSI errors.
Available online , doi: 10.11999/JEIT260926
Abstract:
In recent years, hybrid convolutional neural network (CNN) and spiking neural network (SNN) accelerators have shown great potential for edge intelligence. However, existing designs typically employ fixed time windows (TW) for temporal integration, limiting their adaptability to tasks with different temporal complexities. This work presents a low-power hybrid CNN/SNN accelerator with configurable TWs. First, a configurable TW instruction set is developed based on the spatiotemporal complexity of different tasks to reduce inference latency and computational overhead. Second, a zero-skipping data-reuse processing array with unified CNN/SNN operator mapping is designed to improve computational resource utilization. Finally, an event-driven execution scheme and fine-grained clock-gating strategy are introduced to eliminate redundant operations and memory accesses during computation and data transfer. The proposed CNN/SNN accelerator is fabricated in a 28-nm TSMC CMOS process, occupying a core area of 0.75 mm2. At 0.81 V and 100 MHz, it achieves classification accuracies ranging from 75.3% to 98.4% across MIT-BIH, N-TIDIGITS, N-MNIST, IMAGE-100, and UniMiB SHAR, with a minimum energy consumption of 147.78 nJ per inference. Compared with prior works, the configurable-TW mechanism reduces energy consumption by 1.24×–7.6× while maintaining competitive classification accuracy, significantly improving the overall energy efficiency. Objective The goal of this work is to design an efficient and low-power CNN/SNN hybrid neuromorphic SoC based on the RISC-V architecture, addressing the limitations of conventional SNN hardware with fixed TW under different task requirements. By introducing a configurable TW mechanism, the temporal integration length can be flexibly adjusted according to task complexity, thereby reducing redundant time-step computations and lowering inference latency and energy consumption. Methods To address the varying time-window requirements of different tasks, this work presents a RISC-V-based SCNN processor with a configurable TW. First, a custom instruction set is developed to configure the TW threshold of the underlying hardware according to the complexity of input samples, enabling demand-driven computation while reducing inference latency and computational overhead. Second, a zero-skipping data-reuse processing array is designed to improve computational resource utilization. Finally, an event-driven execution mechanism and fine-grained clock-gating strategy are introduced to eliminate redundant operations and unnecessary memory accesses during computation and data transfer. The chip is fabricated in a 28-nm TSMC CMOS process with a core area of 0.75 mm2. Across MIT-BIH, N-TIDIGITS, N-MNIST, IMAGE-100, and UniMiB SHAR tasks, the accelerator achieves classification accuracies ranging from 75.3% to 98.4%, with a minimum energy consumption of 147.78 nJ per inference. Results and Discussions Experimental results validate the effectiveness of the proposed configurable-TW SCNN processor under diverse workloads. Cross-task analysis (Fig. 7 ) shows that accuracy generally improves with increasing TW, but sensitivity varies significantly across tasks. Event-driven datasets such as N-MNIST and UniMiB SHAR exhibit strong temporal redundancy, achieving near-saturated accuracy (97.8%–98.4%) at small TW (≥3), whereas temporally dependent tasks like N-TIDIGITS require larger TW to maintain performance, indicating that TW demand is inherently data-modality dependent. Single-task evaluation (Fig. 8 and Fig. 9 ) further demonstrates that reducing TW can significantly lower spike rates and computational cost with negligible accuracy degradation. The latency–accuracy trade-off (Fig. 10 ) reveals that blindly increasing TW leads to exponential growth in addition operations and inference latency without proportional accuracy gains. The optimal TW varies across tasks, highlighting the necessity of dynamic TW adaptation. Conclusions This work presents a hybrid hardware accelerator with a configurable TW. By dynamically adjusting the TW, the proposed design addresses the performance and energy-efficiency limitations of fixed temporal resolution in conventional SNN hardware. Implemented in a TSMC 28-nm process, the accelerator occupies a core area of only 0.75 mm2 and supports multimodal tasks, including physiological-signal, speech, visual, and activity recognition. It achieves classification accuracies of 75.3%–98.4% and a minimum energy consumption of 147.78 nJ per inference. Compared with fixed-TW designs, the proposed accelerator offers greater task adaptability, computational efficiency, and energy efficiency, reducing energy consumption by 1.24×–7.6× under the same process technology.
Available online , doi: 10.11999/JEIT261012
Abstract:
Objective The large-scale deployment of Internet of Things (IoT) devices raises the demand for lightweight authentication and dynamic key generation on resource-limited hardware. Physical Unclonable Functions (PUFs) leverage inherent manufacturing variations of integrated circuits to generate unique, unclonable hardware fingerprints, serving as an effective lightweight scheme for device authentication and secure key derivation. Nevertheless, conventional delay-based strong PUFs feature linear delay superposition, rendering their challenge–response relationships susceptible to machine learning modeling. Existing countermeasures against modeling attacks enhance nonlinearity using elaborate feedback loops, multi-stage XOR gates and heterogeneous logic cells. Though these approaches improve security to some extent, they incur substantial hardware overhead, timing convergence problems, degraded reliability and weak environmental robustness. This work presents a Dynamic Spatial Coupling Strong PUF (DSC-PUF) architecture. With restrained hardware cost, the proposed design maintains PUFs’ inherent physical properties and resistance to machine learning attacks, enabling joint optimization of multi-dimensional system metrics. Methods To mitigate modeling vulnerabilities of conventional strong PUFs and resolve the tradeoff between security and reliability, this paper presents a DSC-PUF architecture with dual delay chains and dynamic spatial perturbation. It consists of a primary chain A, an auxiliary chain B, and a dynamic spatial coupling network. As signals propagate, the primary chain captures multi-level delay competition outcomes to produce random hidden states governed by physical process variations. Using these hidden states as dynamic control signals, nonlinear transformations including mask modulation, long-range shift diffusion, and symmetric interleaved XOR are applied to input challenges. This dynamically varies the challenge space seen by the auxiliary chain and breaks the stable linear CRP mapping inherent in traditional PUFs. To suppress dynamic interference to system stability, a master-slave asynchronous triggering mechanism is introduced. It temporally isolates hidden state latching, spatial coupling and auxiliary-chain propagation to prevent glitch propagation and race hazards. A majority voting scheme is also utilized for response denoising to enhance output robustness under harsh operating conditions. Results and Discussions The proposed architecture was implemented on a Xilinx Artix-7 FPGA platform, and its performance was evaluated through multiple standard tests. Experimental results demonstrate that the DSC-PUF significantly enhances the resistance of conventional delay-based PUFs against machine learning modeling attacks. Under typical attacks, including Logistic Regression, Support Vector Machine, Deep Neural Network, Covariance Matrix Adaptation Evolution Strategy, Convolutional Neural Network, and Light Gradient Boosting Machine, the prediction accuracy remains between 50.28% and 53.39%, close to the random-guessing level (Fig. 5 ), indicating that the dynamic spatial reconfiguration mechanism effectively disrupts the stable feature distribution of conventional CRPs. The proposed structure has a minimum reliability of 98.49% in the temperature range of -40℃ to 80℃, satisfying the reliability requirements of IoT devices in complex environments (Fig. 6 ). The generated responses successfully passed the NIST SP800-22 randomness test suite (Table 1 and Fig. 7 ), while the autocorrelation coefficients remained within the 95% confidence bounds of ±0.0141 (Fig. 8 ), demonstrating favorable randomness and low data correlation. In addition, the proposed architecture achieved a uniqueness value of 0.5152 and an average Strict Avalanche Criterion value of approximately 0.517 (Fig. 9 and 10 ), both close to the theoretical ideal values. The hardware overhead of the proposed design is 606 LUTs and 350 FFs (Table 2 ), demonstrating good lightweight implementation efficiency. Compared with existing strong PUF structures, the proposed DSC-PUF achieves a better balance among attack resistance, reliability, randomness, and hardware overhead. Conclusions This paper proposes a dynamic spatial coupling strong PUF architecture with enhanced resistance against machine learning attacks. By introducing a hidden-state-driven dynamic spatial coupling mechanism, the proposed architecture dynamically reconstructs challenge features, thereby disrupting the stable linear characteristics of conventional delay-based PUFs and reducing the statistical learnability of challenge–response mappings. Experimental results demonstrate that the DSC-PUF effectively resists LR, SVM, DNN, CMA-ES, CNN, and LightGBM modeling attacks, with prediction accuracies reduced to near-random guessing levels, while maintaining favorable reliability, randomness, uniqueness, and avalanche characteristics. This architecture provides a reliable hardware security solution for device authentication and secure key generation in resource-constrained IoT environments.
Available online , doi: 10.11999/JEIT261058
Abstract:
Objective Warehouse robot fleets allocate streaming pickup-and-delivery tasks in narrow aisles. Packet loss can hide a healthy robot, a physical failure can invalidate its assignment, and a blockage can propagate queues. Existing studies often separate communication, task allocation, and multi-agent path finding. Scheduling may therefore use simulator-side states that were never delivered, while evaluation may omit failed transmissions, retries, or relay hops. This inconsistent information boundary obscures whether recovery is deployable. This study develops a method whose reconfiguration feasible domain is constrained by delivered-packet evidence. Methods A communication-availability-constrained event-triggered reconfiguration method, termed CAR-ETR, is proposed. Its central contribution is a cross-layer hard gate rather than a new event trigger, Hungarian matcher, congestion weight, or conflict resolver. Delivered heartbeat, retransmission, or relay packets form the availability set \begin{document}$ {\mathcal{A}}_{t} $\end{document} \begin{document}$ \mathcal{R}_{t}^{a} $\end{document} \begin{document}$ {x}_{ik,t} $\end{document} \begin{document}$ {R}_{c} $\end{document} \begin{document}$ {\mathcal{A}}_{t} $\end{document} \begin{document}$ H=16 $\end{document} \begin{document}$ \beta =0.88 $\end{document} \begin{document}$ 31\times 23 $\end{document} \begin{document}$ t $\end{document} \begin{document}$ 5\times 5 $\end{document} \begin{document}$ H $\end{document} \begin{document}$ \beta $\end{document} Results and Discussions CAR-ETR completes \begin{document}$ 128.75\pm 6.55 $\end{document} \begin{document}$ 62.99\pm 1.50 $\end{document} \begin{document}$ 27.85\pm 15.65 $\end{document} 1476 runs. Conclusions CAR-ETR establishes a communication-evidence gate from delivered packets to task-reconfiguration feasibility, with congestion guidance and fixed-point conflict resolution supporting execution. It improves service over local periodic replanning and reduces gateway coordination and online computation relative to strong rolling-horizon baselines, at the cost of more total packet attempts. Evidence remains limited to discrete-event simulation; continuous dynamics, wireless contention, localization error, energy constraints, permanent failures, and component interactions require hardware-in-the-loop and physical-platform validation.
Available online , doi: 10.11999/JEIT261062
Abstract:
Objective Human-Object Interaction (HOI) recognition aims to identify semantic interactions between people and objects and is an important task in high-level scene understanding, with applications in behavior analysis, human-machine interaction, assisted driving, and intelligent monitoring. Existing HOI methods are generally developed in the pixel domain, where compressed images must first be decoded before feature extraction and interaction prediction. This explicit pixel reconstruction may increase the exposure of sensitive visual content in privacy-sensitive scenarios and also makes recognition dependent on successful decoding. Since images are commonly stored and transmitted as compressed bitstreams, it is meaningful to investigate whether high-level interaction semantics can be recognized directly from bitstreams without reconstructing pixels. To this end, the Bitstream Human-Object Interaction Recognition (BHOIR) task is formulated, which directly takes the raw byte sequence of a compressed image as input and predicts interaction categories without pixel decoding or reconstruction. BHOIR is more challenging than conventional byte-level image classification because HOI categories have a large compositional label space and a severe long-tailed distribution. For example, HICO contains 600 interaction categories composed of 117 actions and 80 objects, with many combinations having limited training samples. Moreover, bitstream features do not explicitly correspond to high-level concepts such as actions and objects. Therefore, BHOIR requires effective exploitation of HOI compositional structure and enhanced semantic modeling of byte-level representations. Methods A bitstream-domain HOI recognition framework named ByteHOI is developed for BHOIR, consisting of bitstream feature extraction and interaction classification (Fig. 1 ). ByteFormer-Tiny is adopted as the encoder. Each JPEG image is represented by its raw byte sequence and truncated or zero-padded to 50,000 bytes. The bytes are converted into embeddings, reduced by a one-dimensional convolutional module, combined with positional embeddings, and processed by 12 window-based Transformer blocks. Global average pooling then produces a 192-dimensional bitstream representation without pixel decoding or reconstruction. Two components are introduced for interaction classification. First, action-object compositional decoupling decomposes HOI prediction into separate action and object classification branches, and interaction scores are reconstructed according to predefined action-object combinations. This enables categories sharing the same action or object to share supervision and alleviates the sparsity of rare interaction combinations. Second, a byte-text semantic alignment classifier introduces fixed semantic prototypes generated by the pretrained CLIP ViT-B/32 text encoder. Action and object prompts, such as “a person riding” and “a horse”, are encoded offline and L2-normalized. Two learnable projection networks map the 192-dimensional bitstream feature into separate 512-dimensional action and object semantic spaces. Classification logits are obtained from cosine similarities between projected features and fixed text prototypes, with a temperature coefficient of 0.01. Binary cross-entropy losses are applied to both branches, and final HOI scores are obtained by combining the corresponding action and object predictions. ByteHOI is trained for 50 epochs using AdamW with a batch size of 64. The initial learning rates of the bitstream encoder and classification module are set to \begin{document}$ 1\times {10}^{-5} $\end{document} \begin{document}$ 5\times {10}^{-4} $\end{document} Results and Discussions Experiments are conducted on the HICO and V-COCO datasets to evaluate the effectiveness and robustness of ByteHOI. General byte-level models, including MambaByte, MBLM, MegaByte, and bGPT, achieve Default mAP values below 2% on HICO, showing that generic byte-sequence modeling is insufficient for complex interaction recognition. ByteFormer, which benefits from visual byte-level pretraining, achieves 16.95% Default mAP, whereas ByteHOI improves the performance to 27.34%. Consistent improvements are observed across low-, medium-, and high-frequency interaction categories, particularly in the low-frequency group (Table 1 ). On V-COCO, the Default mAP is improved from 16.71% to 21.38%, and consistent gains are also obtained in Macro-Precision, Macro-Recall, and Macro-F1 on both datasets, demonstrating the cross-dataset effectiveness of the proposed framework (Table 2 ). Ablation experiments further confirm the contributions of the two key designs. On HICO, action-object compositional decoupling increases the Default mAP from 16.95% to 24.81%, while the introduction of CLIP semantic knowledge and cosine-based byte-text alignment further improves the performance to 27.34%. Similar trends are observed on V-COCO (Tables 3 and 4 ). Robustness is further evaluated by randomly flipping bits in the compressed data segment. Although some pixel-domain models achieve higher recognition accuracy on intact inputs, their performance decreases sharply as bitstream corruption increases because damaged files may suffer severe visual degradation or fail to decode. In contrast, ByteHOI maintains relatively stable performance. At a bit error rate of 0.01, it still achieves 24.82% and 19.72% mAP on HICO and V-COCO, respectively, corresponding to decreases of only 2.52 and 1.66 percentage points from the intact-bitstream setting (Fig. 2 ). Additional experiments show that ByteHOI is relatively insensitive to different JPEG encoder implementations and file-header modifications, whereas substantial changes in JPEG quality factor lead to more evident performance degradation, indicating that compression quality has a greater influence on bitstream distribution (Tables 5 and 6 ; Fig. 3 ). Conclusions The BHOIR task is formulated to investigate whether complex human-object interactions can be recognized directly from compressed image bitstreams without reconstructing pixel images during model inference. ByteHOI is developed to address two major difficulties in this setting. Action-object compositional decoupling is used to exploit the inherent compositional structure of HOI labels and alleviate the learning difficulty caused by sparse and long-tailed interaction categories, while byte-text semantic alignment is introduced to establish explicit associations between bitstream features and high-level action and object semantics using fixed CLIP text prototypes. Experimental results on HICO and V-COCO demonstrate that ByteHOI consistently outperforms the compared byte-level methods and achieves Default mAP values of 27.34% and 21.38%, respectively. The effectiveness of the proposed components is further supported by ablation studies. Bitstream corruption experiments show that ByteHOI maintains relatively stable recognition performance even when corrupted files become difficult or impossible to decode through conventional pixel-domain pipelines. By avoiding explicit reconstruction of visible pixel images during recognition, the proposed method reduces the exposure of pixel-level visual content in the recognition pipeline while preserving task-related interaction semantics. This privacy-related benefit is limited to the recognition process and does not constitute encryption or a formal privacy guarantee. The current study focuses on JPEG bitstreams. Future work will extend bitstream-domain HOI recognition to other image formats, including PNG, WebP, and AVIF, and will investigate unified bitstream representations with stronger adaptability across different encoding formats.
Available online , doi: 10.11999/JEIT260210
Abstract:
Objective With their high mobility, flexible deployment, and Line-of-Sight (LoS) links, Unmanned Aerial Vehicles (UAVs) are promising components of future wireless communication systems. However, the broadcast nature of LoS links exposes UAV communications to detection by malicious wardens. Covert communication is therefore used to prevent transmission activity from being detected. Meanwhile, Ultra-Reliable and Low-Latency Communications (URLLC) require short-packet (finite-blocklength) transmission, for which decoding errors are non-negligible and the infinite-blocklength assumption is unsuitable. This makes UAV-assisted short-packet covert communication an important research topic. Nevertheless, most existing studies consider a single UAV and a static environment, limiting communication service capacity and applicability in dynamic scenarios. To address this research gap, this study investigates a multi-UAV-assisted short-packet covert communication system. Methods A system model for multi-UAV-assisted short-packet covert communication is established, in which multiple UAVs serve multiple randomly moving covert users while a moving aerial warden attempts to detect their transmissions. Cooperative jammers are selected from public users to transmit jamming signals that impair the warden’s detection capability. User mobility is modeled using the Gauss-Markov random mobility model. A binary hypothesis-testing model is formulated at the warden, from which the Detection Error Probability (DEP) is derived. An optimization problem is formulated to maximize short-packet covert throughput under the covertness constraint by jointly optimizing UAV trajectories, cooperative jammer selection, and covert users’ transmit powers. A short-packet covert communication algorithm based on Multi-Agent Proximal Policy Optimization (MAPPO), termed MAPPO-SPCC, is proposed. Specifically, MAPPO is used to train multiple agents to learn coordinated UAV trajectory-control policies for dynamic environments. During execution, each agent outputs its UAV’s velocity at each time slot, thereby determining UAV trajectories and user associations. The algorithm then selects cooperative jammers and optimizes covert users’ transmit powers by solving a convex optimization problem. Results and Discussions Simulations are conducted to evaluate the performance of MAPPO-SPCC. The reward curves show stable convergence at an appropriate learning rate and indicate that training remains robust as user mobility increases progressively (Fig. 2 ). The loss curves further support the stability of the training process (Fig. 3 ). The trajectory plots show that the agents adaptively adjust UAV paths in response to the spatial distribution of covert users and the aerial warden’s patrol path (Fig. 4 ). The corresponding covert throughput heatmaps show that high-throughput regions overlap with UAV hovering regions (Fig. 5 ), indicating that the learned policies identify where UAV positioning yields greater covert throughput based on the observed environmental states. (Fig. 6 ) further examines how blocklength and the covertness constraint affect total covert throughput. Relaxing the covertness constraint increases achievable throughput because it permits higher transmit power. Conversely, increasing the blocklength reduces total throughput in the simulations. A larger blocklength allows the warden to collect more detection samples, improving its statistical detection capability and making the covertness constraint more stringent. Covert users must therefore reduce their transmit powers to satisfy the constraint, resulting in lower total throughput. Comparisons with the Multi-Agent Deep Deterministic Policy Gradient (MADDPG), Deep Q-Network (DQN), and Ant Colony Optimization (ACO) algorithms further show that MAPPO-SPCC consistently achieves the highest covert throughput (Fig. 7 ). This advantage is attributable to MAPPO-SPCC’s ability to coordinate UAV trajectories across agents and update policies on-policy as the environment changes. Conclusions This study investigates a multi-UAV-assisted short-packet covert communication system in which multiple UAVs collaboratively serve multiple randomly moving covert users while cooperative jamming impairs a moving aerial warden’s detection capability. The MAPPO-SPCC algorithm jointly optimizes UAV trajectories, cooperative jammer selection, and covert users’ transmit powers to maximize short-packet covert throughput under the covertness constraint. The simulation results show that MAPPO-SPCC adapts UAV trajectories to dynamic user and warden positions and achieves higher covert throughput than the MADDPG, DQN, and ACO baseline algorithms. Future work may integrate Physical Layer Security (PLS) with covert communication to improve both transmission covertness and content confidentiality. Extending trajectory optimization to three dimensions may further improve system flexibility and performance.
Available online , doi: 10.11999/JEIT260273
Abstract:
Objective Mobile Sensor Networks (MSNs) provide flexible deployment, dynamic sensing coverage, and cooperative observation capabilities for target tracking. However, nonlinear target motion, time-varying wireless communication topologies, intermittent measurements, and malicious attacks jointly complicate target tracking in MSNs. In open wireless communication environments, false data injection, denial-of-service, and replay attacks may corrupt measurement information and cause abnormal state estimates to propagate among neighboring nodes, thereby degrading estimation accuracy and potentially compromising filter consistency and stability. In addition, the target state considered in this study comprises orientation and position variables, whose evolution has an intrinsic non-Euclidean geometric structure. Conventional Extended Kalman Filter (EKF) methods generally linearize the system directly in Euclidean coordinates, which can introduce approximation errors and violate the orthogonality of the rotation matrix. Existing resilient distributed estimation methods also commonly rely on static sensor networks or restrictive linear-motion assumptions, limiting their applicability to nonlinear target tracking under dynamically switching network topologies. Therefore, reliable target tracking in MSNs requires a distributed state estimation method that accounts for geometric consistency, attack resilience, and dynamic network connectivity. Methods A distributed resilient invariant extended Kalman fusion filtering method is developed for nonlinear target tracking under network attacks. First, a Lie group-Lie algebra mapping is introduced into the EKF framework to construct an Invariant Extended Kalman Filter (IEKF). The target pose is represented on the two-dimensional Special Euclidean group (SE(2)), and the estimation error is defined through group multiplication. The error is then mapped to the corresponding Lie algebra, where linearization is performed. This geometric error representation preserves the structural properties of the rotation matrix during state propagation and avoids the loss of orthogonality caused by direct Taylor-expansion linearization. Prediction and measurement updates are formulated using the invariant error representation to obtain prior and posterior state estimates and their covariance matrices. Second, a dimensionality-reduced k-means attack detection mechanism is designed to assess local node credibility. For each neighboring node, the deviation between its posterior estimate and the local node’s fused prior estimate is reduced from a high-dimensional state difference to a scalar dissimilarity measure. These scalar dissimilarity measures are then clustered into two groups using k-means, and the cluster with the smaller centroid value is classified as the normal-node group. This mechanism identifies compromised or untrusted nodes without prior knowledge of the attack model and reduces the computational cost of clustering high-dimensional state deviations. The posterior estimates of identified compromised nodes are excluded from the trusted information set. Third, a covariance-intersection-based consensus fusion strategy is developed for time-varying network topologies. At each time step, a sensor node combines its own posterior estimate with the posterior estimates of trusted neighbors identified by k-means and the prior estimates of untrusted neighbors. The fusion weights are determined by the current network topology and node connectivity. Therefore, each node requires only local neighborhood information, and no global network information is needed. The resulting four-stage procedure consists of prediction, measurement update, attack detection, and consensus fusion. Results and Discussions MATLAB simulations are conducted to evaluate the estimation accuracy, stability, computational efficiency, and attack resilience of the proposed method. A mobile sensor network with six sensors is used to track a nonlinear moving target. The communication topology switches every two time steps, as shown in Fig. 1 , and the target and sensor trajectories are presented in Fig. 2 . Each sensor obtains a valid measurement only when the target is within its sensing range. The method is first evaluated in two cases: without attack detection and with the proposed attack detection mechanism. The corresponding state estimates are shown in Figs. 3 and 4 . Without attack detection, malicious measurements cause the state estimates of several sensors to deviate markedly from the true target state. With the proposed detection mechanism enabled, corrupted measurements are screened out, and the state estimates from all six sensors gradually converge to the true target state, indicating improved estimation stability and tracking accuracy. A grouped false data injection attack is then simulated. Every 10 time steps, sensors 1 and 5 are attacked simultaneously, as are sensors 2 and 6, while the remaining sensors operate normally. Figs. 5 and 6 show the tracking trajectories without and with attack detection, respectively. The results indicate that the proposed detection mechanism identifies compromised nodes and prevents their abnormal measurements from dominating the distributed fusion process. Figs. 7 and 8 show the corresponding angular and relative-position estimation errors. Without attack detection, the estimation errors fluctuate significantly when attacks occur. When attack detection is combined with resilient fusion, the errors remain bounded and become more stable, indicating that the proposed method suppresses the propagation of malicious information under dynamic topology switching. For a quantitative comparison, the Root Mean Squared Error (RMSE) values for the angular and position estimates are compared among the proposed method, an adaptive-threshold detection method, and an invariant set-membership filtering method. All three methods are evaluated under the same target motion, network topology, noise, and attack conditions. As shown in Fig. 9 , the proposed method maintains lower estimation errors and converges faster than both comparison methods. This improvement is attributable to two complementary components: the Lie group-Lie algebra-based IEKF, which preserves the geometric structure of the state representation, and the combined attack detection and covariance-intersection-based consensus fusion strategy, which limits the influence of abnormal information. Computational efficiency is also evaluated by running each method independently 100 times under identical conditions. As reported in Table 1 , the average running time of the proposed method is 0.0340 s, compared with 0.0809 s for the adaptive-threshold detection method, 0.0816 s for the invariant set-membership filtering method, and 0.0357 s for the high-dimensional k-means method. The corresponding computation-time reductions are approximately 57.9%, 58.3%, and 4.7%, respectively. These results show that reducing high-dimensional state deviations to scalar dissimilarity measures lowers the computational cost of attack detection while preserving information about deviations caused by malicious measurements. Conclusions A distributed resilient invariant extended Kalman fusion filtering method is proposed for target tracking in MSNs subject to nonlinear target motion, time-varying communication topologies, and malicious measurement attacks. The Lie group-Lie algebra representation aligns the error definition and linearization with the system’s geometric structure, helping preserve consistency in nonlinear state estimation. The dimensionality-reduced k-means mechanism converts high-dimensional estimation deviations into scalar dissimilarity measures, enabling attack identification without a predefined attack model while reducing the computational burden. The covariance-intersection-based consensus fusion strategy combines each node’s own posterior estimate and trusted neighbors’ posterior estimates with untrusted neighbors’ prior estimates according to the current local topology, thereby limiting the influence of compromised nodes while preserving distributed operation. Simulation results show that the proposed method achieves lower estimation errors, more stable tracking performance, and lower computational cost than the comparison methods under the tested attack and dynamic-topology conditions. The proposed framework therefore provides an effective distributed solution for resilient target state estimation in attack-prone MSNs.
Available online , doi: 10.11999/JEIT260339
Abstract:
Objective Disasters in complex, high-rise, and ultra-high-rise buildings occur frequently in China, and their incidence has continued to rise. This trend creates an urgent need for emergency rescue command systems adapted to these environments. The complex structures of these buildings can severely attenuate wireless signals, creating communication blind spots at indoor-outdoor transitions and floor boundaries. To address these coverage limitations, the deployment of “dual-use” emergency communication networks designed for both routine operations and disaster response has become an important strategy for improving urban safety and resilience. Unmanned Aerial Vehicles (UAVs), with their lightweight design, high payload capacity, and mobility, provide a reliable platform for deploying these networks. However, UAV-assisted networks in complex building emergency scenarios remain constrained by limited network coverage and backhaul link capacity. Integrated Access and Backhaul (IAB) technology offers a potential architectural solution by supporting wireless access and backhaul over shared spectrum resources. Nevertheless, deploying UAV-assisted IAB networks in emergency scenarios remains challenging. Concurrent user access creates localized traffic surges, while partial infrastructure damage further intensifies network interference. Dense node deployment, high UAV mobility, and complex channel conditions also cause interference relationships to vary over time and become non-stationary. Existing resource allocation methods generally assume stationary environments. Therefore, their allocation decisions may become mismatched with the evolving network topology and interference structure. This study therefore proposes a joint time-frequency resource optimization method to reduce network conflicts under time-varying and non-stationary interference conditions. Methods A segmented composite channel model is established to characterize electromagnetic propagation in complex buildings affected by disasters. The model accounts for line-of-sight links between IAB nodes within indoor or outdoor environments, as well as penetration links between indoor and outdoor nodes. To capture abrupt changes in channel conditions caused by complex environmental factors, a Rician small-scale fading model is introduced. The Rician factor is assumed to follow a piecewise-constant probability distribution, with abrupt changes at selected system frames and constant values between change points. Graph theory is then used to represent the time-varying network topology and interference relationships as a non-stationary directed conflict graph. Communication links are modeled as vertices, and interference relationships are represented as edges. Two types of constraints are modeled: link-scheduling constraints and resource mutual-exclusion constraints. The joint time-frequency resource allocation problem is subsequently formulated as a non-stationary directed dual-graph coloring problem. A minimum-conflict optimization model with soft constraints is developed, in which both constraint types are incorporated as penalty terms in the objective function. An Exponential Forgetting Upper Confidence Bound (EF-UCB) algorithm is proposed within a multi-player Multi-Armed Bandit (MAB) framework to solve this problem. Each communication link is treated as an independent player, and each joint time-frequency resource combination is treated as an available action. An exponential forgetting mechanism updates historical action counts and accumulated rewards. By assigning exponentially decreasing weights to past observations through a forgetting factor, the algorithm reduces the influence of outdated information and supports adaptive, distributed online decision-making under non-stationary interference conditions. Results and Discussions The proposed EF-UCB algorithm is evaluated through Monte Carlo simulations in a three-dimensional emergency scenario involving complex buildings and compared with the Greedy algorithm, the Color Diffusion Algorithm (CDA), Q-Learning (QL), and an ε-greedy-based MAB algorithm. Simulation results show that increasing the number of time slots reduces all three evaluated metrics: the link-scheduling constraint conflict rate, the resource mutual-exclusion constraint conflict rate, and the average total network conflict rate (Fig. 3 ). As the number of subchannels increases, the resource mutual-exclusion constraint conflict rate and the average total network conflict rate also decrease, whereas the link-scheduling constraint conflict rate changes relatively little (Fig. 4 ). Across the tested resource configurations, EF-UCB achieves the lowest average total network conflict rate among the compared algorithms (Figs. 3 and 4 ). In terms of computational efficiency, EF-UCB completes online decisions at the millisecond scale and has a shorter average decision time than the Greedy and CDA algorithms (Figs. 3 and 4 ). As the number of UAVs increases, the average decision time of EF-UCB remains relatively low and rises only slightly, indicating good scalability for dense emergency deployments (Fig. 6 ). Compared with the other learning-based methods, EF-UCB reduces the average total network conflict rate more rapidly as the number of system frames increases (Fig. 5 ). When the number of environmental change points increases, EF-UCB exhibits the smallest increase in the average total network conflict rate among the compared algorithms (Fig. 7 ). The exponential forgetting mechanism downweights outdated observations, enabling EF-UCB to adapt more effectively to increasingly non-stationary interference conditions (Fig. 7 ). Conclusions This study addresses the joint time-frequency resource optimization problem for UAV-assisted IAB networks subject to time-varying and non-stationary interference in complex building emergency scenarios. By integrating a segmented composite channel model with a non-stationary directed dual-graph coloring framework, the proposed method represents the propagation characteristics and dynamic interference relationships of these environments within a unified resource allocation framework. The EF-UCB algorithm uses exponential forgetting to reduce the influence of outdated observations and adapt resource allocation to changing interference conditions. Simulation results demonstrate that the proposed method reduces the average total network conflict rate while maintaining relatively low computational complexity and short average decision times. Its polynomial per-frame time complexity further supports online decisions within a single system frame. These results indicate that the proposed method can support timely resource allocation in UAV-assisted emergency communication networks operating under non-stationary interference conditions.
Available online , doi: 10.11999/JEIT260344
Abstract:
Objective Data heterogeneity under Non-Independent and Identically Distributed (Non-IID) settings remains one of the major challenges in Personalized Federated Learning (PFL) because it causes inconsistent local optimization, insufficient global knowledge transfer, and degraded personalization performance. To address these challenges, this paper proposes FedFACO, a personalized federated learning method based on the Fisher Information Matrix (FIM) for Adaptive Aggregation (AA) and Client Collaborative Optimization (CO). The proposed method improves model adaptation to heterogeneous client data distributions while preserving effective global knowledge sharing. By integrating AA with CO, FedFACO achieves a better balance between global generalization and local personalization in complex Non-IID environments. Methods FedFACO consists of two complementary components. First, the AA mechanism dynamically adjusts the fusion weights between the global and local models to generate personalized model initialization that better matches local data distributions. Second, the CO mechanism combines Parameter Alignment (PA) and Client Synchronization (CS). The PA strategy performs feature alignment and global regularization to strengthen global knowledge transfer while preserving local feature representations. The CS strategy uses FIM to quantify the information content of each client update and performs weighted aggregation to suppress low-quality updates. The proposed method is evaluated on the MNIST, CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets under Non-IID settings and is compared with representative baseline methods. Convergence behavior, robustness to stragglers, and resilience to low-quality updates are also evaluated. Results and Discussions Experimental results demonstrate that FedFACO consistently outperforms representative baseline methods on all four benchmark datasets, achieving an average accuracy improvement of approximately 3.1% over existing methods (Fig. 1 and Table 2 ). On the more challenging Tiny-ImageNet dataset, FedFACO reduces the total training time required for convergence by approximately 4.8% compared with the best-performing baseline (Table 3 ). Ablation experiments demonstrate that both AA and CO substantially improve performance, and their combined use achieves the highest accuracy (Table 6 ). Furthermore, the FIM-based CS strategy accurately quantifies the information content of client updates in straggler and asynchronous training scenarios, thereby improving aggregation reliability (Figs. 2 and 3 ). The proposed method also exhibits superior robustness in the presence of malicious clients (Fig. 4 ). Conclusions This paper proposes FedFACO, a personalized federated learning method for Non-IID environments based on FIM for AA and CO. The proposed method effectively balances global knowledge sharing and local personalization while improving training stability and robustness under heterogeneous client participation. Experimental results demonstrate its advantages in prediction accuracy, convergence efficiency, and robustness. Future work will investigate lightweight FIM approximation methods and adaptive triggering strategies to reduce computational overhead and integrate the proposed framework with privacy-preserving and security mechanisms for deployment in resource-constrained and high-security environments.
Available online , doi: 10.11999/JEIT261119
Abstract:
Objective Spaceborne dual-beam along-track interferometric synthetic aperture radar (DBATI-SAR) acquires fore- and aft-looking radial velocities along different ground-projected radar line-of-sight (RLOS) directions, providing the directional diversity required for two-dimensional ocean surface current vector retrieval. The look angle is a key geometric design parameter because it affects both the horizontal RLOS projection and the interferometric radial-velocity accuracy. Increasing the look angle generally increases the local incidence angle, thereby strengthening the horizontal RLOS projection and reducing geometric error amplification. Meanwhile, the associated increases in slant range and incidence angle weaken ocean-surface backscatter and degrade the signal-to-noise ratio (SNR) and interferometric coherence. This study investigates this tradeoff and determines an appropriate look-angle range for spaceborne DBATI-SAR current-vector measurements. Methods A unified random-error model is developed by integrating dual-beam radial-velocity projection, linear covariance propagation, spaceborne observation geometry, normalized radar cross section (NRCS), noise-equivalent sigma zero (NESZ), SNR, interferometric coherence, effective number of looks, ATI phase noise, and phase-to-velocity conversion. The square root of the maximum eigenvalue of the two-dimensional velocity-error covariance matrix is adopted as the two-dimensional random current-retrieval error metric. Under orthogonal ground-projected RLOS geometry and approximately symmetric fore- and aft-looking channels, a compact analytical expression is derived to characterize the competition between geometric error reduction and radial-velocity error growth. Numerical simulations employ a Ku-band reference configuration while retaining the actual incidence angle, slant range, NRCS, NESZ, SNR, and radial-velocity error of each channel. Orbit altitude, nominal NESZ, ocean wind speed, and ATI baseline are further varied to examine their effects on the optimum look angle. A near-optimal interval is defined by requiring the error under each operating condition not to exceed 1.2 times its corresponding minimum. Results and Discussions The results confirm the competing effects of the look angle on projection geometry and interferometric velocity accuracy. As the look angle increases, both the mean incidence angle and slant range increase, whereas ocean-surface NRCS decreases and NESZ increases, resulting in progressive SNR degradation at large look angles (Fig. 3 ). The geometric error amplification factor decreases monotonically, whereas the radial-velocity error increases rapidly at large angles; their combined effect therefore produces an interior optimum (Fig. 4 ). Under the reference configuration, the minimum two-dimensional random current error is 0.0054 m/s at a look angle of 28.50°, corresponding to a mean incidence angle of 42.43°. Increasing the orbit altitude from 500 to 1000 km shifts the optimum look angle from 34.75° to 25.00°, while degrading the nominal NESZ from –35 to –25 dB shifts it from 33.00° to 24.00°. Increasing the wind speed from 4 to 12 m/s shifts the optimum from 24.25° to 31.25°, whereas increasing the ATI baseline from 4 to 20 m shifts it from 25.50° to 31.25° (Fig. 5 ). The intersection of the near-optimal intervals for all representative conditions yields a recommended look-angle range of 25.75°–31.00°, within which the worst-case normalized error does not exceed 1.2 (Fig. 6 ). Conclusions The optimum look angle for spaceborne DBATI-SAR is governed by the tradeoff between reduced geometric error amplification and degraded radial-velocity accuracy at large angles. The proposed error framework quantitatively links observation geometry, system sensitivity, and ocean-surface scattering. The resulting look-angle range of 25.75°–31.00° provides a practical reference for the joint design of observation geometry and system sensitivity for spaceborne DBATI-SAR ocean surface current vector measurements.
Available online , doi: 10.11999/JEIT260956
Abstract:
Objective This paper investigates the cooperative pursuit-evasion problem of multiple unmanned aerial vehicles (UAVs) against a highly maneuverable evader in low-altitude urban environments. With the rapid growth of low-altitude UAV applications, unauthorized UAVs pose increasing threats to airspace security. Compared with conventional counter-UAV measures, cooperative pursuit by multiple UAVs offers a more flexible solution; however, when the evader outperforms each pursuer in maneuverability, capture can only be achieved through coordinated decision-making that restricts the evader's escape space. The objective is therefore to develop an efficient multi-UAV cooperative pursuit strategy achieving robust capture under dynamics constraints, obstacle interference, and speed disadvantage. Methods To address this problem, a cooperative pursuit-evasion scenario is constructed in a bounded urban area with static rectangular obstacles, where the motion of all UAVs is described by a damped double-integrator model that captures the inertia and velocity-attenuation characteristics of UAV dynamics. The pursuit problem is then formulated as a Markov decision process in which each UAV acts on local observations. The pursuers are trained with the Multi-Agent Soft Actor-Critic (MASAC) algorithm under the centralized training and decentralized execution framework, while the evader is trained simultaneously with Soft Actor-Critic (SAC), forming an adversarial learning process. To improve exploration efficiency, a progressive adversarial imitation learning framework is introduced: expert demonstrations are generated offline via Apollonius-circle analysis under speed disadvantage, Generative Adversarial Imitation Learning provides intrinsic imitation rewards in early training, and an exponential decay mechanism gradually shifts the pursuer policy from imitation-driven to autonomous reinforcement learning. Furthermore, a velocity-ratio-based curriculum progressively increases the evader's maximum speed, enhancing policy adaptability against highly maneuverable targets. Results and Discussions Simulation experiments are conducted on a Python-based multi-UAV pursuit-evasion platform, where the urban airspace contains static rectangular buildings, three pursuer UAVs (maximum speed 3 m/s), and one evader UAV (3.9 m/s). Results show that the proposed framework outperforms baseline reinforcement learning methods and converges to a stable reward level after approximately 2000 episodes. Ablation studies demonstrate that Apollonius-circle based expert demonstrations significantly accelerate early-stage learning, while velocity-ratio curriculum learning enhances final pursuit performance; their combination yields a synergistic improvement in both convergence speed and policy quality. Experiments under speed ratios of 1:1, 1:1.3, and 1:1.6 further verify robustness: the capture success rate approaches 100% without a speed disadvantage and remains high as the evader's speed advantage increases. Conclusions This paper proposes a progressive adversarial imitation learning framework for multi-UAV cooperative pursuit of a highly maneuverable evader. By integrating MASAC-based cooperative decision-making, adversarial evader training, Apollonius-circle expert demonstrations with GAIL imitation rewards, and velocity-ratio curriculum learning, the framework significantly improves learning efficiency and capture performance. Simulation results verify faster convergence, higher capture success rates, and stronger adaptability across different evader speeds, providing an effective learning-based solution for low-altitude cooperative interception of highly maneuverable targets.
Available online , doi: 10.11999/JEIT260875
Abstract:
Objective Low-altitude intelligent platforms support aerial inspection, smart logistics, emergency search and rescue, and disaster monitoring by transmitting image observations to edge or ground receivers. Reliable scene recognition depends on preserving the information needed for these tasks over wireless links with limited bandwidth and changing channel conditions. Conventional image transmission emphasizes accurate recovery of pixels or bits, which can consume communication resources for image content that contributes little to the downstream task. Task-oriented semantic communication instead prioritizes discriminative information for scene perception. However, platform motion, occlusion, and multipath propagation can disturb the transmission link, while small input perturbations can shift the extracted semantic features and change the predicted category. Low-altitude images also contain fine local textures and complex spatial layouts, so reliable scene perception requires both local detail and global structural information. A reliable semantic transmission framework for low-altitude image perception is therefore proposed to preserve task-relevant features under dynamic channels and semantic noise. The objective is to improve classification accuracy at the receiver through the joint design of robust feature extraction, compact semantic transmission, and a shared semantic knowledge base. Methods The framework models semantic noise as a constrained task perturbation generated by the Fast Gradient Sign Method (FGSM) to increase classification loss while keeping visual changes small. The framework contains a Robust Feature Extraction Module (RFEM), a Joint Source-Channel Coding (JSCC) encoder and decoder, a shared Semantic Knowledge Base (SKB), and a scene perception module, with the transmitter placed at the low-altitude platform and the receiver at an edge or ground node (Fig. 1 ). RFEM combines a local multi-scale branch and a global State Space Duality (SSD) branch. The local branch uses depthwise separable convolutions with kernel sizes of 3 by 3, 5 by 5, and 7 by 7 to extract textures, edges, and local structures at different scales. The global branch employs non-causal SSD modeling to capture long-range spatial dependencies among image regions. Both branches provide 256-dimensional features, which are combined through an attention fusion block with eight heads to form a representation less sensitive to semantic noise (Fig. 2 ). The JSCC encoder contains five convolutional layers to compress features, while the corresponding decoder uses five deconvolutional layers to recover task-relevant semantic representations at the receiver (Fig. 3 ). The SKB contains trainable semantic primitive vectors shared by the transmitter and receiver. Continuous encoded features are matched to the nearest primitive vectors, and their indices are transmitted for retrieval at the receiver. During training, a straight-through estimator approximates gradients through the non-differentiable nearest-neighbor mapping. Feature-to-primitive alignment and reduced similarity between different primitive vectors are encouraged during training to strengthen discriminability and limit semantic feature confusion. The recovered representation is perceived using global average pooling, a fully connected layer, and a Softmax function (Fig. 4 ). The system is optimized end to end using cross-entropy classification loss and SKB-related constraints. Evaluation uses UCMerced-LandUse, containing 21 scene categories, and NWPU-RESISC45, containing 45 categories. Results are averaged over three independent runs. Results and Discussions Classification accuracy is evaluated over Additive White Gaussian Noise (AWGN) and Rayleigh fading channels. The compared methods include Joint Photographic Experts Group (JPEG) image compression with Low-Density Parity-Check (LDPC) channel coding, DT-JSCC with discrete task representations and digital modulation, and ASE-JSCC with attention-based semantic encoding and decoding. At an SNR of 3 dB on UCMerced-LandUse without semantic noise, the proposed method achieves accuracies of 87.07% and 85.62% over AWGN and Rayleigh fading channels, respectively. These values exceed ASE-JSCC by 2.83% and 1.69%. With semantic noise, the corresponding accuracies are 83.01% and 80.68%, exceeding ASE-JSCC by 3.75% and 3.25% (Table 1 ). On NWPU-RESISC45 under the same channel conditions, the accuracies of our method without semantic noise are 76.12% and 73.31%, with gains of 5.28% and 3.74% over ASE-JSCC. With semantic noise, the framework achieves 72.54% and 71.42%, compared with 61.62% and 59.03% for ASE-JSCC, giving gains of 10.92% and 12.39% (Table 2 ). The larger gains on this dataset indicate the benefit of maintaining discriminative semantics across a wider range of scene categories. Ablation experiments at an SNR of 3 dB establish the contributions of the two principal modules. On UCMerced-LandUse, removing RFEM reduces accuracy to 77.38% and 75.79% under AWGN and Rayleigh fading, while removing SKB reduces it to 71.47% and 69.05%. Removing both modules gives 63.85% and 57.96%, compared with 87.07% and 85.62% for the complete framework (Table 3 ). On NWPU-RESISC45, the corresponding results are 66.27% and 62.93% without RFEM, 61.03% and 58.74% without SKB, and 52.38% and 50.44% without both modules, compared with 76.12% and 73.31% for the complete framework (Table 4 ). These reductions support the complementary contributions of stable feature extraction and discrete semantic representation, with removal of SKB producing the larger loss on both datasets. Across all SNRs, the framework achieves the highest classification accuracy among the compared methods on UCMerced-LandUse with and without semantic noise (Fig. 5 , Fig. 6 ). The same advantage can be observed on NWPU-RESISC45 under both channel models (Fig. 7 , Fig. 8 ). As semantic noise intensity increases from 0 to 0.036, classification accuracy generally decreases, but the framework retains the highest accuracy on both datasets, demonstrating robustness over different noise perturbation (Fig. 9 , Fig. 10 ). Complexity measurements show that the proposed model contains 4.77 million parameters, lower than 11.28 million for DT-JSCC and 14.54 million for ASE-JSCC. Its computational cost is 6.45 billion Floating-Point Operations (FLOPs), and its inference latency is 6.93 ms. Although its computational cost and latency exceed those of ASE-JSCC, its parameter count is lower and its classification accuracy is higher, showing the tradeoff between model size, computation, and task reliability (Table 5 ). Conclusions The proposed framework improves task-oriented semantic transmission for low-altitude image perception by preserving local discriminative details and global structural information, compressing task-relevant features through JSCC, and reinforcing their representation through a shared SKB. The experimental comparisons show higher classification accuracy than JPEG+LDPC, DT-JSCC, and ASE-JSCC over the AWGN and Rayleigh fading channels with semantic noise. The ablation results substantiate the contributions of RFEM and SKB. These results demonstrate the effectiveness of the framework for reliable low-altitude scene recognition under different communication conditions.
Available online , doi: 10.11999/JEIT261031
Abstract:
Objective With the rapid development of the low-altitude economy, unmanned aerial vehicles (UAVs) have been widely used in aerial photography, logistics, environmental monitoring, and emergency rescue. However, unauthorized UAV operations pose increasing challenges to public safety and low-altitude airspace management. Vision-based anti-UAV detection has attracted considerable attention due to its relatively low cost, rich visual information, and convenient deployment. Compared with conventional objects, UAV targets in low-altitude surveillance images often exhibit small scales, weak textures, significant pose variations, and complex background interference. During repeated feature downsampling, the edge and structural information of these targets may be weakened or lost, resulting in missed detections, false alarms, and inaccurate localization. Furthermore, practical anti-UAV detection systems must balance detection accuracy, computational complexity, and real-time performance under limited hardware resources. Existing lightweight detection networks may struggle to preserve discriminative spatial details while maintaining computational efficiency, particularly when UAV targets occupy only a small proportion of the input image. To address these challenges, a lightweight anti-UAV object detection network named DCP-YOLO is proposed based on YOLOv26n. The objective is to enhance the feature representation and detection accuracy of tiny UAV targets while maintaining a compact model structure and real-time inference capability. Methods The proposed DCP-YOLO is developed based on the lightweight YOLOv26n framework to improve the detection performance of tiny UAV targets while maintaining real-time inference. The network consists of three collaboratively optimized components. In the backbone network, a lightweight feature extraction module named DPC3k2 is designed by integrating depthwise separable convolution (DSConv) and pinwheel-shaped convolution (PConv) to enhance multi-directional feature representation with reduced computational complexity, enabling more discriminative feature extraction for small UAV targets. DSConv reduces redundant spatial computation through the separation of spatial feature extraction and channel mapping, while PConv employs asymmetric convolutional branches to capture directional edge and contour information. During feature fusion, a learnable reverse upsampling module named CUS replaces the conventional interpolation-based upsampling operation to recover high-frequency edge and texture information, thereby improving feature reconstruction quality. In the neck network, a contextual attention enhancement module named C2PSCAA is constructed by incorporating Context Anchor Attention (CAA) into the C2PSA structure to strengthen long-range spatial dependency modeling and multi-scale contextual representation with only a slight computational overhead. Through the collaborative optimization of lightweight feature extraction, feature reconstruction, and contextual attention modeling, the proposed network effectively enhances the representation capability of tiny UAV targets while maintaining low computational complexity and high inference efficiency. Results and Discussions Extensive experiments are conducted on the DUT Anti-UAV, TIB-Net, and VisDrone2019 datasets. The ablation results in Tables 2 and 4 demonstrate that the proposed DPC3k2, CUS, and C2PSCAA modules contribute to lightweight feature extraction, high-frequency detail reconstruction, and contextual feature modeling, respectively. On the DUT Anti-UAV dataset, their collaborative integration improves Precision, Recall, mAP@50, and mAP@50-95 to 98.7%, 90.0%, 92.4%, and 61.7%, respectively, while reducing the number of parameters to 2.29 M and maintaining an inference speed of 210.71 FPS. Similar performance improvements are observed on the TIB-Net dataset, confirming the effectiveness and complementarity of the three modules under different target scales and background conditions. The visualization results in Fig. 6 further show that the proposed method reduces missed detections and false alarms for tiny UAV targets in cluttered scenes. The comparison results in Table 5 indicate that DCP-YOLO outperforms representative YOLO-series detectors in detection accuracy while maintaining lower model complexity and a smaller model size. Compared with the baseline YOLOv26n, mAP@50 and mAP@50-95 are improved from 90.0% and 59.4% to 92.4% and 61.7%, respectively, whereas the computational complexity is reduced from 5.2 GFLOPs to 5.0 GFLOPs. The convergence curves in Fig. 8 show that DCP-YOLO maintains higher Precision and mAP@50-95 during training. Furthermore, Table 6 demonstrates that the proposed method achieves competitive or superior performance compared with recent anti-UAV and small-object detection methods while preserving a lightweight architecture. The detection results in Fig. 7 visually confirm its stronger ability to identify tiny UAV targets under complex backgrounds. The cross-dataset results in Tables 7 and 8 show that DCP-YOLO also achieves improved detection performance on VisDrone2019 while reducing the number of parameters, computational complexity, and model size. Although VisDrone2019 differs substantially from anti-UAV datasets in target categories and imaging viewpoints, the proposed method still improves mAP@50-95 by 1.1% over YOLOv26n, demonstrating favorable cross-scene adaptability and generalization capability. To investigate practical deployment feasibility, the models are further evaluated on an NVIDIA Jetson Orin Nano platform using TensorRT acceleration under a 25 W power setting, as illustrated in Fig. 9 . As shown in Table 9 , DCP-YOLO achieves an mAP@50 of 94.59%, an mAP@50-95 of 62.70%, and an inference speed of 116.02 FPS in FP16 mode, with a memory consumption of 3.00 GB. Although its inference speed is lower than that of the baseline on the edge platform, DCP-YOLO maintains real-time inference capability while achieving higher detection accuracy. These findings indicate that reductions in theoretical computational complexity do not necessarily translate into proportional improvements in actual inference speed, particularly when frequency-domain operations are involved. Conclusions A lightweight anti-UAV object detection network named DCP-YOLO is proposed to address the challenges of weak feature representation, high-frequency detail degradation, and limited computational resources in low-altitude UAV detection. By jointly optimizing lightweight feature extraction, learnable upsampling, and contextual attention modeling, the proposed method enhances the detection accuracy of tiny UAV targets while maintaining a compact network structure. Experimental results on DUT Anti-UAV, TIB-Net, and VisDrone2019 demonstrate the effectiveness of the proposed modules and their applicability under different scene conditions. Compared with the baseline network, DCP-YOLO achieves higher detection accuracy with fewer parameters and lower theoretical computational complexity. Edge deployment experiments further demonstrate that the proposed method can maintain real-time inference capability on a resource-constrained platform, despite the additional computational overhead associated with its feature reconstruction operations. Nevertheless, the detection of extremely small targets under severe occlusion, adverse weather, and low-light conditions remains challenging. In addition, the current network primarily relies on single-frame visual information and does not explicitly exploit temporal relationships between consecutive frames. Future work will investigate the integration of target tracking, temporal feature modeling, and multimodal perception, together with further optimization of computationally intensive operations, to improve detection robustness and deployment efficiency in complex low-altitude environments.
Available online , doi: 10.11999/JEIT260998
Abstract:
Objective Integrated sensing and communication (ISAC) is an important technology for future wireless networks because it enables communication and sensing functions to share spectrum, hardware, and signal processing resources. Orthogonal frequency division multiplexing (OFDM) is a promising ISAC waveform due to its flexible spectrum allocation and compatibility with existing communication systems. However, conventional OFDM suffers from a high peak-to-average power ratio (PAPR), which reduces power amplifier efficiency under peak-power-limited transmission. Although constant-modulus OFDM (CM-OFDM) can avoid power back-off, existing iterative CM-OFDM methods may produce severe frequency-domain power fluctuations and deep spectral notches. In element-wise-cancellation-based OFDM radar processing, these notches amplify the equivalent noise and degrade the range-velocity map. To address this problem, a generalized Boolean function (GBF)-based CM-OFDM waveform design method is proposed for sensing-prioritized ISAC scenarios with low communication payloads. Methods The proposed method constructs a pair of Golay complementary sequences from a special quadratic GBF. The quadratic term determines the complementary structure, while the affine term embeds communication information. The two complementary sequences are transformed into OFDM time-domain branches by inverse fast Fourier transform. Then, a sample-wise phase correction factor is introduced according to the phase difference between the two branches, and the corrected branches are orthogonally superposed to generate a strict constant-modulus OFDM signal. Two consecutive OFDM time slots with opposite signs in the second branch are transmitted to support receiver-side component separation. For sensing, element-wise cancellation and two-dimensional discrete Fourier transform are used to obtain the range-velocity map. Spectral flatness measure, deep notch rate, harmonic mean, and SNR efficiency are introduced to analyze noise enhancement caused by non-flat transmit spectra. A thresholded element-wise cancellation model is also considered to avoid direct division on low-power subcarriers. For communication, the receiver adds the paired time slots, removes the known quadratic term, and detects the affine information by fast Walsh-Hadamard transform in the BPSK case. Results and Discussions Simulation results show that the proposed GBF-CM-OFDM method achieves 0 dB PAPR, the same as iterative CM-OFDM, but provides a much flatter final transmit spectrum. When the normalized power threshold is 0.1, the deep notch rate of iterative CM-OFDM is mainly around 0.20, while that of the proposed method is about 0.01–0.03. The spectral flatness measure of the proposed method is mainly above 0.80, whereas that of the iterative method is around 0.3–0.4. Under naive element-wise cancellation, the average SNR efficiency is improved from 0.00732 for iterative CM-OFDM to 0.80604 for GBF-BPSK, and the average effective SNR is improved from –11.35 dB to 9.06 dB. With a −20 dB thresholded cancellation model, the iterative method is improved by discarding low-power subcarriers, but GBF-BPSK still achieves higher sensing performance, increasing the average effective SNR from 0.51 dB to 9.06 dB. Multi-target range-velocity maps further demonstrate that the proposed waveform has lower background noise and clearer target peaks. For communication performance, when \(N=256\), GBF-BPSK and GBF-QPSK carry 9 bits and 18 bits per paired waveform, respectively. The bit error rate results show that the proposed method has better recovery stability than iterative CM-OFDM under the considered low-communication-payload setting. Conclusions This paper proposes a GBF-based CM-OFDM waveform design method for ISAC systems. By using GBF-Golay complementary sequence construction, phase correction, and orthogonal branch superposition, the proposed method achieves strict 0 dB PAPR and supports communication data embedding. Compared with iterative CM-OFDM, the proposed method has better spectral flatness, fewer deep notches, and higher SNR efficiency after element-wise cancellation, thereby reducing noise enhancement and improving sensing performance under peak-power-limited conditions. The proposed method is more suitable for sensing-prioritized ISAC scenarios in which the communication function is used to convey a limited amount of auxiliary information, such as control, identification, synchronization, or status information, rather than to support high-throughput data transmission. Future work will investigate additional structural degrees of freedom in the GBF construction, such as variable permutations and complementary-sequence construction rules, to explore additional information embedding while preserving constant modulus and spectral flatness.
Available online , doi: 10.11999/JEIT251093
Abstract:
Objective In Magnetic Resonance Imaging (MRI), undersampling of k-space data is an effective approach to accelerating image acquisition. Reconstructing high-quality images from undersampled data is clinically important for improving diagnostic efficiency. Dual-domain reconstruction methods that jointly exploit spatial- and frequency-domain features are now widely used. However, existing dual-domain MRI reconstruction methods lack differentiated feature extraction strategies for the two domains and sufficient physical prior constraints, leading to the loss of original information during reconstruction. Methods To address these issues, this study proposes a Dual-Domain Differentiated Feature Extraction Network (DDF-Net) for MRI reconstruction. In the spatial domain, an interlaced row-column self-attention mechanism is combined with depthwise convolution to accurately capture anisotropic structures and texture details, addressing the limitations of conventional convolutional feature extraction. In the frequency domain, amplitude and phase characteristics are modeled independently to capture intensity and spatial location information, respectively. A frequency-domain feature enhancement module is used to fully exploit spectral information and improve reconstruction fidelity. Finally, a data consistency layer and a cross-domain adjustment module are incorporated to integrate physical priors and cross-domain information, reinforce measurement consistency, and stabilize reconstruction. Results and Discussions The proposed DDF-Net is evaluated on two publicly available MRI datasets, CC359 and IXI, and compared with six representative reconstruction algorithms, including DAGAN, KIKI-Net, MD-Recon-Net, SwinMR, Reconmer, and KTMR. As shown in Fig. 6 and Fig. 7 , DDF-Net achieves the most faithful anatomical reconstruction under a one-dimensional Gaussian undersampling mask with a 30% sampling rate. It produces clearer cortical edges and finer texture details and effectively suppresses aliasing artifacts. The error maps exhibit more uniform residual distributions, indicating better consistency between the reconstructed and reference images. Quantitative comparisons (Table 1 and Table 2 ) show that DDF-Net attains the highest Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) values across all sampling patterns. Specifically, on the CC359 and IXI datasets, respectively, the improvements over the second-best method are 0.17 dB and 0.23 dB in PSNR and 0.0037 and 0.0029 in SSIM, yielding the best overall performance. These results demonstrate that differentiated feature extraction in the spatial and frequency domains enables DDF-Net to use complementary information from both domains for more accurate reconstruction. To further assess robustness, a noise experiment is conducted following the procedure used for SwinMR. Gaussian noise of varying intensities is added to the k-space data to vary the Noise Level (NL). As illustrated in Fig. 8 and summarized in Table 3 , DDF-Net is more robust to noise perturbations than SwinMR, maintaining higher PSNR and SSIM values under all tested noise conditions. Even at NL = 80%, DDF-Net effectively preserves most structural details, confirming that amplitude-phase separation and multiple data consistency layers jointly improve noise robustness and reconstruction stability. Conclusions This study proposes DDF-Net, an end-to-end network that addresses the limitations of existing dual-domain MRI reconstruction methods. These methods often lack differentiated feature extraction strategies for the two domains and sufficient physical prior constraints, which can lead to information loss during reconstruction. In the spatial domain, an interlaced row-column self-attention mechanism is designed to more effectively model local structures and texture details. In the frequency domain, amplitude-phase separation and a frequency-domain feature enhancement module improve the decoupling and utilization of spectral information. Furthermore, integrating a cross-domain adjustment module with multiple data consistency layers enhances cross-domain interaction and measurement consistency. Experimental results on the CC359 and IXI datasets demonstrate that DDF-Net achieves superior quantitative performance and visual reconstruction quality, recovering fine anatomical details. In addition, noise experiments confirm its robustness under complex conditions, showing that the network effectively preserves image details across different noise levels. Future work will focus on extending DDF-Net to unsupervised or semi-supervised learning frameworks to reduce dependence on labeled data and further enhance its potential for clinical applications in fast MRI reconstruction.
Available online , doi: 10.11999/JEIT260089
Abstract:
Objective The rapid growth of the Internet of Things (IoT) has enabled many emerging applications with stringent latency requirements. In Multi-access Edge Computing (MEC), vertical task offloading from terminal devices to an Edge Node (EN) is a key approach to meeting these requirements. However, each EN serves a limited coverage area, preventing terminal devices outside that area from readily accessing its computing resources and resulting in load imbalances across ENs. Existing vertical task offloading methods primarily optimize terminal-side metrics, such as latency and energy consumption, while giving insufficient attention to load balancing among ENs. Although Multi-Agent Deep Deterministic Policy Gradient (MADDPG) is widely used for vertical task offloading, it is better suited to environments involving both cooperation and competition, whereas horizontal task offloading among ENs constitutes a fully cooperative scenario. Furthermore, conventional vertical task offloading strategies based on Internet Protocol (IP) networks face challenges related to partial observability and limited global information because of the distributed architecture of IP infrastructure. To address these limitations, Software-Defined Networking (SDN) is introduced to provide a global network view and mitigate information asymmetry. A task offloading strategy based on Collaborative Multi-Agent Proximal Policy Optimization (Co-MAPPO) is proposed to support horizontal task offloading and improve load balancing across ENs. Methods An SDN-assisted MEC framework for horizontal task offloading is established, comprising a multi-hop routing communication architecture and a multi-timeslot computation framework. The SDN controller assigns a virtual agent to each EN. A load index is defined as the task data volume per unit frequency of the Central Processing Unit (CPU) for each EN at each timeslot, expressed in Mb/GHz. The variance of the EN load indices is used as the load-balancing index, with a smaller variance indicating more balanced loads across ENs. A multi-objective optimization problem that jointly minimizes system-wide task latency and the load-balancing index is formulated as a stochastic integer programming problem. Subsequently, the offloading process is formulated as a Partially Observable Markov Decision Process (POMDP), and a Co-MAPPO-based task offloading strategy is developed. Because cooperative agents share a common state-value function, a shared critic network is used to estimate the global state-value function. To accommodate agent-specific observations and actions, a shared actor network maps each agent’s encoded local observation to an action probability distribution. An observation one-hot encoding scheme places each agent’s local observation in an agent-specific segment of a zero-padded vector, allowing the shared actor network to distinguish among ENs. During training, the global system state is concatenated with each agent’s local observation and supplied to the shared critic network. This state augmentation enables the critic to account for agent-specific information and mitigates the effects of partial observability on training stability. The SDN controller aggregates information on task data volumes, transmission-queue delays, computation-queue delays, network topology, and link bandwidths to construct the global system state. It computes agent-specific actor and critic losses, aggregates the local losses into global losses, and updates the shared network parameters online. These mechanisms iteratively refine the offloading policy to reduce latency while balancing computational loads across ENs. Results and Discussions The Co-MAPPO strategy exhibits superior convergence performance and outperforms the comparison strategies under different task data volumes and network conditions. (1) Compared with QMIX and the baseline strategies, Co-MAPPO converges faster, exhibits more stable policy updates, and achieves better converged performance. Its clipped proximal policy optimization mechanism limits the magnitude of policy updates and helps stabilize training (Fig. 3 ). Co-MAPPO also enables horizontal collaboration by offloading computationally intensive tasks from ENs with lower computing capacity to those with higher capacity (Fig. 4 ). (2) As task data volume increases, Co-MAPPO maintains lower average latency than the comparison strategies and exhibits the smallest increase in latency. Under the highest task loads, it reduces latency by 40.3%–84.3% relative to the other strategies. Its load-balancing index also remains lower than those of the comparison methods, without abrupt increases under heavy workloads (Fig. 5 ). (3) As transmission bandwidth decreases, system latency increases, but Co-MAPPO retains a clear performance advantage. Under the lowest-bandwidth conditions, it reduces latency by 62.2%–88.5% relative to the other strategies. Its load-balancing index increases gradually as bandwidth becomes constrained but remains lower than those of the comparison methods (Fig. 6 ). (4) When tasks are offloaded from ENs with lower computing capacity to those with higher capacity, both latency and the load-balancing index decrease, indicating a positive correlation between these two metrics. In contrast, when the policy prioritizes load balancing without adequately accounting for transmission bandwidth and network topology, communication latency can increase even as the load-balancing index decreases, indicating a negative correlation between the two metrics (Figs. 7 and 8 ). Conclusions The SDN-assisted Co-MAPPO strategy uses a global network view to adapt to changes in network topology and maintain low latency while balancing EN loads under different task data volumes and transmission bandwidths. The results further show that the relationship between system latency and the load-balancing index depends on offloading conditions. (1) Positive-correlation phase: Offloading tasks from ENs with lower computing capacity to those with higher capacity reduces both latency and the load-balancing index, indicating a positive correlation between the two metrics. (2) Negative-correlation phase: Prioritizing load balancing without adequately accounting for transmission bandwidth limitations and network topology can increase communication latency even when the load-balancing index decreases, indicating a negative correlation between the two metrics. These findings indicate that effective horizontal task offloading requires joint consideration of computing capacity, network conditions, and load distribution.
Available online , doi: 10.11999/JEIT260073
Abstract:
Objective Large Language Models (LLMs) can assist with Register Transfer Level (RTL) code generation. However, processor-level designs require coordination among independently generated control, datapath, top-level integration, and verification modules. Inconsistencies in port names, directions, widths, or signal types can prevent compilation and require manual diagnosis for repair. This study proposes V-Agent, a protocol-driven multi-agent framework that converts implicit interface assumptions into explicit, machine-checkable contracts and returns tool-generated diagnostics to the agent responsible for the affected module. Four specialized agents handle architecture specification, control logic, datapath integration, and verification. A single-cycle RISC-V processor integrating a predefined Intellectual Property (IP) core for 32-bit signed combinational multiplication serves as a controlled benchmark task for evaluating multi-module generation, localized error correction, and differences among LLM backends. The framework addresses the gap between isolated code generation and executable processor-level verification by exposing structural assumptions early enough to support bounded regeneration of individual modules. Methods V-Agent receives a hierarchical JavaScript Object Notation (JSON) configuration specifying the selected LLM backend, generation and retry parameters, architectural attributes, task operations, data widths, and test cases. The Architect Agent generates the top-level ports, control signals, datapath components, and opcode mappings. The Control Unit Agent produces synthesizable instruction-decoding and control logic. The Datapath Agent generates arithmetic, register-file, program-counter, routing, and top-level integration logic and instantiates the predefined multiplier IP core. The Verification Agent generates memory models, testbenches, Verilator drivers, and execution scripts. The processor and accelerator organization is shown in Fig. 1, and the multi-agent and verification feedback loops are summarized in Figs. 2 and 3 . JSON interface contracts record port names, directions, widths, and types. A static analyzer parses the Verilog Abstract Syntax Tree (AST) and compares the generated port signatures with the corresponding contracts. Compilation and simulation provide additional diagnostics on syntax, connectivity, and functional behavior. When a check fails, the resulting diagnostic is returned to the responsible agent to trigger regeneration. The maximum number of retries is set to 3 (max_retries = 3). A fallback template is used only when repeated generation reaches the retry limit, allowing the remaining integration and verification steps to continue. The evaluated design supports two standard instructions and three experimental custom operations: MLOAD, PMUL, and PMAC. Six LLM backends, DeepSeek-V3, Claude Sonnet 4.5, Doubao-1.5 Pro, Gemini 1.5 Flash, Qwen3-235B, and GPT-5.1, are compared using module-compilation results, automatic-test performance, Code Quality Score, line-level code changes during manual repair, and Power, Performance, and Area (PPA) metrics from downstream implementation analysis. Fully automated results are recorded before manual modification, whereas Human-in-the-Loop (HITL) results are measured after localized repair. Results and Discussions The outputs of the monolithic, one-shot baseline fail system-level compilation because of context, interface, connection, or incomplete-state-logic problems. By contrast, the V-Agent pipeline produces designs that can proceed to system-level compilation (Table 4 ). This finding is limited to the present benchmark. In the fully automated evaluation, DeepSeek-V3 and Claude Sonnet 4.5 complete all five task operations, compile all eight RTL modules, and pass 70 of 78 automatic tests (Table 6 ). Doubao-1.5 Pro and Gemini 1.5 Flash compile 6 of 8 modules, GPT-5.1 compiles 7 of 8, and Qwen3-235B compiles 4 of 8. The generated designs support different numbers of executable automatic tests. Therefore, the results retain the raw counts and denominators rather than treating all test totals as directly comparable. Figure 4 presents the task-operation completion and automatic-test pass ratios. An interface-width mismatch illustrates how the feedback mechanism supports localized correction. The Datapath Agent initially defines MemToReg as a one-bit signal, while the additional multiplier-result path requires a two-bit selector. The contract checker identifies the mismatch and returns the diagnostic to the responsible agent. The agent regenerates the affected module in approximately 15 s, after which compilation and testing pass (Table 7 ). In the complete case study, the four agents generate eight RTL modules and 17 verification files in about 5 min. All eight modules compile, and four system-level programs produce the expected final register states and waveforms. The generated arithmetic execution and writeback path is shown in Fig. 5 . After localized HITL edits, all six designs compile all eight RTL modules and pass the same set of 78 tests. The average line-level code-change rate is 4.37%, ranging from 1.3% to 9.5% (Table 9 ). PPA analysis shows that Claude Sonnet 4.5, DeepSeek-V3, and Gemini 1.5 Flash complete synthesis and timing analysis. GPT-5.1 completes gate-level mapping, but its timing analysis fails to converge because of a combinational loop. The unusually small implementations produced by Doubao-1.5 Pro and Qwen3-235B exhibit large-scale logic pruning during synthesis (Table 8 ). Among 126 classified errors at the generation and compilation stages, 54 are Verilog semantic errors, 36 are interface-width errors, 26 are cascading failures, and 10 are custom-operation semantic errors (Table 10 ). Conclusions V-Agent structures processor RTL generation as a coordinated workflow among specialized agents connected by explicit interface contracts and tool feedback. The evaluated RISC-V arithmetic-IP task shows that structural interface mismatches can be localized and corrected without regenerating the entire design. HITL results further distinguish initial generation quality from engineering repairability. The principal contribution is a coordination mechanism that converts cross-module assumptions into machine-checkable contracts and returns AST, compilation, and simulation diagnostics to the responsible agent. The framework complements existing synthesis, simulation, high-level synthesis, and verification tools rather than replacing them. The evidence is limited to a single-cycle RISC-V core, a predefined 32-bit multiplier IP core, task configurations containing 5–15 operations, and nonuniform executable test workloads. Verilator cannot expose every timing defect, and the PPA study requires more complete activity information, supply conditions, and netlist checks. Future work will extend the workflow to pipelined processors and coverage-guided automated repair.
Available online , doi: 10.11999/JEIT260326
Abstract:
Objective Radio Frequency (RF) switches are critical front-end components in communication and radar systems, especially in active phased array Transmit/Receive (T/R) modules, where their insertion loss, isolation, and power-handling capability directly affect system performance. Positive-Intrinsic-Negative (PIN) diodes are widely used in high-power microwave switches because of their high power-handling capability, good linearity, and mature fabrication process. However, low-cost plastic-packaged PIN diodes suffer from considerable parasitic effects at X-band frequencies. The junction capacitance, package inductance, and grounding-via inductance introduce deviations from ideal short- or open-circuit conditions in shunt branches, resulting in increased signal leakage, degraded isolation, and higher insertion loss. Therefore, this paper proposes a low-cost high-isolation X-band Single-Pole Double-Throw (SPDT) RF switch design method based on parasitic-parameter compensation to improve the high-frequency performance of low-cost PIN diode switches. Methods A step-wise synergistic matching strategy, termed “OFF-state prioritized determination and ON-state structural compensation”, is proposed. OFF-state isolation and ON-state transmission conditions are treated separately to reduce coupling between isolation and transmission objectives. First, the OFF-state isolation condition is determined by combining the diode parasitic parameters, compensation stub, and grounding-via inductance to construct a series-resonant network. By compensating the parasitic reactance of the isolation branch, the required impedance condition is satisfied, thereby suppressing signal leakage through the isolated path. After determining the compensation-stub parameters, ON-state structural compensation is further implemented by introducing equivalent capacitance through T-junction edge effects and a stepped-impedance main line. The introduced structural capacitance compensates for the equivalent inductance of the shunt branch and forms a parallel-resonant network, allowing the branch impedance to approach an open-circuit condition at the operating frequency. Thus, signal shunting is reduced and low-insertion-loss transmission is achieved without introducing additional lumped compensation components. Results and Discussions The proposed switch topology is established, and a full-wave simulation model is developed (Fig. 3, Fig. 4(a)). Simulation results demonstrate that the designed switch achieves a return loss better than 20 dB, an isolation higher than 20 dB, and an insertion loss of approximately 0.2 dB within 8.57～9.34 GHz (Fig. 4(b)). A prototype is fabricated and experimentally characterized (Fig. 5). Measurement results indicate that, within 9.5～10.4 GHz, the return loss is better than 10 dB, the insertion loss is lower than 2 dB, and the isolation exceeds 20 dB (Fig. 6). Although the measured operating band shifts upward by approximately 1 GHz compared with the initial simulation results, the measured variation trends agree well with the theoretical analysis. To investigate the frequency shift, the effects of OFF-state junction capacitance and ON-state parasitic inductance are further analyzed. When the OFF-state junction capacitance decreases from 0.24 pF to 0.195 pF, the high-isolation frequency band shifts toward higher frequencies (Fig. 7(a)). With the OFF-state capacitance fixed at 0.195 pF, reducing the ON-state parasitic inductance from 0.7 nH to 0.2 nH also shifts the return-loss response toward higher frequencies (Fig. 7(b)). After correcting the parasitic parameters, the simulated return loss, insertion loss, and isolation show improved agreement with the measured results (Fig. 6). Compared with existing SPDT RF switch designs, the proposed design employs low-cost plastic-packaged PIN diodes and a simple microstrip compensation structure, eliminates the need for additional lumped compensation components, reduces dependence on the intrinsic high OFF-state impedance of PIN diodes, and provides an effective trade-off between manufacturing cost and high-frequency switch performance (Table 1). Conclusions A low-cost high-isolation X-band SPDT RF switch based on parasitic-parameter compensation is proposed and experimentally verified. The proposed design combines OFF-state series resonance with ON-state structural capacitance compensation to simultaneously achieve high isolation and low insertion loss. The fabricated switch achieves an isolation higher than 20 dB and an insertion loss lower than 2 dB over 9.5～10.4 GHz. Owing to its simple structure, low cost, and ease of integration, the proposed switch provides a practical solution for low-cost high-isolation RF switches for T/R modules in active phased array systems.
Available online , doi: 10.11999/JEIT260709
Abstract:
Objective Three-Dimensional (3D) Synthetic Aperture Radar (SAR) tomography can overcome layover, perspective foreshortening, and shadowing encountered in two-dimensional SAR imaging. It can also resolve multiple scatterers within the same azimuth-range resolution cell in the elevation direction, enabling high-precision 3D imaging of targets. High-resolution and high-precision SAR tomographic imaging requires repeated observations and data accumulation over the same location. Conventional satellite systems must therefore conduct repeated overpasses and frequently adjust their orbits to obtain the required observation geometry, incurring substantial system costs. Even with these measures, temporal decorrelation cannot be eliminated. Distributed satellite formations can obtain simultaneous multi-baseline observations of the same region through coordinated flight of multiple satellites. However, existing distributed satellite formation configurations are not optimized for 3D SAR tomography. This study proposes a formation-design method tailored to tomographic imaging. By appropriately designing satellite orbits, the method establishes a distributed observation array in space and enables high-resolution, high-precision 3D SAR imaging in a single pass. Formation-maintenance strategies are also developed based on the stability of in-orbit configurations, demonstrating the practical engineering value of the proposed method. Methods To meet the elevation-direction array-imaging requirements of 3D SAR tomography, a mathematical model relating the three-dimensional baselines of distributed SAR satellites to their relative motion is established. The model is expressed in the Radial-Transverse-Normal (RTN) coordinate system to derive the elevation-direction synthetic aperture, normal baseline, and along-track baseline required for tomographic imaging. This establishes a direct relationship between the three-dimensional baselines and the six orbital elements of the formation satellites, allowing the desired baseline distribution to be achieved by directly designing their orbital parameters. To meet the linear-array observation requirements, two types of multi-satellite formations are designed: a nested-wheel side-by-side formation with vertically aligned satellites and a transverse side-by-side formation with laterally aligned satellites. The satellite trajectories are designed to maintain a linear array at every instant of the orbital period. The resulting spatial baselines satisfy the array-observation requirements of 3D SAR tomography. Formation-configuration divergence is also analyzed through simulations, and corresponding formation-maintenance strategies are proposed, demonstrating the engineering feasibility of the method. Results and Discussions A nested-wheel side-by-side multi-satellite formation is designed to generate a vertical-aperture baseline, with the normal baseline set to zero and the arguments of latitude of the main and follower satellites set equal ua=uc. The semi-major axes, eccentricities, and arguments of perigee of the satellites are selected according to the required vertical-aperture baseline length. The argument of latitude is then determined based on the required observation position, ensuring that the maximum radial baseline occurs at the desired position. The resulting maximum radial-baseline amplitude is \begin{document}$ a{e}_{a} $\end{document} Figs. 3 and 4 ). To generate a normal baseline, a transverse side-by-side multi-satellite formation is designed with the radial baseline set to zero. The eccentricity of the follower satellite ea is set to zero, and the difference in the right ascension of the ascending node Δ\begin{document}$\varOmega $\end{document} \begin{document}$ |{a}_{c}({\varOmega }_{a}-{\varOmega }_{c})\sin {i}_{c}| $\end{document} Figs. 5 and 6 ). Both configurations provide stable baselines for linear-array observations and meet the requirements of 3D SAR tomography. Using the nested-wheel side-by-side four-satellite formation as an example, configuration divergence under atmospheric drag is analyzed. Because atmospheric drag causes semi-major-axis decay and the resulting along-track drift, a formation-maintenance strategy is proposed based on the difference between the drift of the centers of the relative-motion ellipses of the two satellite groups. When \begin{document}$ \left| {\varDelta }_{1}-{\varDelta }_{2}\right| \leq {L}_{\max } $\end{document} \begin{document}$ \left| {\varDelta }_{1}-{\varDelta }_{2}\right| \geq {L}_{\max } $\end{document} Fig. 12 ). Conclusions The proposed nested-wheel side-by-side four-satellite formation provides a stable vertical-aperture baseline for 3D SAR tomographic imaging. Its formation-maintenance strategy is straightforward to implement and offers good engineering feasibility. Distributed satellite formations enable multiple observations in a single pass, overcoming the temporal-decorrelation limitation of conventional multi-pass observations and supporting high-resolution, high-precision 3D SAR tomography. The nested-wheel side-by-side and transverse side-by-side configurations provide stable baselines for linear-array observations, with the former designed to generate the vertical-aperture baseline and the latter to generate the normal baseline. Simulation results show that the divergence of the nested-wheel side-by-side formation during coplanar flight is primarily caused by semi-major-axis decay. Formation stability can be maintained by controlling the rate of semi-major-axis decay, and the analytical results agree with the formation simulations. Imaging simulations further demonstrate that the proposed four-satellite formation provides stable tomographic baselines, reduces coherence loss caused by along-track separation between formation satellites, and improves 3D imaging performance. These findings provide a technical reference for the design of next-generation 3D SAR tomography systems.
Available online , doi: 10.11999/JEIT260379
Abstract:
The 3rd Generation Partnership Project (3GPP) Technical Specification Group Service and System Aspects Working Group 4 (TSG SA Working Group 4, SA4) is currently investigating Artificial Intelligence (AI)-based speech codecs. Based on the Descript Audio Codec (DAC) under investigation by SA4, this paper proposes three AI-enabled transmission schemes: a Total Power Constrained (TPC) DAC Joint Source-Channel Coding and Modulation (DAC-JSCCM) scheme, a Constant-Envelope Constrained (CEC) DAC-JSCCM scheme, and a DAC Joint Source-Channel Coding (DAC-JSCC) scheme. The proposed schemes are evaluated using the typical physical-layer parameter configuration for 3GPP Geostationary Earth Orbit (GEO) voice communications. The results provide useful guidance for future research on GEO voice communications. Objective Joint Source-Channel Coding (JSCC) has attracted increasing attention in both academia and industry. Academic research has primarily focused on video, image, and audio transmission, whereas industrial research has mainly considered Channel State Information (CSI) as the source. The limited industrial interest in JSCC for video and image transmission is mainly attributed to two factors. First, video and image coding standards are generally developed by external organizations, such as the Moving Picture Experts Group (MPEG) and the Joint Photographic Experts Group (JPEG), rather than by 3GPP. Second, application-layer JSCC is limited by architectural challenges because the channel information available at the application layer is often outdated, reducing its potential gain. Audio differs from video and images because speech coding is developed within 3GPP, and SA4 is currently investigating AI-based speech codecs. In GEO voice communications, satellites remain nearly stationary relative to ground stations, resulting in slowly varying channels. Therefore, physical-layer CSI can be reliably delivered to the application layer, avoiding outdated channel information. These characteristics make JSCC for audio a promising research direction within 3GPP. Methods The typical physical-layer parameter configuration for 3GPP GEO voice communications and the DAC under investigation by SA4 are first adopted as the evaluation baseline to ensure a fair performance comparison. Three AI-enabled transmission schemes are then proposed for low-bitrate audio communications. First, the TPC DAC-JSCCM scheme jointly optimizes source coding, channel coding, and modulation to maximize transmission gain. Second, to reduce the high Peak-to-Average Power Ratio (PAPR) associated with the TPC scheme, the CEC DAC-JSCCM scheme is proposed. Finally, because transmitting modulation symbols within the Real-time Transport Protocol (RTP) payload requires substantial modifications to the existing protocol, the DAC-JSCC scheme is proposed to retain bit-sequence transmission in the RTP payload while remaining compatible with the current protocol architecture. Results and Discussions Baseline scheme 1 is first evaluated through physical-layer link-level simulations to establish the relationship between Signal-to-Noise Ratio (SNR) and Block Error Rate (BLER) for 64-bit packets encoded with a rate-1/3 Turbo code. Assuming that the error rate of an audio frame is equal to the corresponding BLER, the transmission error rate of an RTP packet containing L audio frames is calculated. According to the existing transmission mechanism, an RTP packet is discarded if any audio frame within the packet is received incorrectly. In real-time communication, Voice over Internet Protocol (VoIP), and speech enhancement applications, a Perceptual Evaluation of Speech Quality (PESQ) score of 2.5 is generally regarded as the minimum acceptable speech quality threshold. Therefore, PESQ = 2.5 is selected as the reference point for GEO voice communications. At this threshold, both the TPC DAC-JSCCM and CEC DAC-JSCCM schemes achieve an approximately 4 dB coverage gain over baseline scheme 1 (Fig. 4 ), whereas the DAC-JSCC scheme achieves a 2.5 dB coverage gain. These gains result from application-layer JSCC, which eliminates the cliff effect associated with Turbo coding at the physical layer. At a Complementary Cumulative Distribution Function (CCDF) of 10−3, the Peak-to-Average Power Ratio (PAPR) of the TPC DAC-JSCCM scheme is approximately 4 dB higher than that of Quadrature Phase Shift Keying (QPSK), resulting in sharper signal peaks, stricter linearity requirements for the power amplifier, and greater susceptibility to nonlinear distortion. By contrast, the CEC DAC-JSCCM scheme exhibits a PAPR comparable to that of QPSK. Conclusions Based on the DAC under investigation by SA4, three AI-enabled transmission schemes are proposed. Simulations are performed using the typical physical-layer parameter configuration for 3GPP GEO voice communications. At the minimum acceptable speech quality threshold (PESQ = 2.5), the proposed TPC DAC-JSCCM and CEC DAC-JSCCM schemes each provide an approximately 4 dB coverage gain over the conventional DAC-based transmission scheme, whereas the DAC-JSCC scheme provides a 2.5 dB coverage gain. As SA4 continues to investigate more advanced GEO speech codecs, future work will extend the proposed framework to these next-generation codecs.
Available online , doi: 10.11999/JEIT260523
Abstract:
Objective Practical semantic communication should satisfy two requirements: compatibility with existing digital communication infrastructures and robustness under varying channel conditions. Semantic-Oriented Modulation (SOM) maps continuous semantic features to layered digital constellation symbols, enabling semantic transmission over conventional digital systems. However, the digitization process introduces structured quantization distortion, which increases the difficulty of receiver-side recovery. Although diffusion models provide channel-adaptive semantic recovery, their direct application to SOM-based digital semantic communication is limited by this structured distortion. Therefore, a channel-adaptive receiver is designed for digital semantic communication to compensate for SOM-induced structured distortion before subsequent recovery. Methods An SOM-based digital semantic communication system for image transmission over an Additive White Gaussian Noise (AWGN) channel is considered. The proposed receiver adopts a two-stage structure comprising a Quantization Noise Predictor (QNP) and a diffusion recovery module. In the first stage, QNP estimates and compensates for the structured quantization distortion introduced by SOM from layer-wise soft received symbols. In the second stage, the compensated semantic representation is further refined by a diffusion denoiser, with the number of inference steps adaptively selected according to the estimated Signal-to-Noise Ratio (SNR). The QNP consists of a shared feature-extraction front end, a classification branch that exploits discrete SOM constellation priors, and a regression branch that performs fine-grained compensation for quantization distortion in the continuous domain. A Feature-wise Linear Modulation (FiLM) mechanism is used to incorporate SOM parameters and channel state information, enabling a single QNP to adapt to different modulation configurations and channel conditions. In addition, a composite loss comprising classification, regression, and distribution regularization terms is designed to improve the statistical properties of the compensated residual noise. Results and Discussions Experiments are conducted on the CLIC dataset using Peak Signal-to-Noise Ratio (PSNR) and Multi-scale structural similarity (MS-SSIM) as evaluation metrics. The proposed method is compared with a Variational autoencoder (VAE), VAE+Diff, VAE+SOM, VAE+SOM+QNP, VAE+SOM+Diff, and the Joint coding-modulation (JCM) method. Direct SOM-based digitization causes noticeable performance degradation, whereas the proposed method consistently improves image reconstruction quality over digital semantic baselines. QNP alone improves PSNR by an average of 1.77 dB and MS-SSIM by 26.08% compared with the VAE+SOM baseline under different channel conditions. Moreover, compared with VAE+SOM and VAE+SOM+Diff, the complete method improves PSNR by 3.67 dB and 1.89 dB and MS-SSIM by 58.79% and 18.64%, respectively. These results demonstrate that QNP provides effective compensation independently of diffusion recovery and that its combination with diffusion recovery provides complementary gains. Two training strategies, namely independent QNP training and diffusion-assisted fine-tuning, are further compared. Diffusion-assisted fine-tuning improves PSNR and MS-SSIM by only 0.83 dB and 3.30% on average, respectively, but substantially increases training complexity and resource consumption. Therefore, independent training is adopted for the main experiments. Under different SOM configurations, the proposed method maintains stable gains across different modulation orders and numbers of SOM layers. Latency analysis shows that QNP requires only 4.127 ms for one forward pass, whereas one diffusion step requires approximately 85.746 ms. Thus, inference latency is mainly determined by diffusion iterations, and the diffusion-step schedule is selected according to the trade-off between reconstruction quality and latency. Finally, Gaussianity analysis based on Kullback-Leibler divergence and Wasserstein distance shows that QNP compensation makes the residual noise distribution more consistent with a Gaussian distribution. Without distribution regularization, the two distances decrease by 97.20% and 89.26% on average, respectively, compared with the uncompensated case. With distribution regularization, the corresponding reductions reach 99.26% and 98.47%, respectively. Conclusions A channel-adaptive receiver for digital semantic communication is proposed by combining QNP-based front-end compensation with diffusion-based semantic recovery. The QNP provides lightweight compensation for SOM-induced structured quantization distortion before subsequent recovery and remains effective when used independently. Experimental results show that stable digital semantic recovery gains are obtained under different SNR conditions and SOM configurations. When combined with diffusion recovery, the proposed method achieves the best overall performance. The proposed receiver therefore provides an effective approach for improving semantic recovery quality and channel adaptability while maintaining compatibility with existing digital communication infrastructures.
Available online , doi: 10.11999/JEIT260408
Abstract:
Objective With the continued expansion of satellite constellations and the emergence of new services, satellite computing is expected to substantially improve the service capabilities of large-scale satellite networks. However, limited onboard computing resources make it challenging to determine how many computing nodes to deploy and where to place them. In highly dynamic satellite networks, topology changes affect satellite connectivity, computing-resource distribution, service coverage, and the costs associated with deployment and reconfiguration. An effective deployment strategy therefore requires a graph-based model that captures these coupled effects. To address this problem, a method combining a Graph Neural Network (GNN) and Proximal Policy Optimization (PPO), termed GNN-PPO, is proposed. The GNN encodes neighborhood information to represent inter-satellite connectivity and the spatial distribution of computing resources. Based on this representation, the method determines the number of computing nodes to deploy and selects their locations, providing a basis for computing-resource deployment in future satellite constellations. Methods The GNN-PPO framework uses the GNN to encode the dynamic states of satellite nodes and Inter-Satellite Links (ISLs). The input state includes satellite topology, available computing resources, task-related computing demand, node energy and load, candidate-satellite status, computing-node deployment status, and the ratio of deployed computing nodes to the total number of satellites. Link features include bandwidth and link availability. Through message passing, the GNN aggregates neighborhood topology and resource information into a fixed-length graph representation that serves as input to the PPO agent. The agent uses hierarchical actions: the upper-level action determines whether to increase, decrease, or maintain the number of computing nodes, whereas the lower-level action determines the magnitude of the change. After the target number of computing nodes is determined, a heuristic procedure selects their deployment locations based on candidate-satellite degree, remaining computing resources, and degree centrality. The reward function combines a hop-count penalty, energy penalties for reconfiguration and transmission, a computing-node-count penalty, and positive rewards for computing-service coverage and coverage efficiency. This design enables the agent to balance service performance against deployment and reconfiguration costs. Results and Discussions The proposed GNN-PPO method converges more effectively than the standalone PPO, Deep Q-Network (DQN), and Advantage Actor-Critic (A2C) baselines (Fig. 3 ). At a network scale of 700 satellites, the method reduces the number of deployed computing nodes by approximately 29% compared with Particle Swarm Optimization (PSO) and lowers reconfiguration energy consumption by approximately 35.8% compared with DQN (Fig. 5 ). Across the tested hop constraints (h=1, 2, and 3) and Computing Structure Graph (CSG) types, the average management hop count ranges from approximately 2.5 to 3.3 hops, while the average reconfiguration energy consumption ranges from 0.9 to 1.2 J (Table 4 ). The average transmission energy consumption of the tree CSG is approximately 15 times that of the star and fully connected CSGs. These results indicate that star and fully connected CSGs offer more favorable trade-offs between deployment costs and management performance in the simulated large-scale satellite networks. Conclusions The proposed GNN-PPO method converges more effectively than the standalone PPO, Deep Q-Network (DQN), and Advantage Actor-Critic (A2C) baselines (Fig. 3 ). At a network scale of 700 satellites, the method reduces the number of deployed computing nodes by approximately 29% compared with Particle Swarm Optimization (PSO) and lowers reconfiguration energy consumption by approximately 35.8% compared with DQN (Fig. 5 ). Across the tested hop constraints (h=1, 2, and 3) and Computing Structure Graph (CSG) types, the average management hop count ranges from approximately 2.5 to 3.3 hops, while the average reconfiguration energy consumption ranges from 0.9 to 1.2 J (Table 4 ). The average transmission energy consumption of the tree CSG is approximately 15 times that of the star and fully connected CSGs. These results indicate that star and fully connected CSGs offer more favorable trade-offs between deployment costs and management performance in the simulated large-scale satellite networks
Available online , doi: 10.11999/JEIT260221
Abstract:
Objective With the rapid development and extensive application of Unmanned Aerial Vehicle (UAV) observation platforms, remote sensing video data with high spatiotemporal resolution are increasing rapidly. Dynamic relation understanding in UAV videos is a challenging task in intelligent remote sensing analysis. Current research on video scene graph generation focuses mainly on natural videos, whereas studies on remote sensing videos remain at an early stage. Benchmark datasets with high-order dynamic relation annotations are also lacking. Conventional visual models are difficult to apply directly to remote sensing scenes, which are characterized by substantial viewpoint changes, large variations in object scale, and dense object distributions. To address these limitations, a UAV Video Relation Dataset (UAVREL) with high-order dynamic relation annotations is constructed. A spatiotemporal hypergraph-enhanced relation understanding method tailored to UAV videos is also proposed. In addition, a unified benchmark evaluation system for video scene graph generation in UAV remote sensing is established. These efforts facilitate the transition of intelligent remote sensing interpretation from static object recognition to global dynamic scene understanding. Methods This study consists of three main components: dataset construction, model design, and comprehensive experimental evaluation. First, the Unmanned Aerial Vehicle Benchmark: Object Detection and Tracking (UAVDT) dataset is selected as the basic data source. Fifty representative UAV videos are finely annotated with 12 object categories, resulting in 876,687 object bounding boxes. Twenty relation categories are further annotated based on the object bounding boxes, resulting in 459,164 relation triplets. A semiautomatic annotation strategy integrating object tracking, multi-person collaboration, and cross-checking is developed to reduce subjective bias and improve annotation efficiency. Relations are divided into three levels according to the number of frames required for accurate determination, with high-order dynamic relations prioritized during annotation. Second, a spatiotemporal hypergraph-enhanced relation understanding method (STHG) is proposed. The method comprises six core modules: basic feature extraction, pairwise context encoding, spatiotemporal dual-hypergraph enhancement, multi-source feature fusion, global temporal modeling, and relation prediction. A multi-label classifier is then used for final relation prediction. In the experimental evaluation, an object detection task and three video scene graph generation subtasks, namely Predicate Classification (PredCls), Scene Graph Classification (SGCls), and Scene Graph Detection (SGDet), are conducted on the UAVREL dataset. Representative object detection models and video scene graph generation methods are selected as baselines. Mean Average Precision (mAP) and mAP@50 are used to evaluate object detection, whereas Recall@K (R@K) and mean Recall@K (mR@K) are used to evaluate video scene graph generation. Results and Discussions In the object detection task, eight models with different architectures are compared to evaluate their detection performance (Table 2 ). The results show that model architecture affects mAP, a core detection metric. Single-stage anchor-free models represented by VFNet and DDOD achieve the best overall detection performance. All models perform poorly on small and easily occluded object categories, such as pedestrians and non-motorized vehicles. In the video scene graph generation task, five methods are evaluated on three subtasks (Tables 3 –5 ). STHG achieves improved performance in SGDet, with R@20, R@50, and R@100 increasing by 2.2%, 1.2%, and 0.4%, respectively, compared with the STTran baseline. In PredCls, STHG performs comparably to STTran, whereas its mR@100 in SGCls is higher than that of STTran. In SGDet, STHG improves mR@20, mR@50, and mR@100 by 4.6%, 4.2%, and 2.2%, respectively, over STTran. The mR@K values are consistently lower than the corresponding R@K values, indicating substantial challenges in modeling long-tailed relations. These results also demonstrate the difficulty of relation prediction in complex UAV video scenes. Conclusions This study focuses on dynamic relation understanding in UAV videos. The UAVREL benchmark dataset is constructed to provide data support for high-order dynamic relation understanding in UAV videos. A spatiotemporal hypergraph-enhanced relation understanding method is also proposed. Experimental results demonstrate its effectiveness across multiple evaluation metrics, particularly for SGDet and long-tailed relation recognition. The dataset and method provide a basis for advancing remote sensing video analysis from static object recognition to high-level dynamic scene understanding. Future research will focus on two directions: improving temporal modeling for dynamic remote sensing scene graph generation to capture long-term evolution and complex relations, and expanding the scene coverage and diversity of relation categories in the dataset to support further development and application of scene graph generation methods.
Available online , doi: 10.11999/JEIT260912
Abstract:
Objective With the rapid development of marine transportation, ship-radiated noise recognition has become essential for maritime safety. However, existing deep learning methods rely on large-scale labeled training data. In underwater acoustic scenarios, ship-radiated noise data are difficult to collect because hydrophone deployment and sea trials are difficult to control. Under this condition, conventional supervised recognition methods tend to suffer from performance degradation. Although contrastive learning can exploit unlabeled data, generic augmentation strategies may destroy critical spectral structures. In addition, conventional fine-tuning fails to fully use the similarity structure among ships. To address these challenges, a two-stage recognition method based on contrastive pre-training and progressive self-distillation is proposed. This research aims to improve the utilization of unlabeled data and limited labeled data and enhance the accuracy of ship-radiated noise recognition. Methods The ship-radiated noise signals are processed by the Constant-Q Transform (CQT) to obtain CQT spectra, which are then used as the input. The proposed method is carried out through contrastive pre-training and progressive self-distillation. In the contrastive pre-training stage, each unlabeled sample is transformed into two augmented views by a combined data augmentation strategy. This strategy includes time-axis flipping, random signal truncation and additive noise injection. These augmentations reflect variations in observation order, observation windows, and background interference. Time-axis flipping preserves the frequency positions of spectral lines, while signal truncation retains part of the waveform and restores its length through zero padding. These operations introduce view diversity while seeking to preserve structures relevant to ship recognition. The augmented views generated from the same sample are treated as a positive pair, and the views generated from other samples in the same mini-batch are treated as negative samples. An encoder and a projection head are used to map the CQT spectra into an embedding space. The contrastive loss is used to increase the similarity of positive pairs and decrease the similarity between negative ones. After contrastive pre-training, the projection head is discarded, and the encoder parameters are retained for the following recognition stage. In the progressive self-distillation stage, the pre-trained encoder is connected with a classifier to form the recognition model. The shallow layers of the encoder are frozen to preserve general acoustic features, while the deeper layers and the classifier are updated using limited labeled samples. The training objective includes the cross-entropy loss and the soft-label distillation loss. The soft labels are generated from the historical predictions of the model itself, and the weight of the distillation loss increases linearly with the training epoch. This progressive weighting strategy reduces the influence of unreliable soft labels in the early training stage and gradually introduces similarity information contained in historical predictions. Results and Discussions The proposed two-stage method is evaluated on the VTUAD dataset under five labeled-data scales, including 100%, 10%, 5%, 2%, and 1% labeled samples (Table 1 ). The supervised baseline achieves an accuracy of 94.28% and an F1-score of 91.90% when all labeled samples are used. However, its performance decreases as the labeling scale is reduced, and the accuracy and F1-score drop to 72.60% and 58.08%, respectively, when only 1% labeled training samples are used (Table 3 ). After contrastive pre-training, standard fine-tuning improves the recognition performance over the supervised baseline. By introducing progressive self-distillation, the proposed method further improves the recognition performance (Table 3 ). Compared with CRNN, VGG, CNN-Transformer, and ECAPA-TDNN, the proposed method achieves the best recognition performance in all five settings. In the 1% labeled-data case, it improves the accuracy by 11.33% and the F1-score by 15.08% compared with the second-best method (Table 4 ). The comparison of different self-distillation strategies shows that progressive self-distillation performs better, indicating that the progressive weighting strategy can reduce the influence of noisy soft labels at the early training stage and make better use of category similarity information (Table 6 ). The shallow-layer freezing experiment shows that freezing the first two layers of the encoder achieves the best balance between preserving pre-trained features and adapting to the recognition task (Fig. 8 ). Noise robustness is further evaluated by adding Gaussian white noise to the test recordings under the 1% labeled training data setting. At a signal-to-noise ratio of −5 dB, the proposed method achieves 62.18% accuracy and a 43.82% F1-score, compared with 45.00% and 37.68%, respectively, for the supervised baseline (Table 7 ). The confusion matrices and feature visualizations show that the proposed method improves the recognition performance and produces more compact intra-class distributions and clearer inter-class separation (Fig. 9 , Fig. 10 ). A cross-dataset transfer experiment is further conducted on the ShipsEar dataset. In this experiment, the encoder pre-trained on VTUAD is adapted using labeled ShipsEar training samples. The 1% training subset contains only 40 samples. Across the five labeled-data scales, the proposed method improves accuracy and F1-score over the supervised baseline by averages of 5.27% and 5.51%, respectively. Although VTUAD and ShipsEar differ in data distribution, and ShipsEar contains fewer labeled training samples, the proposed method outperforms the supervised baseline across different labeled-data scales, demonstrating cross-dataset transferability (Table 8 ). Conclusions This paper addresses the performance degradation of ship-radiated noise recognition with limited labeled data by proposing a two-stage method based on contrastive pre-training and progressive self-distillation. The proposed method improves the utilization of unlabeled and limited labeled data in three aspects: (1) A two-stage recognition framework is constructed, in which contrastive pre-training is used to learn general acoustic representations from unlabeled ship-radiated noise samples, and self-distillation is used to optimize the recognition model with limited labeled samples; (2) A combined data augmentation strategy is designed for contrastive pre-training to introduce different views; (3) A progressive self-distillation mechanism is introduced in the recognition stage, where the historical predictions are used as soft labels and are gradually incorporated into training to supplement hard-label supervision. The method is evaluated on the VTUAD dataset, and the experimental results show that it achieves average improvements of 6.17% in accuracy and 8.11% in F1-score over the supervised baseline across five labeled-data scales. The transfer experiment on ShipsEar further demonstrates that the proposed method remains superior to the supervised baseline when the source and target datasets differ in data distribution and the target dataset contains fewer labeled training samples.
Available online , doi: 10.11999/JEIT260984
Abstract:
Objective Few-weight linear codes have attracted considerable attention in coding theory due to their applications in secret sharing schemes, authentication codes, association schemes, and strongly regular graphs. In recent years, significant progress has been made in the construction of few-weight linear codes from defining sets, leading to some classes of linear codes with good parameters. Most existing constructions are based on specific types of defining sets or particular algebraic structures. Therefore, it is of interest to develop more general defining set constructions and establish corresponding criteria for code optimality. Motivated by this, we investigate linear codes over finite fields constructed from defining sets given by the complements of unions of subgroup cosets. As a consequence, several classes of few-weight linear codes are obtained, and explicit computable criteria are established for determining their optimality with respect to the Griesmer bound. Methods The main technique of this paper is a unified defining set construction based on the complements of unions of additive or multiplicative subgroup cosets. Linear codes are constructed within the Ding–Niederreiter framework, and character sum theory is employed to investigate their parameters. For additive subgroups, the orthogonality relations of additive characters are used to derive explicit expressions for the Hamming weights of codewords. For multiplicative subgroups, multiplicative characters and Gauss sums are applied to characterize the Hamming weights and determine the parameters of linear codes. By specializing the subgroup structures and exploiting explicit evaluations of character sums, several classes of linear codes with few weights are obtained. Furthermore, the Griesmer bound is employed to establish explicit computable criteria for optimality, which lead to the construction of several infinite families of optimal and distance-optimal linear codes. Results and Discussions The motivation of this paper is to search for optimal or distance-optimal linear codes with few weights. The results and discussions can be summarized as follows. Let \begin{document}$ G $\end{document} \begin{document}$ {\mathbb{F}}_{{{q}^{k}}} $\end{document} \begin{document}$ \mathcal{D}={\mathbb{F}}_{{{q}^{k}}}\backslash \cup _{i=1}^{h}({\alpha }_{i}+G) $\end{document} \begin{document}$ {\mathbf{c}}_{x} $\end{document} \begin{document}$ {\mathcal{C}}_{\mathcal{D}} $\end{document} \begin{document}$ G=\{g\in {\mathbb{F}}_{{{q}^{k}}}\colon {\text{Tr}}_{{{q}^{k}}/q}(g)=0\} $\end{document} \begin{document}$ {\mathcal{C}}_{\mathcal{D}} $\end{document} \begin{document}$ G={\langle {{\gamma }_{1}},{{\gamma }_{2}},\cdots ,{{\gamma }_{s}}\rangle }_{{{\mathbb{F}}_{{{q}^{l}}}}} $\end{document} \begin{document}$ {\mathcal{C}}_{\mathcal{D}} $\end{document} \begin{document}$ G $\end{document} \begin{document}$ \mathbb{F}_{{q}^{k}}^{*} $\end{document} \begin{document}$ \mathcal{D}=\mathbb{F}_{{q}^{k}}^{*}\backslash \cup _{i=1}^{h}{\alpha }_{i}G $\end{document} \begin{document}$ {\mathcal{C}}_{\mathcal{D}} $\end{document} \begin{document}$ {\mathbf{c}}_{x} $\end{document} \begin{document}$ {\mathcal{C}}_{\mathcal{D}} $\end{document} Table 1 summarizes some optimal or distance-optimal linear codes satisfying the Griesmer bound. Conclusions Given a linear code\begin{document}$ {\mathcal{C}}_{\mathcal{D}} $\end{document} \begin{document}$ \mathcal{D}={\mathbb{F}}_{{{q}^{k}}}\backslash \cup _{i=1}^{h}({\alpha }_{i}+G) $\end{document} \begin{document}$ \mathcal{D}=\mathbb{F}_{{q}^{k}}^{*}\backslash \cup _{i=1}^{h}{\alpha }_{i}G $\end{document} \begin{document}$ {\mathcal{C}}_{\mathcal{D}} $\end{document} \begin{document}$ G $\end{document}
Available online , doi: 10.11999/JEIT260763
Abstract:
Objective Unmanned aerial vehicle (UAV)-assisted mobile edge computing (MEC) extends network coverage and improves service flexibility by exploiting the high mobility and rapid deployment capability of UAV platforms. However, the same mobility can also be exploited by an adversary, allowing a malicious UAV to act as a dynamic jammer and continuously degrade the reliability and security of computation offloading links. Existing optimization-based solutions usually rely on convex approximation, block-coordinate iteration, or staged decomposition, and therefore have difficulty responding efficiently when the jammer moves rapidly and communication, computation, and mobility variables are tightly coupled. To address this problem, a secure transmission system for UAV-assisted MEC under malicious flight jamming is investigated. The objective is to maximize the system secure computing capacity by jointly coordinating the service UAV trajectory, user transmit power, and computation-task allocation, thereby improving the robustness of offloading and processing in a highly dynamic hostile environment. Methods A system composed of one MEC-enabled service UAV, one malicious aerial jammer, and multiple ground users is established (Fig. 1 ). The service period is divided into equal time slots, and one user is served in each slot. Line-of-sight air-to-ground and air-to-air channels are adopted, while obstacle blockage, jammer interference power, receiver noise, and path loss are incorporated into the achievable transmission-rate model. A partial-offloading mechanism is used, in which each user divides its workload between local execution and UAV-assisted computation. Transmission delay, edge-computing delay, local-computing delay, UAV computing energy, propulsion energy, and user transmission energy are jointly modeled. The resulting optimization problem maximizes the sum of locally processed and securely offloaded tasks subject to user association, blockage-state, delay, flight-region, maximum-speed, transmit-power, UAV-energy, and minimum secure-computing-capacity constraints. Because the problem is high-dimensional, non-convex, and strongly coupled, it is reformulated as a Markov decision process. The state contains the service UAV position and residual energy, jammer position, user locations, remaining workloads, and blockage states. The action contains the user transmit power, UAV velocity, and task-allocation decision. A Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm is then designed within an actor–critic framework (Fig. 2 ). Two critic networks are used to reduce value overestimation, delayed actor updates are introduced to improve policy stability, and soft target-network updates are employed to support smooth convergence. Simulations are implemented in Python 3.8 with PyTorch, and the main parameters are listed in Table 1 . The proposed method is compared with Deep Deterministic Policy Gradient (DDPG) and a genetic algorithm. Results and Discussions The learned trajectories show that the service UAV can identify the spatial threat caused by the moving jammer and adaptively detour when approaching the interference region, while still visiting users and completing offloading tasks (Fig. 3 ). This behavior demonstrates that trajectory planning and resource allocation are coordinated rather than optimized independently. The convergence comparison shows that the DDPG baseline exhibits larger fluctuations and converges to a lower secure computing capacity, whereas the TD3-based scheme converges faster and more smoothly (Fig. 4 ). The improvement is mainly attributed to the twin-critic structure and delayed policy update, which suppress value overestimation and reduce unstable policy drift. The results further indicate that a higher UAV computing capability increases the secure computing capacity because the processing time per offloaded bit is reduced, encouraging users to offload more tasks (Fig. 5 ). A lower flight altitude also improves performance by shortening the service-UAV-to-user distance and strengthening the communication channel, thereby increasing the amount of data that can be securely transmitted and processed within each time slot (Fig. 6 ). Under different malicious jamming powers, the proposed scheme consistently achieves the highest secure computing capacity, while the genetic algorithm shows noticeable performance variation and weaker adaptation to time-varying channel conditions (Fig. 7 ). These findings confirm that the proposed learning-based design remains effective across changes in computing capability, flight altitude, and interference intensity. Conclusions A TD3-based intelligent trajectory and resource-cooperative secure transmission scheme is developed for UAV-assisted MEC under malicious flight jamming. By transforming the coupled trajectory, power, and task-allocation problem into a continuous-control learning process, the scheme enables autonomous trajectory planning and dynamic communication–computation resource scheduling. Simulation results verify that it provides faster convergence, better stability, and higher secure computing capacity than DDPG and genetic-algorithm baselines. The proposed approach therefore offers a practical solution for secure UAV-assisted edge computing in dynamic and adversarial environments.
Available online , doi: 10.11999/JEIT260772
Abstract:
Objective Traditional phased array radar suffers from beam squint during wide-band and wide-angle scanning, a problem that severely constrains the scanning accuracy, real-time performance, and reliability of the radar system, thereby compromising its practical application effectiveness in demanding scenarios such as military reconnaissance, aerospace, and intelligent transportation systems. In military contexts in particular, this limitation makes it difficult for existing radar systems to meet the requirements of modern combat and surveillance for high precision and wide coverage. In view of this, there is an urgent need to develop an innovative beamforming mechanism to overcome the limitations of traditional approaches in terms of hardware resource redundancy, scanning flexibility, and anti-jamming capability. Therefore, a microwave photonic dual-beam transceiver-shared beamforming system is proposed in this paper. By optimizing the hardware architecture and incorporating dispersive delay and wavelength division multiplexing (WDM) techniques, this system achieves simultaneous dual-beam scanning without beam squint, thereby providing critical theoretical support and a technical pathway for enhancing the overall performance of radar systems. Methods The microwave photonic dual-beam transceiver-shared beamforming system consists of three parts: the radio frequency (RF) front-end, the optically controlled beamforming network, and the broadband antenna. In this system, the optically controlled beamforming network adopts a dispersion-based time-delay method, in which two optical signals with different central wavelengths can generate two distinct time delays within the same dispersive medium. Therefore, based on the WDM approach, by properly setting the initial wavelength values and adjusting the wavelengths of two tunable lasers at equal intervals, two sets of independent proportional delays can be obtained, thereby enabling independent control of two beams. Furthermore, by rationally designing the architecture of the receiving, transmitting, and transceiver modules within the RF front-end, and employing strategies such as duplexing and time-division switching, transceiver sharing can be achieved using the same system. Finally, a broadband Vivaldi antenna is employed for the reception and radiation of wide-spectrum signals. In the future, the number of beams can be expanded by increasing the number of optical sources and RF front-end modules. Results and Discussions The microwave photonic dual-beam transceiver-shared beamforming system enables time-division multiplexing of dual-channel transmitting and receiving beamforming based on a single set of hardware. A 2×8 transceiver-shared platform is constructed for experimental validation. Experimental results demonstrate that the system achieves independent dual-beam spatial scanning without beam squint over a frequency range of 15 GHz to 26 GHz, with a scanning coverage of ±52°. In the transmitting state, when Beam 1 operates at a frequency of 22 GHz, the main-to-sidelobe suppression ratio of the radiation pattern is approximately 13 dB for beam steering directions of ±52°, ±30°, and ±4° (Fig. 11 ). In the transmitting state, when Beam 1 and Beam 2 operate at frequencies of 15 GHz and 26 GHz, respectively, the radiation pattern at a steering direction of 35° exhibits no beam squint (Fig. 12 ), demonstrating wideband operation capability. In the receiving state, with the operating frequencies of Beam 1 and Beam 2 unchanged, the radiation pattern at a steering direction of –42° similarly shows no beam squint (Fig. 13 ). Conclusions This paper employs the same set of hardware resources to achieve dual-beam transceiver-shared beamforming based on dispersion-based delay and WDM technology, and develops a prototype system, thereby addressing the beam squint issue inherent in traditional phased array radar under wideband wide-angle scanning. Experimental results demonstrate that:(1) The optical true time delay method based on the dispersion regime offers real-time tunability, and its combination with WDM technology can effectively expand the number of delay channels;(2) The beamforming system based on optical true time delay exhibits no frequency-dependent beam squint and is capable of wideband (15 GHz–26 GHz) and wide-angle (±52°) operation;(3) The proposed optically controlled beamforming architecture enables time-division switching between transmitting and receiving modes, and dual-beam transmission or dual-beam reception scanning can be achieved through the expansion of optical sources.
Available online , doi: 10.11999/JEIT260899
Abstract:
Objective High-mobility communication scenarios, including low-Earth-orbit satellite links, high-speed railway communications, and unmanned aerial vehicle networks, require reliable control transmission and efficient service-data delivery. Different services have different reliability and throughput requirements, which are difficult to accommodate using a single information mapping structure. Orthogonal Time Sequency Multiplexing (OTSM) operates in the Delay-Sequency (DS) domain and uses the Walsh-Hadamard Transform (WHT) to support transmission over doubly dispersive channels with low implementation complexity. Index Modulation (IM) conveys information through resource indices in addition to constellation symbols. Existing OTSM-based IM schemes improve transmission reliability through sparse activation, but inactive resources limit spectral efficiency. Moreover, applying the same mapping strategy to all information bits does not exploit differences in their importance. To address these limitations, a Double-Layer Multi-Mode OTSM scheme with IM, termed DLMM-OTSM, is proposed. The scheme combines a sparse base layer with a fully activated enhancement layer to support differentiated transmission. Its receiver design balances detection complexity and error performance through layer-specific detection and a limited number of index candidates. Methods Each group of information bits is divided between a base layer and an enhancement layer, which occupy two adjacent rows of the DS-domain grid (Fig. 1 ). The base layer carries reliability-sensitive information using sparse single-mode IM. Its index bits select a subset of available carriers, and its symbol bits are mapped to constellation A on the selected carriers. The remaining carriers are inactive. The enhancement layer carries throughput-oriented information using full-carrier dual-mode IM. Its index bits determine which carriers use constellation B, while the remaining carriers use constellation C. Both carrier subsets transmit symbols, so all enhancement-layer carriers remain active. Constellations B and C are disjoint to distinguish their modes, and their power allocation and relative rotation help separate the constellation points (Fig. 2 ). For each layer, index patterns are selected from a shared legal mapping table to establish a one-to-one correspondence between index bits and transmission patterns. The resulting DS-domain symbols undergo the inverse WHT and vectorization before transmission. For eight available carriers and two selected carriers, there are 28 possible position combinations. Sixteen combinations are used to represent four index bits, and the same mapping is available at both the transmitter and receiver (Table 1 ). With 4-QAM in all three constellations, each group carries eight base-layer bits and twenty enhancement-layer bits. This legal-pattern constraint is applied during detection, preventing the receiver from selecting combinations that do not correspond to transmitted index bits. At the receiver, Maximum Likelihood (ML) detection searches legal index patterns and constellation symbol combinations and serves as a performance benchmark. To reduce the search burden, a Log-Likelihood Ratio (LLR) detector is developed following Minimum Mean Square Error (MMSE) equalization. The detector calculates per-carrier metrics for activation in the base layer and constellation membership in the enhancement layer. These metrics are accumulated over legal index patterns, followed by pattern selection and symbol decisions. A single selected pattern may exclude the correct transmitted vector when individual LLR values are unreliable. Therefore, a Top-K candidate detector retains the K legal patterns with the largest accumulated metrics for each layer. Candidate signal vectors are constructed, and their Euclidean distances are evaluated using the received signal and equivalent channel matrix. The candidate with the minimum distance is selected (Algorithm 1). When K = 1, the method reduces to LLR detection. The additional search effort is controlled by the number of retained candidate patterns. Detection complexity and spectral efficiency are analyzed, and an analytical approximation to the Average Bit Error Probability (ABEP) union bound is derived. Results and Discussions Simulations use an Extended Vehicular A (EVA) multipath Rayleigh fading channel with nine propagation paths, a carrier frequency of 4 GHz, and a subcarrier spacing of 15 kHz. The equivalent channel matrix is assumed known at the receiver. Unless otherwise specified, the relative speed is 500 km/h. Spectral efficiency is calculated from the information bits carried by the data grid, excluding cyclic-prefix overhead, using the same convention for all schemes. For M=N=8, Jb=Je =2, and 4-QAM, DLMM-OTSM achieves a spectral efficiency of 1.75 bit/s/Hz. This is 75% higher than OTSM-IM and approximately 16.7% higher than Pairwise Sequency Index Modulation with OTSM (PSeIM). With the carrier-selection ratios Jb/N= Je/N maintained, its spectral efficiency reaches 2 bit/s/Hz for a 64 × 64 DS-domain grid (Table 2 ). At a signal-to-noise ratio (SNR) of 25 dB, the bit error rates (BER) of the base and enhancement layers are approximately 1.59 × 10–4 and 4.29 × 10–4, respectively. Thus, the base-layer BER is approximately 37% of the enhancement-layer BER, demonstrating differentiated reliability through the two IM structures (Fig. 5 ). Increasing K from 1 to 4 reduces BER by approximately 13.1% at 10 dB and 33.1% at 20 dB (Fig. 3 ). The analytical ABEP approximation approaches the ML simulation results in the high-SNR region. Compared with ML detection, the Top-K detector incurs SNR losses of approximately 0.5 dB and 0.7 dB at BERs of 10–3 and 10–4, respectively, while reducing detection complexity (Fig. 6 ). Under equal spectral-efficiency conditions, DLMM-OTSM achieves SNR gains of approximately 4.0 dB, 3.4 dB, and 3.3 dB over OTSM-IM, PSeIM, and conventional OTSM, respectively, at a target BER of 10–3 (Fig. 7 ). A modulation-order comparison gives spectral efficiencies of 1.125, 1.75, and 2.375 bit/s/Hz for BPSK, 4-QAM, and 8-QAM, respectively. Moving from BPSK to 4-QAM increases spectral efficiency by approximately 55.6%, with an SNR loss of approximately 0.55 dB at a BER of 10–3. Moving from 4-QAM to 8-QAM increases spectral efficiency by approximately 35.7%, but requires approximately 4.5 dB of additional SNR at the same target BER (Fig. 8 ). Conclusions DLMM-OTSM combines layered transmission and multiple IM modes to support heterogeneous data requirements in high-mobility channels. By assigning reliability-sensitive information to a sparse base layer and throughput-oriented information to a full-carrier dual-mode enhancement layer, the proposed scheme achieves flexible unequal error protection and improved spectral efficiency. The LLR detector reduces receiver complexity, while the Top-K mechanism recovers much of the performance loss caused by single-pattern LLR decisions with comparatively low complexity. Theoretical and simulation results demonstrate that the proposed DLMM-OTSM provides an effective tradeoff among reliability, spectral efficiency, and detection complexity.
Available online , doi: 10.11999/JEIT260945
Abstract:
Significance With the rapid development of the low-altitude economy, emergency communications, and related industries, multi-Unmanned Aerial Vehicle (UAV) communication systems have shown significant application value in disaster rescue, low-altitude monitoring, and other scenarios, owing to their advantages in wide-area coverage and link compensation. However, affected by strong link coupling, dynamic topology variation, and enlarged Line-of-Sight (LoS) interference ranges, resource allocation in multi-UAV systems is characterized by high dimensionality, time variation, and multi-agent cooperation. Traditional model-driven optimization methods, such as convex optimization and game theory, are difficult to satisfy the requirements of real-time adaptive decision-making in complex dynamic environments. Therefore, Artificial Intelligence (AI)-enabled intelligent resource allocation has become an important direction for improving robust transmission, autonomous decision-making, and collaborative optimization in multi-UAV communication systems. Based on this, a survey is presented on the architectures, models, and challenges of intelligent resource allocation for multi-UAV communication systems. Progress This paper presents a comprehensive survey of recent advances in intelligent resource allocation for multi-UAV communication systems, with an in-depth discussion of typical network architectures, including cellular cooperative communication, Reconfigurable Intelligent Surface (RIS)-assisted communication, Integrated Sensing and Communication (ISAC), and Mobile Edge Computing (MEC). Among them, cellular cooperative architectures enhance coverage and link connectivity through UAV-enabled aerial relays or aerial base stations. RIS-assisted architectures improve blocked link quality by reconstructing the wireless propagation environment. ISAC architectures support the coordination of communication coverage, target localization, and environmental sensing. MEC architectures deploy computing resources at ground edge servers or UAV nodes to support task offloading and proximate computing. Different architectures vary in node functions, link forms, and resource coupling relationships, thereby determining different optimization variables and intelligent decision-making requirements. Then, typical intelligent models for multi-UAV resource allocation are introduced. Deep Neural Networks (DNNs) can learn nonlinear mappings from system states to continuous resource variables. Convolutional neural networks are suitable for extracting spatial features from channels, topologies, and user distributions. Recurrent neural networks can characterize temporal variations in channel states and UAV positions. Deep Reinforcement Learning (DRL) is applicable to long-term sequential decision-making problems, such as UAV trajectory control and power optimization. Federated Learning (FL) enables distributed collaborative training without sharing raw data. Furthermore, traditional optimization methods and intelligent methods are compared in terms of their applicable characteristics, and it is indicated that these methods should be selected according to scenario dynamics, model availability, real-time requirements, and cooperation scale. Finally, existing studies on intelligent resource allocation are compared across typical architectures, including cellular communication, RIS-assisted communication, ISAC, and MEC. In cellular communication systems, multi-agent proximal policy optimization, multi-agent Deep Deterministic Policy Gradient (DDPG), and other methods are used for trajectory control, frequency allocation, routing optimization, and power allocation. In RIS-assisted systems, double deep Q-network, DDPG, and other methods are used for the joint optimization of UAV trajectories, RIS phase shifts, beamforming, and task offloading. In ISAC systems, generative diffusion-driven multi-agent DRL and adaptive DNN splitting mechanisms are used for sensing-communication resource coordination, robust scheduling, and integrated sensing-communication-computation task processing. In MEC systems, DRL and FL are applied to task offloading, user association, power control, device selection, and priority-aware task scheduling. Conclusions This paper provides a comprehensive survey of intelligent resource allocation for multi-UAV communication systems, with particular emphasis on typical architectures, intelligent models, and current research progress. First, typical architectures of multi-UAV communication systems are systematically reviewed, and the node functions under different architectures are analyzed. Then, typical intelligent models for multi-UAV communication systems are summarized in terms of their mechanisms and applicable characteristics in state representation, feature extraction, sequential decision-making, and distributed cooperation. Furthermore, intelligent resource allocation studies for multi-UAV communication systems are compared across different architectures. Finally, existing challenges are identified, including difficulties in multi-objective tradeoffs under three-dimensional mobility, complex cooperative decision-making under LoS interference, and cross-layer resource coupling under multifunctional integration. Relevant research directions are also proposed. These directions are expected to provide valuable references for future research on intelligent resource allocation in multi-UAV communication systems. Prospects Although intelligent resource allocation methods have made progress in link recovery, coverage enhancement, RIS-assisted link optimization, sensing-communication resource coordination, and MEC task offloading, several challenges still remain in complex dynamic multi-UAV scenarios. First, the three-dimensional mobility of UAVs dynamically couples coverage range, service delay, and energy consumption. Existing methods with fixed reward weights are difficult to adapt to changes in task priorities. Therefore, multi-objective Reinforcement Learning (RL) and dynamic reward weighting can be combined to construct adaptive tradeoff mechanisms. Second, air-to-air and air-to-ground LoS links enlarge the interference propagation range, making spectrum reuse and user association more complex. Graph neural networks and interference prediction mechanisms can be introduced to improve the representation capability of dynamic interference topology. Finally, with the integration of RIS, ISAC, and MEC, UAVs are gradually evolving into multifunctional nodes. Hierarchical RL and cross-layer joint modeling can be further explored to realize the collaborative optimization of communication, sensing, computing, and flight-control resources.
Available online , doi: 10.11999/JEIT260618
Abstract:
Objective Torus Fully Homomorphic Encryption (TFHE) is an important cryptographic technique for privacy-preserving computation, as it enables Boolean circuits to be evaluated directly over encrypted data. In TFHE, programmable bootstrapping refreshes ciphertext noise while evaluating a user-defined lookup-table function, thereby providing a flexible foundation for secure inference, encrypted data processing, and privacy-preserving cloud services. However, bootstrapping remains the major performance bottleneck of TFHE. A complete bootstrapping procedure consists of modulus switching, blind rotation, sample extraction, and identity key switching. Among these operations, blind rotation and key switching involve intensive polynomial arithmetic and frequent accesses to large bootstrapping keys and key switching keys. Consequently, software implementations usually suffer from high latency and limited throughput, which limits the deployment of TFHE in practical high-throughput applications. To overcome this limitation, this paper proposes an FPGA-based accelerator for TFHE programmable bootstrapping. The objective is to improve bootstrapping execution efficiency by combining a dedicated streaming data path, an optimized Number Theoretic Transform (NTT)-based polynomial multiplication module, and a batching mechanism oriented toward bootstrapping key reuse. Methods A fully hardware-based streaming architecture is designed to map the main stages of TFHE programmable bootstrapping onto dedicated FPGA. The accelerator consists of a test vector generation module, a polynomial rotation module, a decomposition module, an element-wise multiplication module, a sample extraction module, and an identity key switching module. These modules are connected through on-chip buffers and streaming interfaces, enabling the bootstrapping data path to execute in a pipelined manner. Compared with CPU-like control-oriented architectures, the proposed architecture reduces instruction scheduling and centralized control overhead, and is suitable for continuous processing of multiple ciphertexts. To accelerate the external product in the CMUX operation, a Radix-32 Four-Step Number Theoretic Transform (NTT) module is developed for polynomial multiplication. The required 1024 -point NTT is decomposed into smaller 32-point NTT operations, reducing the hardware complexity of directly implementing a large-size NTT. A negative-cycle NTT method is further adopted to avoid zero-padding and reduce redundant computation. In addition, a generalized Mersenne prime, also known as the Goldilocks prime, is used as the NTT modulus. By exploiting its special form, part of the twiddle-factor multiplication and modular reduction operations can be converted into shifts, additions, and subtractions. For the 192-bit intermediate data generated in the Radix-32 NTT pipeline, a large-word modular reduction method is designed to avoid costly division-based modular reduction. To alleviate off-chip memory traffic caused by bootstrapping key accesses, a ciphertext batching mechanism is introduced. Multiple TLWE ciphertexts sharing the same bootstrapping key are organized as a batch. After a portion of the bootstrapping key is loaded into the on-chip buffer, it can be reused by multiple ciphertexts in the same batch. Meanwhile, the key data required for subsequent computation are prefetched to partially hide off-chip memory access latency. Results and Discussions The proposed accelerator is implemented and evaluated on a Xilinx Alveo U280 FPGA platform. The platform provides HBM2 memory with multiple independent channels to support high-bandwidth accesses to input TLWE ciphertexts, bootstrapping keys, key switching keys, and output ciphertexts. Different HBM channels are allocated according to the bandwidth requirements of different data types. Several channels are used for bootstrapping key transfer to match the throughput demand of the NTT-based polynomial multiplication module, while more channels are assigned to the key switching key to support the high access rate of the identity key switching module. In the physical implementation, the element-wise multiplication module is placed across different super logic regions to balance resource utilization and reduce timing pressure. Experimental results show that the proposed accelerator correctly completes TFHE programmable bootstrapping on the FPGA platform. Under the same TFHE parameter setting, it achieves higher bootstrapping efficiency than existing CPU and GPU implementations. Benefiting from the streaming pipeline and ciphertext batching mechanism, the proposed design also improves bootstrapping throughput compared with previous FPGA-based accelerators. These results indicate that the proposed architecture reduces polynomial multiplication cost and improves bootstrapping key reuse. The batching mechanism amortizes the loading cost of the bootstrapping key over multiple ciphertexts, thereby alleviating the off-chip memory bottleneck. Resource utilization results show that the accelerator consumes a moderate amount of LUT, register, BRAM, and DSP resources on the Alveo U280 platform, leaving room for further parallel extension. Power evaluation also shows that the proposed design achieves a reasonable trade-off among throughput, resource consumption, and energy efficiency. Although the absolute latency is still affected by the sequential dependency of blind rotation iterations and off-chip key accesses, the overall throughput is significantly improved in batch-processing scenarios. Conclusions A dedicated FPGA accelerator for TFHE programmable bootstrapping is presented in this paper. The proposed design introduces a fully hardware-based streaming bootstrapping architecture, an efficient Radix-32 Four-Step NTT module with large-word modular reduction, and a ciphertext batching mechanism for bootstrapping key reuse. These techniques jointly reduce control overhead, polynomial multiplication cost, and repeated off-chip memory accesses. The implementation results on the Alveo U280 FPGA platform demonstrate that the accelerator correctly executes TFHE programmable bootstrapping and achieves high throughput under practical hardware resource constraints. Therefore, the proposed accelerator provides an effective hardware implementation approach for high-throughput privacy-preserving computation based on TFHE. Future work may further explore larger-scale parallelism, more flexible parameter support, and tighter integration with privacy-preserving application workloads.
Available online , doi: 10.11999/JEIT260619
Abstract:
Objective With the rapid development of the Internet of Things, massive terminal devices deployed in physically open environments are highly vulnerable to physical probing, side-channel attacks, and bus sniffing. Current IoT security solutions exhibit significant limitations: software-based protocol implementations suffer from severe performance bottlenecks and lack physical protection, while dedicated hardware accelerators based on Finite State Machine (FSM) configurations experience state explosion, poor scalability, coarse control granularity, and the critical risk of sensitive data exposure on unprotected buses due to frequent data movement. Therefore, it is of great necessity and importance to construct an efficient, lightweight, and security-analyzable hardware protection system that supports fine-grained operator orchestration and cross-layer trust closure for resource-constrained IoT devices. Methods To address the aforementioned issues, an Instruction-Based Secure Coprocessor Architecture for Cross-Layer Trust Closure (IBSCA) is proposed. The architecture abandons the traditional FSM hard-wired control mode and introduces a Reduced Instruction Set Computer concept to construct a processor-like “Decode-Dispatch-Execute” pipeline. A dedicated 32-bit reduced instruction set architecture is defined, where all instructions are divided into five functional fields: Control Bits, Functional Unit Selection, Operand Addresses, Destination Address, and Reserved Bits. The Control Bits explicitly define data flow direction and operational attributes, where the source selection bit and result write-back bit are key to realizing the “data residency” security policy. The hardware microarchitecture consists of four core modules: Interface, Control Logic, Functional Unit, and Memory. The Interface Module serves as the sole communication channel to minimize the attack surface; the Control Logic Module dynamically generates micro-operation control signals; the Functional Unit Module integrates heterogeneous operators like PUF, ECC, and Hash; and the Memory Module guarantees data confidentiality through on-chip data residency. The execution flow comprises three coupled stages: instruction sequence generation, fetch and decode, and instruction execution. The host directly injects instruction streams via the standard bus, abstracting heterogeneous operators into independently addressable atomic operations to decouple control logic from data paths. Furthermore, a PUF-based lightweight WiFi authentication protocol is mapped into the IBSCA to verify the design. Results and Discussions Based on the FPGA platform, the IBSCA secure IP is implemented and evaluated. Functionally, waveform-level simulation verifies that when a PUF instruction is issued, the coprocessor accurately decodes it, activates the PUF module, and writes the response back to the on-chip register, explicitly demonstrating that sensitive data like PUF responses and keys reside and flow entirely within the chip (Fig. 4 ). End-to-end protocol tests are conducted by mapping the PUF-based WiFi authentication protocol into the IBSCA. The protocol steps are successfully translated into binary instruction sequences, executing continuously to output the correct Response and Session Key (Fig. 5 ). Performance evaluations indicate that the access latency of IBSCA is 6802 ns, which is 23.95% shorter than the software baseline (8944 ns) and only 12.2% longer than the FSM baseline (6062 ns) (Table 2 ). The slight latency gap compared to FSM stems from implementation details such as handshake bubbles rather than paradigm flaws. Resource overhead analysis shows that the IBSCA control module consumes only 800 LUTs and 1550 FFs, achieving a significant reduction of 67.8% and 57.8% respectively compared to the FSM control module (2483 LUTs and 3676 FFs), as its general scheduling logic absorbs complexity into instruction encoding rather than explicit state nodes. Furthermore, SAIF-annotated post-route power analysis (100% nets matched, high confidence) shows that, at the timing-verified 100 MHz clock, the total on-chip power of IBSCA is 0.365 W, of which the control module contributes only about 5 mW of dynamic power (4.1% of the total dynamic power). Conclusions An instruction-based secure coprocessor architecture for cross-layer trust closure is proposed to address the limitations of existing IoT security solutions. By introducing a reduced instruction set paradigm and constructing a processor-like pipeline, the architecture successfully decouples control logic from data paths, eliminating the FSM state explosion problem and providing runtime dynamic programming capabilities. The fine-grained orchestration of heterogeneous operators and the on-chip data residency mechanism ensure that sensitive data form a closed loop on-chip, constructing a robust hardware closed-loop trust chain from physical root of trust generation to runtime data encryption protection. FPGA prototype verification and resource analysis confirm that IBSCA achieves a superior balance among protocol extension flexibility, hardware resource overhead, and security performance, providing a flexible, efficient, and reliable hardware security foundation for resource-constrained IoT terminal devices.
Available online , doi: 10.11999/JEIT260718
Abstract:
Objective With the growing deployment of digital identity systems, protecting user privacy and reducing excessive identity disclosure has become a key requirement.Existing digital identity frameworks, such as eIDAS 2.0 and W3C Verifiable Credentials, rely on cryptographic mechanisms to enable selective disclosure. However, current anonymous credential schemes face two major limitations. First, most existing schemes are built on classical number-theoretic assumptions, which are vulnerable to Shor’s quantum algorithm and cannot resist future quantum computing attacks. Second, high-performance privacy-preserving schemes generally follow the Signature with Efficient Protocol (SEP) paradigm, which requires identity issuers to adopt specific nonstandard signature primitives and results in poor compatibility with existing identity infrastructures. Although the zk-creds paradigm improves compatibility by using general-purpose zero-knowledge proofs to encapsulate standard credentials, existing implementations mainly depend on pairing-based proof systems, such as Groth16, and therefore lack post-quantum security. Meanwhile, lattice-based solutions introduce large proof sizes and high communication overhead due to the complexity of high-dimensional matrix operations. Therefore, developing an anonymous credential scheme that maintains compatibility with existing credential systems while achieving post-quantum security remains an important challenge. Methods A lattice-based anonymous credential scheme named LBAC (Lattice-Based Anonymous Credentials) is proposed under the zk-creds paradigm. The scheme adds a zero-knowledge proof layer between users and verifiers to decouple presentation from the original credential issuance.Existing digital credentials are directly used as private inputs, and no modification or interaction with credential issuers is required. A hybrid proof framework based on the lattice-based LaBRADOR interactive proof system is constructed. First, a random masking mechanism based on Ajtai commitments is introduced to achieve zero-knowledge and non-interactive proof generation. The rejection sampling method is adopted to blind the original witness, and the Fiat-Shamir heuristic is applied to convert the interactive protocol into a non-interactive zero-knowledge argument. Second, the Sumcheck protocol is introduced as a proof compressor to reduce the verification complexity of arithmetic circuits generated from standard signature verification. By transforming large-scale circuit satisfiability statements into smaller polynomial evaluation and dot-product proof tasks, the computational overhead of LaBRADOR is reduced. Finally, based on the design concept of the Crescent scheme, the improved LaBRADOR proof generation process is divided into an offline Prepare phase and an online Show phase. The computationally intensive operations, including Sumcheck compression and the initial D–1 rounds of recursive folding, are completed during the Prepare phase, where intermediate states and base vectors are stored locally. During the Show phase, the prover only performs lightweight lattice vector re-randomization using a fresh nonce and cached data, enabling efficient credential presentation with multi-show unlinkability. Results and Discussions The security properties of the LBAC scheme are analyzed and proven under the random oracle model. Experimental evaluations are conducted on an Ubuntu 20.04 platform equipped with an Intel Xeon Platinum 8352V processor. The results demonstrate that the Prepare-Show architecture effectively improves credential presentation efficiency. By transferring computationally intensive operations to the Prepare phase, the Show phase achieves a presentation time of 5.27 ms, which is lower than that of existing quantum-resistant schemes based on Ligero and remains competitive with classical non-quantum-resistant schemes (Fig. 4 ). For a circuit scale of 220 constraints, the generated proof size is 50.17 kB, showing lower communication overhead than existing quantum-resistant alternatives (Fig. 5 ). Unlike conventional quantum-resistant schemes whose proof sizes increase rapidly with statement complexity, LBAC maintains a logarithmic proof size through the recursive amortization mechanism of the modified LaBRADOR system. The verification time is 621 ms, which is acceptable considering the post-quantum security provided by the scheme and the fact that verification is generally performed on powerful servers. This overhead can be further reduced through batch verification techniques and GPU acceleration for Number Theoretic Transform (NTT). Conclusions A post-quantum secure anonymous credential scheme for existing digital credentials is proposed by integrating the zk-creds paradigm with an optimized LaBRADOR proof system and the Sumcheck protocol. The proposed framework improves compatibility with existing credential systems while providing post-quantum security, compact proof sizes, and efficient credential presentation. The correctness, presentation unforgeability, and anonymity with multi-show unlinkability are formally proven. Future research will focus on simulation extractability, hardware-based device binding, and traceability mechanisms to further improve security in adaptive adversarial environments and support credential management while preserving privacy.
Available online , doi: 10.11999/JEIT260255
Abstract:
Objective Multimodal sentiment analysis is an important research topic in the field of human-computer interaction because it integrates information from multiple modalities, such as text, audio, and video, to understand human emotions. However, existing methods still have several limitations. First, many approaches neglect the temporal dependencies among multimodal features, resulting in insufficient utilization of temporal information. Second, conventional multimodal fusion strategies usually rely on simple feature concatenation, which limits cross-modal interaction and may introduce redundant features. Moreover, sentiment expressions are continuous and ambiguous, whereas traditional hard contrastive learning strategies generally model sample relationships in a binary manner, making it difficult to capture subtle semantic differences among samples. To address these problems, this paper proposes a multimodal sentiment analysis method based on adversarial optimization and triple soft contrastive learning. The proposed framework enhances multimodal representation learning by modeling temporal dependencies and introducing a soft contrastive learning mechanism that assigns continuous similarity weights between samples. Meanwhile, adversarial optimization is incorporated to improve the robustness and discriminative capability of the learned representations. Methods The proposed method constructs a triple soft contrastive learning framework with adversarial optimization. The overall architecture consists of three components: multimodal feature extraction, triple soft contrastive representation learning, and adversarial optimization. First, modality-specific encoders are employed to extract representations from textual, acoustic, and visual inputs. The extracted features are mapped into a shared representation space to facilitate multimodal interaction. To capture temporal dependencies within multimodal data, a temporal contrastive learning strategy is introduced to model the relationships between sequential features and preserve temporally consistent representations. Second, a soft contrastive learning mechanism is adopted to overcome the limitations of conventional hard contrastive learning. Instead of using binary labels to distinguish positive and negative samples, the proposed method assigns continuous similarity weights according to the sentiment similarity between samples, enabling the model to better capture gradual sentiment transitions and alleviate the influence of ambiguous sentiment boundaries. Third, an adversarial optimization strategy is introduced to further improve the robustness of representation learning. By constructing a dynamic adversarial training process between representation learning and adversarial perturbation, the model learns more discriminative and stable multimodal representations. The overall training objective combines the soft contrastive loss, temporal contrastive loss, and adversarial loss to achieve effective multimodal representation learning. The overall framework of the proposed model is illustrated in (Fig. 2 ). Results and Discussions The effectiveness of the proposed method is evaluated through extensive experiments on widely used multimodal sentiment analysis datasets. The experimental results demonstrate that the proposed adversarial optimization-based triple soft contrastive learning framework improves the performance of multimodal sentiment analysis. Compared with representative baseline models, the proposed method achieves competitive performance in terms of classification accuracy, F1-score, and regression metrics. The comparisons with existing multimodal methods are summarized in (Tables 5 and 6 ). The results indicate that soft contrastive learning enables the model to capture fine-grained sentiment relationships between samples, while adversarial optimization improves the robustness of representations against noisy inputs. Furthermore, ablation studies verify the contribution of each component to performance improvement. In particular, the soft contrastive learning mechanism effectively alleviates the limitations of hard boundary assumptions in traditional contrastive learning, while temporal contrastive learning preserves temporal sentiment information. The adversarial optimization module further enhances the discriminative capability of learned representations. These results demonstrate the effectiveness of combining adversarial optimization with contrastive representation learning for multimodal sentiment analysis. Conclusions This paper proposes a multimodal sentiment analysis method based on adversarial optimization and triple soft contrastive learning. The proposed approach introduces a soft contrastive learning strategy to model continuous sentiment similarity between samples, while temporal contrastive learning captures temporal dependencies in multimodal data. In addition, adversarial optimization is incorporated to enhance representation robustness and improve generalization capability. Experimental results on benchmark datasets demonstrate that the proposed method achieves competitive performance and effectively improves multimodal sentiment representation learning. The proposed framework provides a potential solution for robust and fine-grained multimodal sentiment analysis.
Available online , doi: 10.11999/JEIT260719
Abstract:
Objective Functional safety has become a critical requirement in automotive System-on-Chip (SoC) design in compliance with the ISO 26262 functional safety standard. Conventional exhaustive fault injection faces prohibitive computational costs in large-scale circuits, which may result in insufficient identification of critical fault regions or overly conservative protection strategies. Existing learning-based methods still suffer from limited cross-scale transferability, insufficient integration of circuit-domain knowledge, and inadequate support for downstream hardening decisions. This work aims to develop an accurate, efficient, and transferable framework for node-level fault-criticality prediction and low-overhead selective hardening to support automotive SoC functional safety verification. Methods A primitive-level fault dataset is constructed from 40 DesignWare (DW) library circuits covering four functional families: arithmetic and logic, communication and encoding, storage and control, and error correction and detection. Each circuit is converted into a unified directed graph with structural, connectivity, timing, and activity-related features. Node-level fault-criticality labels are generated through controlled stuck-at fault injection with ASIL-aware threshold calibration. An expert-guided graph learning model is proposed, which adopts a Graph Attention Network (GAT) with residual connections to learn transferable fault-sensitive representations. A four-layer expert rule calibration module is further developed to incorporate circuit-domain knowledge. The propagation-oriented rules emphasize nodes with strong fault propagation capability, such as high fan-out and critical-path nodes; the safety-context rules enhance the evaluation of clock-domain crossings and reset-related logic; the confidence-correction rules reduce unstable predictions and false positives; and the circuit-adaptive rules dynamically adjust rule intensity according to different functional families. Furthermore, an explanation-based node ranking mechanism combining structural and feature attribution analyses is established to form a prediction-explanation-hardening closed-loop framework for targeted protection (Fig. 1 , Fig. 2 ). Results and Discussions The proposed method is evaluated on nine benchmark circuits with gate counts ranging from 273 to 15 290. The results show that the method achieves an accuracy of up to 99.4% and outperforms GAT, CNN, and three state-of-the-art methods on all test circuits (Fig. 3 , Table 4 ). The introduction of expert rules increases the average accuracy and F1 score by 1.77 and 1.98 percentage points, respectively, with greater improvements observed in complex circuits (Fig. 7 , Fig. 8 ). Ablation experiments demonstrate that the four functional families provide complementary fault knowledge and jointly enhance cross-scale generalization capability (Fig. 5 , Fig. 6 ). Compared with traditional fault simulation, the proposed framework achieves a geometric mean speedup of 166.4× and a maximum speedup of 1 119× on I2C, while the explainability module introduces only approximately 20% additional computational overhead (Fig. 9 ). The selective hardening strategy reduces area overhead by 8.35%～14.95% compared with global Triple Modular Redundancy (TMR), while maintaining the Single-Point Fault Metric (SPFM) above 99.3% and the Latent Fault Metric (LFM) above 89.0%. All nine circuits achieve the SPFM target value associated with ASIL-D, eight circuits satisfy both SPFM and LFM target values associated with ASIL-D, and the remaining circuit satisfies the corresponding ASIL-C target values (Fig. 10 , Fig. 11 ). Conclusions The proposed framework enables effective cross-scale fault-criticality prediction by learning transferable fault-sensitive patterns from primitive circuits without retraining on target designs. It achieves high prediction accuracy, substantial acceleration, and low hardware overhead while satisfying ISO 26262-related functional safety requirements. Future work will expand primitive dataset coverage, incorporate causal inference to reduce dependence on manually designed expert rules, and explore extensions toward heterogeneous chiplet architectures.
A High-Parallelism Simulated Adiabatic Bifurcation Processor for Combinatorial Optimization Problems
Available online , doi: 10.11999/JEIT260780
Abstract:
Objective Combinatorial Optimization Problems (COPs) arise in numerous applications, including network optimization, autonomous driving path planning, VLSI design, and computational biology. Their solution space grows exponentially with problem size, making it difficult for conventional von Neumann architectures (e.g., CPUs) to obtain high-quality solutions within practical time limits. Quantum-inspired Ising machines have emerged as a promising computing paradigm for solving COPs. However, existing CMOS-based Ising machines still face significant challenges in simultaneously achieving high speed, high energy efficiency, and high solution accuracy. Discrete-time Ising machines typically require thousands of iterations and depend on random number generators, leading to large chip area and long Time to Solution (TTS). Continuous-time Ising machines converge more rapidly but generally exhibit inferior solution quality because they are prone to becoming trapped in local minima. To address these limitations, a high-parallelism Simulated Adiabatic Bifurcation (SAB) processor is designed and fabricated using 65 nm CMOS technology. Methods The proposed processor implements the SAB algorithm, which is inspired by quantum adiabatic optimization. Unlike simulated annealing, SAB does not require Gibbs sampling or random number generators to escape local minima. The algorithm models each spin as a nonlinear oscillator and solves a set of ordinary differential equations to simulate the adiabatic evolution of a classical nonlinear Hamiltonian system exhibiting bifurcation behavior. The proposed processor adopts a hardware architecture that supports a fully connected spin topology and exploits the inherent fully parallel spin-update capability of the SAB algorithm. To enable bubble-free iterative computation, a three-stage pipelined spin-update strategy is proposed, consisting of coupling-coefficient access, momentum update, and position update. To reduce the storage overhead associated with the fully connected coupling matrix, a folded coupling-coefficient storage array is designed by exploiting matrix symmetry to eliminate redundant storage. The momentum update and position update units are implemented using 8-bit fixed-point arithmetic (2 integer bits and 6 fractional bits), enabling SAB evolution with low hardware overhead. The fabricated processor occupies an area of 0.47 × 1.25 mm2 and operates at a clock frequency of 200 MHz with a 1 V supply voltage. Results and Discussions The fabricated processor consumes only 14 mW, with the spin evolution module accounting for 58% of the total power consumption. The folded coupling-coefficient storage array reduces storage area by 52% compared with full coupling-matrix storage, while its rectangular organization alleviates routing congestion and layout voids during physical implementation (Fig. 6 ). The parallel loading access mechanism enables all coupling coefficients to be read and distributed within a single clock cycle, eliminating the memory-access bottleneck associated with serial loading. For predefined Max-Cut problems configured as an 8 × 8 grid (64 nodes) with 2-bit coupling coefficients, the processor converges to the global optimum within five update cycles and achieves a final Ising energy (Hamiltonian) of −4 032 (Fig. 9 ). This result agrees with the analytical expression n4−n2, confirming the minimum TTS for these Max-Cut instances. For larger Max-Cut problems (108 × 108 nodes), the SAB algorithm achieves 100% solving accuracy relative to the theoretical ground state (Fig. 10 ). Monte Carlo simulations over 1,000 independent randomly generated Max-Cut instances show that SAB achieves an average Hamiltonian of –3 617.42, outperforming simulated annealing, which yields an average Hamiltonian of –3 352.12 (Fig. 11 ). For 3-SAT problems containing 40 clauses and 8 variables, the processor solves instances with Clause-to-Variable Ratio (CTV) values of 3, 4, and 5 in approximately 6 ms, 10 ms, and 16 ms, respectively (Fig. 12 ). These results agree with the theoretical phase-transition behavior and demonstrate the effectiveness of the proposed processor across problems of different complexity. Conclusions A high-parallelism SAB processor that enables fully parallel spin updates without duplicating spin replicas is presented. A three-stage pipeline integrating coupling-coefficient access, momentum update, and position update provides bubble-free parallel updating and low-latency solving. A folded coupling-coefficient storage architecture significantly reduces storage area overhead, while 8-bit fixed-point representations for momentum and position balance computational accuracy with hardware efficiency. Compared with previously reported fully connected Ising machines, the proposed processor achieves 100% solving accuracy for the evaluated benchmark problems while providing higher energy efficiency and lower hardware overhead, demonstrating strong potential for edge combinatorial optimization applications.
Available online , doi: 10.11999/JEIT260771
Abstract:
Objective Orthogonal Delay-Doppler Division Multiplexing (ODDM) is a delay-Doppler (DD) domain multicarrier modulation scheme that transmits information symbols according to the resolution of the DD plane. By exploiting orthogonal pulses in the DD domain, ODDM achieves a structured input-output relationship and a sparse equivalent channel representation in high-mobility wireless communication scenarios. However, the orthogonality of ODDM typically relies on the ideal Nyquist property of the pulse shaping filter. In practical systems, finite-length pulses cannot strictly satisfy this condition, and their non-ideal Nyquist characteristics introduce additional energy leakage. In the presence of carrier frequency offset (CFO), this energy leakage further couples with inter-carrier interference to form residual interference. If such residual interference is neglected in receiver design, the detector operates based on a mismatched equivalent channel model, resulting in degraded symbol detection performance. To address this issue, this paper investigates the interference characteristics of ODDM systems with non-ideal pulse shaping under CFO, and proposes an iterative detection algorithm to compensate for the performance loss caused by neglecting residual interference under the ideal Nyquist assumption. Methods First, a DD-domain equivalent interference model is established for ODDM systems with non-ideal pulse shaping under CFO, characterizing the residual interference that is neglected under the ideal Nyquist pulse assumption. Second, based on the structure of the derived DD-domain interference matrix, the equivalent interference is decomposed into a nominal interference component captured by the ideal pulse model and a residual interference component introduced by non-ideal pulse shaping. Building on this decomposition, a linear residual interference cancellation (LRC) detection algorithm is designed to suppress residual interference through a small number of iterations, thereby compensating for the performance loss incurred by neglecting such interference. Finally, the accuracy of the adopted interference model based on practical pulses and the effectiveness of the proposed detection algorithm are validated through Signal-to-Interference-plus-Noise Ratio (SINR) and Symbol Error Rate (SER) simulations. Results and Discussions A comprehensive evaluation of the non-ideal pulse interference modeling framework and the proposed LRC detection algorithm is conducted in this paper. Simulation results demonstrate that the SINR computed by the non-ideal pulse interference model maintains consistently closer agreement with Monte Carlo simulations, with a deviation of no more than 0.05 dB, whereas the ideal Nyquist pulse model overestimates SINR by up to 0.7 dB across the simulated SNR range (Fig. 3 ). This discrepancy is particularly pronounced at low residual CFO and high SNR. In terms of detection performance, the Message Passing (MP) detector commonly used under the ideal pulse assumption exhibits an increasing performance gap from the MMSE detector based on the full interference model as the normalized CFO increases. At an SNR of 19 dB, the SER ratio between MP and the full-model MMSE detector rises from approximately 2.2 at zero CFO to approximately 24.5 when the normalized CFO is 0.1, whereas the proposed LRC detector remains close to the full-model MMSE detector, with its SER staying within about 1.5 times that of the full-model MMSE detector under the same conditions (Fig. 5 ). This confirms that model mismatch incurs non-negligible performance degradation. The proposed LRC detection algorithm effectively compensates for this performance loss and closely approaches the detection performance of the full-model MMSE detector with as few as 2 iterations across all tested CFO values. Consistent conclusions are further corroborated by results obtained under 16QAM, 64QAM, and 256QAM modulations (Fig. 6 ): as the modulation order increases, the performance gap between the MP detector and the full-model MMSE detector widens, whereas LRC maintains close approximation to the full-model MMSE across different modulation orders. Furthermore, under doubly selective channels, LRC still achieves SER performance close to that of the full-model MMSE detector and outperforms MP detection (Fig. 7 ). When imperfect channel state information (CSI) is further considered, the SERs of all three detection methods increase as the channel estimation SNR decreases. Among them, the performance degradation of LRC and the full-model MMSE is relatively limited, and their detection results remain close to each other, whereas MP is more significantly affected by channel estimation errors. These results further demonstrate that the proposed method exhibits good robustness and broad applicability in interference-limited systems with non-ideal pulse shaping. Conclusions This paper investigates the interference analysis and signal detection problem in ODDM systems with non-ideal pulse shaping under CFO. Based on the characterization of the equivalent interference structure in the delay-Doppler domain under the joint effect of non-ideal pulse shaping and CFO, a linear residual interference cancellation detection algorithm is proposed. The algorithm compensates for the residual interference neglected under the ideal Nyquist pulse assumption, thereby recovering the associated performance loss. Simulation results demonstrate that the adopted non-ideal pulse interference model more accurately characterizes the system interference behavior. Based on this model, the proposed LRC detection algorithm effectively compensates for the performance loss caused by model mismatch, achieving close approximation to the detection performance of the full-model MMSE detector while maintaining lower computational complexity.
Available online , doi: 10.11999/JEIT261132
Abstract:
Objective Low Earth orbit (LEO) integrated communication and navigation (ICAN) networks can provide simultaneous communication and positioning services through shared satellite resources. In regions subject to directional jamming, however, interference affects satellite-ground links unevenly. A reduction in signal-to-interference-plus-noise ratio (SINR) lowers the achievable communication rate and increases pseudorange measurement uncertainty, thereby degrading positioning accuracy. This coupling is particularly consequential for mission-critical users, who require simultaneous access to multiple satellites for communication redundancy and navigation. Satellite selection is further complicated by competition for low-interference satellites with favorable observation geometry, as each satellite has limited capacity to serve multiple users. Existing approaches based on link quality, geometric dilution of precision (GDOP), or joint communication-navigation utility do not fully account for the combined effects of directional interference, observation degradation, and shared-resource constraints. Moreover, conventional nonoverlapping coalition models cannot adequately represent users participating in multiple satellite service relationships. To address these issues, this paper proposes an anti-jamming satellite selection method based on an overlapping coalition formation game (OCFG). The objective is to reconstruct satellite associations in response to interference while balancing communication performance and positioning accuracy and accounting for competition for satellite resources. Methods A snapshot-based LEO-ICAN model is developed for mission-critical ground users in a localized region exposed to directional jamming. Satellite geometry and interference conditions are treated as fixed within each scheduling interval, while binary association variables describe multisatellite access under per-satellite service-capacity limits. The interference model accounts for the directional coverage of the jammer and satellite-ground channel conditions to determine link SINR and jamming-to-noise ratio (JNR). Communication performance is measured by the sum of bandwidth-normalized rates over selected links. For navigation, SINR-dependent pseudorange error variances determine the observation weights used to evaluate weighted geometric dilution of precision (WGDOP) and positioning root-mean-square error (RMSE). The model thus captures the coupled effects of jamming on link throughput and navigation observation quality while retaining the influence of satellite geometry. A joint satellite selection problem is formulated to maximize normalized communication benefits while penalizing positioning error, jamming exposure, and satellite coalition load. Feasible associations must satisfy visibility, the prescribed number of satellites per user, satellite service capacity, minimum link SINR, and maximum positioning-error requirements. The selection problem is then modeled as an OCFG, in which each satellite and its served users form a coalition and users may belong to multiple coalitions. The joint objective serves as the system potential function, with user payoffs defined by their marginal contributions. A jamming-aware reconfiguration algorithm (Algorithm 1) employs feasible single-satellite replacements guided by local potential gains. Each candidate switch is evaluated based on changes in the utilities of affected users and the loads of the outgoing and incoming satellite coalitions. Only switches with strictly positive potential gains are accepted, and the candidate with the largest gain is preferred. At each scheduling interval, the algorithm updates link and positioning metrics and repairs infeasible associations inherited from the preceding interval. Initial selection involves all users, whereas post-jamming reconstruction focuses on users meeting the SINR, JNR, or positioning-error trigger conditions. Iteration terminates when no improving switch remains or the prescribed iteration limit is reached. The marginal-contribution payoff definition yields an exact potential game because each unilateral payoff change equals the corresponding change in system potential. As the feasible strategy space is finite and each accepted switch strictly increases this potential, the switching process converges in finitely many steps to an exchange-stable coalition structure if continued until no improving move remains. During event-triggered reconstruction, stability applies only to the affected-user search set. Local replacement evaluation also avoids exhaustive enumeration of complete satellite subsets. Simulations are performed using Satellite Tool Kit (STK) and MATLAB with a 792-satellite Walker constellation and eight ground users. Jammer power ranges from 0 to 1000 W, with four to six selected satellites per user and service capacities of two to four users per satellite. The comparison methods comprise an integrated communication-navigation coalition formation game (ICAN-CFG) baseline, a joint SINR-GDOP heuristic, a maximum tetrahedral-volume method, and a non-reconfiguration strategy. Apart from representative snapshots and convergence plots, results are averaged over 5000 independent Monte Carlo trials. Results and Discussions The representative jammed snapshot shows that the proposed OCFG method replaces heavily interfered links with lower-interference links to visible satellites while maintaining favorable navigation geometry (Fig. 3 ). The potential-gain switching criterion achieves a higher system utility at convergence than self-interested and Pareto-constrained switching rules (Fig. 4 ), indicating the benefit of accounting for shared-resource competition while retaining sufficient switching flexibility. As jammer transmit power increases, the user-average bandwidth-normalized sum rate decreases and positioning RMSE increases for all evaluated methods. Nevertheless, the proposed method consistently achieves the highest communication rate and lowest positioning RMSE, with its communication advantage becoming more pronounced under strong interference (Fig. 5 ). These improvements reflect its joint consideration of communication link quality, interference-dependent navigation observations, and satellite sharing constraints. Satellite service capacity also influences reconstruction performance. Increasing the maximum number of users served per satellite from two to three improves both communication and positioning performance, whereas further increasing it to four yields diminishing benefits (Fig. 6 ). Increasing the number of selected satellites generally improves communication performance, but does not necessarily produce a corresponding improvement in positioning accuracy because the latter also depends on observation geometry, link quality, and resource competition (Fig. 7 ). These results demonstrate the importance of coordinating interference avoidance and satellite resource sharing in joint communication-navigation selection. Conclusions This paper develops an OCFG-based anti-jamming satellite selection method for multiuser LEO-ICAN networks under directional jamming. By incorporating interference-dependent communication and navigation performance together with satellite-load penalties, the proposed method enables association reconstruction through local potential-gain switching. The finite exact potential-game formulation establishes the convergence of strictly improving switching sequences to an exchange-stable coalition structure. Simulation results demonstrate improved communication rates and positioning accuracy compared with the evaluated baselines. Future work will investigate robust satellite reconstruction under time-varying interference, user mobility, and delayed state information.
Available online , doi: 10.11999/JEIT260599
Abstract:
Objective With the widespread deployment of WiFi in industrial Internet of Things (IoT), enterprise wireless local area networks, and smart home scenarios, security vulnerabilities have become increasingly prominent. Radio Frequency Fingerprint IDentification (RFFID) leverages hardware-induced variations introduced during manufacturing to authenticate devices, providing an effective solution for physical-layer security. However, existing RFFID methods often face challenges in balancing recognition robustness under dynamic channel conditions and computational efficiency for deployment on resource-constrained platforms. This paper aims to develop a lightweight and efficient RFFID method that achieves high recognition accuracy under varying channel conditions and device scales while maintaining low computational complexity for embedded systems. Methods This paper proposes WiFi-RF-LTCN, a lightweight temporal convolutional network for WiFi RFFID. The method processes the Legacy Long Training Field (L-LTF) sequence from the WiFi physical-layer preamble. First, multi-dimensional features, including I/Q components, amplitude, and phase, are extracted to form a four-channel input. The proposed network consists of two parallel branches: a temporal dilated convolution branch with residual connections to capture long-range temporal dependencies in the signal, and a multi-scale learnable convolution branch with different kernel sizes to enhance sensitivity to short-term local waveform distortions. Subsequently, a channel attention mechanism based on the squeeze-and-excitation block adaptively fuses features from both branches, highlighting discriminative RF fingerprint features while suppressing redundant information. Finally, the model is trained and evaluated on a simulated WiFi dataset incorporating diverse transmitter impairments and multipath channel effects. Results and Discussions Extensive experiments demonstrate that WiFi-RF-LTCN achieves superior performance under various conditions. It achieves recognition accuracy of 93.12% at 0 dB Signal-to-Noise Ratio (SNR) and 99.27% at 20 dB SNR. The model maintains robust performance with limited training data, achieving an accuracy of 82.21% with only 50 samples per device, and exhibits good scalability as the number of devices increases. Ablation experiments verify the effectiveness of each component, with the complete model achieving an average accuracy of 96.30% with an inference time of 0.41 ms. Compared with Transformer, ResNet50, TCN, 1D-CNN, and LSTM, WiFi-RF-LTCN achieves accuracy improvements of 2.84, 2.36, 1.40, 2.62, and 3.61 percentage points, respectively. Moreover, it requires only 0.17 M parameters, substantially reducing computational overhead compared with ResNet50 (23.54 M) and Transformer (0.85 M), whose inference times are 3.9381 ms and 0.8953 ms, respectively. Conclusions This paper presents a lightweight temporal convolutional network, WiFi-RF-LTCN, for WiFi device identification using RFFID. By integrating multi-scale feature extraction, temporal modeling, and channel attention, the proposed method effectively captures subtle hardware-induced distortions and demonstrates strong robustness to noise and channel variations. Experimental results verify that WiFi-RF-LTCN achieves a favorable balance between recognition accuracy and computational cost, outperforming existing deep learning methods. Its low parameter count and short inference time make it suitable for real-time deployment on resource-constrained embedded platforms, providing an effective solution for enhancing physical-layer security in IoT and other wireless applications.
Available online , doi: 10.11999/JEIT260503
Abstract:
Objective With the rapid expansion of data-driven services and increasing concerns over privacy protection, publishing privacy-preserving synthetic data has become an important task for balancing analytical utility and individual confidentiality. While numerous Differential Privacy (DP) data synthesis mechanisms have been proposed to generate high-utility datasets, they are mainly designed for single-user and single-release scenarios. Applying these methods to multi-user and multiple data publishing scenarios still faces two major challenges: (1) the lack of systematic privacy budget tracking and fine-grained management mechanisms, which makes cumulative privacy risks across multiple releases difficult to control and audit; and (2) the inability to customize synthetic data according to the specific analytical queries of different users, resulting in reduced utility for targeted analyses. Therefore, an effective framework is required to achieve hierarchical privacy budget management and on-demand data synthesis while ensuring both privacy protection and data utility in complex multi-user publishing scenarios. Methods A customizable differential privacy data synthesis framework, CuSyn, is proposed for multi-user and multiple data publishing scenarios. CuSyn integrates customization requirements into the standard select-measure-generate paradigm through well-defined interfaces and consists of two key components. First, a scalable pool-based privacy budget management strategy is designed to provide strict and traceable control over privacy expenditure. For each original dataset, a dedicated privacy budget pool is maintained and initialized with a predefined global limit, and a user-level privacy budget pool can be maintained for each authorized user to support differentiated access control. A BudgetEval function is developed to validate and allocate the budget of each synthesis request (Algorithm 1) through a configurable four-level budget constraint system, including the per-synthesis budget limit, the per-request user budget limit, the remaining budget of the dataset budget pool, and the remaining budget of the user budget pool. This hierarchical management mechanism provides multi-level control from the global system level to the user level and dataset level, ensuring that cumulative privacy leakage across all users and releases remains bounded even under worst-case collusion. When a request exceeds some constraints but remains partially feasible, the function adaptively negotiates the maximum admissible budget instead of directly rejecting the request. A formal privacy composition analysis is conducted under the zero-Concentrated Differential Privacy (zCDP) framework. Each approved budget is converted into a zCDP parameter, and the cumulative privacy loss of all synthesis operations is bounded by the sum of these parameters. Since the privacy budget pool is never refilled, the cumulative privacy loss cannot exceed the privacy guarantee derived from the predefined global budget and remains independent of user identities, making the guarantee valid even when users collude and combine their synthetic datasets. Second, a customizable synthesis mechanism is developed to extend the standard synthesis pipeline. Users submit user interest query sequences, and CuSyn first performs standard iterative query selection and measurement to capture general statistical characteristics, followed by additional measurements on user interest queries (Algorithm 2 and Algorithm 3). The noisy measurements from both stages are jointly used to estimate the final Probabilistic Graphical Model (PGM) and generate customized synthetic datasets. The allocated budget is shared between the iterative process and user interest query measurements, and the Gaussian noise scale is adjusted accordingly, enabling customization without additional privacy cost. For long user interest query sequences, the number of standard iterations can be reduced to avoid excessive budget dilution and maintain query accuracy. Results and Discussions The performance of CuSyn is evaluated on multiple real-world datasets with different characteristics and scales (Table 1) and compared with two state-of-the-art probabilistic graphical model (PGM)-based mechanisms, AIM and MWEM-PGM. Three experiments are conducted: varying the length of the user interest query sequence (LEN), varying the privacy budget (BGT), and simulating multiple data publishing by multiple users under a shared privacy budget pool (MUL), with the experimental settings summarized in Table 2. The user interest queries of each user are randomly selected from the workload containing all two-dimensional marginal queries of each dataset. In the LEN experiments (Fig. 2), CuSyn achieves lower error on user interest queries than the baseline mechanisms across most datasets and sequence length settings. For example, on the NLTCS dataset with a user interest query sequence length of 10, the error of user interest queries is reduced by approximately 67% compared with the baseline, while the error on the overall workload remains comparable to that of the baseline methods. In some configurations, such as the NLTCS dataset with a sequence length of 15, the overall workload error is also reduced substantially. However, increasing the number of user interest queries does not always reduce the error because the privacy budget is shared between the standard iterative phase and the user interest query measurement phase. When the user interest query sequence occupies a large proportion of the workload, the available budget for capturing general statistical characteristics decreases, resulting in higher overall workload error. Therefore, the sequence length should balance the utility of user interest queries and overall data availability. In the BGT experiments (Fig. 3), CuSyn maintains similar advantages under different privacy budget levels. The average workload error over all attributes decreases as the privacy budget increases, and the utility of user interest queries is higher than that of the baseline methods in most settings. For AIM, the adaptive budget allocation strategy reduces the influence of different privacy budget levels on performance, whereas CuSyn maintains improved utility on user interest queries while preserving overall workload accuracy. In the MUL experiments (Fig. 4 and Fig. 5), four independent users with different user interest query sequences are simulated under a shared total budget limit. CuSyn maintains its utility advantage on user interest queries over 5, 10, and 15 rounds of multiple data publishing, while the overall workload error remains comparable to that of the single-round synthesis case on most datasets. The results demonstrate that CuSyn provides an effective balance between user-specific query utility and overall synthetic data quality, confirming its capability to customize synthetic data for targeted analyses in multi-user and multiple data publishing scenarios. The current privacy budget allocation strategy is static, and all synthesis operations are treated as new releases. Future work will focus on adaptive privacy budget allocation strategies based on data sensitivity and user query patterns, as well as more efficient budget recovery mechanisms for incremental publishing. Conclusions A customizable differential privacy data synthesis framework, CuSyn, is proposed to address privacy-preserving data publishing in multi-user and multiple data publishing scenarios. The pool-based privacy budget management strategy provides a practical and provably secure approach for fine-grained hierarchical management of privacy expenditure and reduces the risk of cumulative privacy leakage. The customizable synthesis mechanism incorporates user interest queries into the data generation process, improving the utility of synthetic data for targeted analytical tasks. Experimental results demonstrate that CuSyn improves utility on user interest queries while maintaining comparable or better performance on the overall workload in both single-round and multiple data publishing scenarios. The proposed framework provides an effective solution for privacy-preserving synthetic data publishing in complex multi-user environments.
Available online , doi: 10.11999/JEIT261138
Abstract:
Objective With the rapid development of low-altitude intelligent networks, uncrewed aerial vehicles (UAVs) are increasingly used in inspection, logistics, and emergency response. These applications require continuous state sensing, wireless transmission, and interaction with digital twin (DT) models to support real-time decision-making. However, existing UAV DT systems generally separate sensing and communication, resulting in inefficient resource utilization, increased update latency, and limited synchronization accuracy. Integrated Sensing and Communication (ISAC) enables sensing and communication to share spectrum, transmit power, and hardware resources. Nevertheless, the coupling among UAV trajectory control, ISAC resource allocation, and DT synchronization performance remains insufficiently investigated. Therefore, this paper investigates the joint optimization of DT synchronization performance in ISAC-enabled low-altitude UAV systems. Methods An ISAC-enabled low-altitude UAV DT architecture consisting of the physical layer, ISAC module, and DT layer deployed on a mobile edge computing (MEC) server is established. Communication latency, sensing error, and DT mapping error models are developed to characterize the coupling among bandwidth allocation, transmit power, channel conditions, UAV trajectory, and synchronization performance. A joint optimization problem is formulated to minimize the long-term average DT mapping error subject to UAV mobility, wireless resource, and synchronization delay constraints. To solve the resulting non-convex problem, a collaborative framework integrating Digital Twin-aware Proximal Policy Optimization (DT-aware PPO) and Alternating Optimization (AO) is proposed. DT-aware PPO optimizes long-term UAV trajectory decisions, while the AO-based resource optimization module dynamically allocates bandwidth and transmit power at each time slot. The resulting DT synchronization performance is fed back to the PPO agent to coordinate long-term trajectory control and short-term resource allocation. Results and Discussions Simulation experiments are conducted on a MATLAB-based low-altitude UAV digital twin platform under a dynamic inspection scenario. Compared with the Non-ISAC-PPO scheme, the proposed method reduces the DT mapping error from 2.009 to 1.084 and the average communication delay from 13.102 s to 3.200 s (Fig. 4 ). Compared with the Random Trajectory and Equal Resource Allocation schemes, the proposed method achieves lower system objective value, synchronization delay, and DT mapping error by jointly optimizing UAV trajectory and wireless resources (Fig. 5 ). The average reward gradually converges during training, indicating stable learning of the DT-aware PPO policy (Fig. 6 ). Compared with Non-ISAC A*, ISAC-Greedy, and ISAC-RRT*, the proposed method achieves lower system cost, faster decision making, and improved performance recovery under dynamic disturbances (Fig. 7 ). Parameter sensitivity analysis further shows that increasing bandwidth and transmit power improves DT synchronization performance, whereas excessive task arrival rates degrade performance due to queue accumulation (Fig. 8 ). Conclusions This paper proposes a joint optimization method for DT synchronization in ISAC-enabled low-altitude UAV systems. A unified communication-sensing-DT coupling model is established to characterize the interactions among UAV trajectory, wireless resource allocation, communication latency, sensing accuracy, and DT synchronization performance. A collaborative framework integrating DT-aware PPO and AO-based resource allocation is developed for long-term trajectory control and short-term resource allocation. Simulation results demonstrate that the proposed method effectively reduces DT mapping error and communication latency while maintaining stable learning and a certain degree of adaptability to dynamic disturbances.
Available online , doi: 10.11999/JEIT261177
Abstract:
Significance Low-altitude embodied intelligence is becoming an important technological foundation for intelligent inspection, logistics delivery, security monitoring, emergency rescue, and other low-altitude applications. Unlike conventional unmanned aerial systems that mainly rely on preset routes, remote control, or isolated perception modules, low-altitude embodied agents are required to continuously perceive the surrounding environment, exchange information with other agents and infrastructure, perform task-oriented inference, generate decisions, execute control actions, and update their behaviors according to environmental feedback. Typical application scenarios include urban infrastructure inspection, dynamic target tracking, autonomous logistics delivery, and disaster-response operations (Fig. 1 ). Such tasks are characterized by open and highly dynamic operating spaces, rapidly varying air-ground channels and network topologies, heterogeneous multimodal sensing data, and strict constraints on onboard payload, energy, computing capability, and storage. Sensing errors, wireless transmission delays, computing latency, and control feedback delays are therefore strongly coupled and may accumulate along the task chain. Optimizing sensing accuracy, communication rate, or computing latency separately cannot guarantee that environmental information is transformed into reliable control actions within the required time. Integrated sensing, communication, and computing (ISCC) provides a promising pathway for coordinating information acquisition, transmission, processing, and decision-making under shared spectrum, power, energy, and computing-resource constraints. Existing surveys mainly discuss network architectures, functional integration, resource management, or intelligent enhancement, whereas the complete autonomous task process of low-altitude embodied intelligence has received less systematic attention. Therefore, this review takes the closed-loop task process as the main thread and examines ISCC from the perspective of the complete perception–transmission–computation–decision–control–feedback chain, with particular attention to task-level metrics such as closed-loop latency, information freshness, task success rate, tracking performance, and control stability. Progress Research on integrated sensing, communication, and computing (ISCC) is evolving from pairwise functional integration toward task-oriented coordination. Communication–sensing integration has progressed from spectrum coexistence to waveform and hardware sharing, bidirectional assistance, multimodal perception, and networked cooperative sensing (Fig. 2 ). Communication–computation integration has developed from conventional cloud and edge offloading toward joint communication–computing design, over-the-air computation, federated edge learning, and distributed intelligence, reflecting a transition from serial transmission-and-computation processing toward deeper functional integration (Fig. 3 ). Sensing–computation integration further supports multimodal fusion, state estimation, feature extraction, and task-oriented information selection. These developments collectively provide the technical basis for coordinating perception, transmission, and processing around unified task objectives.For low-altitude embodied intelligence, a cloud–edge–device cooperative architecture can be adopted to support the closed-loop process of environmental perception, information transmission, computation and inference, decision generation, control execution, and state feedback (Fig. 4 ). Onboard nodes undertake time-critical functions such as local state estimation and emergency response, edge nodes support cooperative perception and latency-sensitive inference, and cloud platforms provide global coordination, model training, and long-term planning. Since control actions continuously change sensing viewpoints, wireless channels, network topology, and computing demands, the different functional modules are dynamically coupled throughout the task process. Meanwhile, the exchanged information is evolving from raw sensing data and complete tasks toward intermediate features, model parameters, and task-relevant semantic information, thereby reducing redundant transmission and processing while preserving information that is useful for perception, inference, and control.Current research can be summarized into four main directions. First, environmental information acquisition is shifting from single-node sensing toward multimodal and multi-agent cooperative perception, where complementary observations from aerial agents, ground infrastructure, and heterogeneous sensors are fused to improve environmental awareness. Second, sensing-task offloading is evolving from coarse-grained migration of complete tasks toward fine-grained cloud–edge–device collaborative inference. Model partitioning and intermediate-feature transmission allow inference workloads to be dynamically distributed according to channel quality, computing load, and task deadlines. Representative results show that joint optimization of deep neural network partitioning, sensing/communication beamforming, and multi-tier computing resources can reduce inference latency by more than 55% compared with local inference under the considered setting (Fig. 5 ). Third, multidimensional resource coordination jointly considers spectrum, transmit power, beamforming, computing resources, task partitioning, and UAV trajectories. Multi-resource coordination can achieve lower sensing-task completion latency than schemes optimizing only part of the available resources, especially under higher sensing-data compression ratios (Fig. 6 ). Finally, system optimization is extending toward decision and control feedback, with increasing attention to closed-loop latency, information freshness, tracking error, control stability, task success rate, and energy consumption. Overall, ISCC is gradually moving from resource sharing and functional integration toward task-driven optimization of the complete perception–transmission–computation–decision–control–feedback chain. Conclusions Low-altitude embodied intelligence requires an end-to-end transition from function-oriented integration to task-driven closed-loop coordination. Sensing acquires task-related environmental and agent states; communication exchanges sensing data, cooperative information, intermediate features, and control commands; and computation transforms the acquired information into environmental understanding, task inference, and decision support. Decisions are then converted into control actions that change the physical state of aerial agents and, in turn, reshape the conditions of subsequent sensing, transmission, and computation. Therefore, the essential role of ISCC is not simply to combine three functions or improve their individual efficiencies, but to coordinate the complete closed-loop process according to the final task objective. Task deadlines and control risks should determine sensing-update intervals, information granularity, offloading locations, communication priorities, and computing-resource allocation, while control-induced state variations should be fed back to subsequent sensing modes, link selection, task placement, and resource scheduling. The reviewed studies further indicate that low-altitude ISCC is evolving along three related dimensions: from pairwise integration to multi-functional coordination, from static single-node processing to cloud–edge–device and multi-agent cooperation, and from module-level performance metrics to task-level closed-loop objectives. This evolution provides a technical basis for supporting safer, more autonomous, and more scalable operation of low-altitude embodied agents. Prospects Several challenges remain before large-scale deployment of low-altitude embodied intelligence can be supported. First, unified sensing–communication–computation–control models are required to characterize the propagation and coupling of sensing uncertainty, communication unreliability, computing and inference errors, control feedback delays, and system dynamics within a common framework. Environmental state, information freshness, semantic value, inference accuracy, control cost, and system stability should be jointly represented so that task requirements can be mapped backward to sensing, communication, and computing constraints. Second, real-time coordination under highly dynamic environments remains difficult because aerial mobility, intermittent blockage, interference, changing network topology, and time-varying computing loads can rapidly invalidate static resource allocation. Predictive scheduling, lightweight edge inference, dynamic model partitioning, elastic cloud–edge–device migration, digital twins, and online learning are promising directions for rapid adaptation under incomplete information and resource fluctuations. Third, security, privacy, and trustworthy intelligence need to be incorporated throughout sensing, transmission, inference, and control. Identity authentication, trusted access, privacy-preserving computation, anomaly detection, uncertainty evaluation, interpretable decision-making, and secure control should be jointly considered because attacks on data, models, wireless links, or control commands may directly affect physical safety. Finally, large-scale heterogeneous swarms require scalable hierarchical and distributed coordination. As the number of agents increases, communication overhead, state-space dimension, and decision complexity grow rapidly, while purely centralized solutions face bandwidth bottlenecks, single-point failures, and excessive computational burden. Hierarchical and distributed architectures, graph learning, multi-agent reinforcement learning, swarm intelligence, semantic communication, event-triggered interaction, and knowledge sharing may reduce redundant information exchange and improve self-organization, self-recovery, and task reconfiguration.
Available online , doi: 10.11999/JEIT260840
Abstract:
Objective Biased source data can be compressed before channel transmission to reduce the equivalent information rate. At finite block-lengths, however, BP decompression may fail for some source frames. In conventional compression-assisted transmission, a compressed sequence is typically passed directly to channel coding and transmission without verifying whether the current source frame can be correctly recovered by the adopted BP decompression decoder. To address this issue, this paper proposes an adaptive low-density parity-check (LDPC)-coded transmission scheme driven by transmitter-side decoding decisions. Methods The proposed scheme employs LDPC codes for source compression and channel protection. After source compression, the transmitter performs local belief propagation (BP) decompression decoding using the compressed sequence and source prior information. If local decompression succeeds, the compressed sequence is channel-coded and transmitted; otherwise, the original source sequence is channel-coded and transmitted. These two alternatives are referred to as the compressed and uncompressed transmission schemes, respectively. The transmission scheme is selected on a frame-by-frame basis according to the transmitter-side local decompression result. At the receiver, two decoding algorithms are developed. Parallel decoding performs BP decoding under both transmission-scheme hypotheses, whereas threshold decoding determines the transmission scheme from the received soft information and then performs only one LDPC channel decoding operation. A weighted combination bound is proposed to estimate the frame error rate (FER), and a genie-aided bound is introduced for comparison. Results and Discussions Under the simulated conditions, the proposed scheme with either parallel decoding or threshold decoding achieves a lower FER than the uncompressed transmission scheme. The resulting FERs agree well with the weighted combination bound and approach the genie-aided bound. The results further show that the probability of selecting the compressed transmission scheme decreases markedly as θ increases. The corresponding probabilities are approximately 0.935, 0.666, and 0.030 for \begin{document}$ \theta=0.07 $\end{document} \begin{document}$ 0.1 $\end{document} \begin{document}$ \theta=0.07 $\end{document} \begin{document}${\text{SNR}}=-1\text{dB}$\end{document} 0.0157 and 0.0143 , respectively. Conclusions The proposed scheme performs transmitter-side local BP decompression to adaptively select between the compressed and uncompressed transmission schemes on a frame-by-frame basis. Its performance advantage is more pronounced for strongly biased sources and when the transmitter-side local BP decompression success probability is high. Its implementation requires one additional local BP decompression operation at the transmitter, while the decision threshold used in threshold decoding needs to be determined for the given source statistics, channel SNR, and coding parameters.
Available online , doi: 10.11999/JEIT260907
Abstract:
Objective Featuring superior material properties including high thermal conductivity, high critical breakdown electric field, and high electron saturation velocity, 4H-SiC MOSFETs break through the performance bottlenecks of conventional silicon-based MOSFETs and satisfy the core requirements of power electronic systems for high frequency, miniaturization, and high efficiency. Among them, the double-trench 4H-SiC MOSFET (DT-MOS) developed by ROHM has been widely adopted in power electronic systems such as three-phase inverters for new energy vehicles and wind power generation. Nevertheless, the conventional DT-MOS (CON-DTMOS) suffers from prominent performance limitations. Restricted by the wide-bandgap characteristics of 4H-SiC materials, its parasitic body diode exhibits poor reverse freewheeling capability and high conduction loss, accompanied by severe bipolar degradation. In addition, the drawbacks of CON-DTMOS, including large on-resistance, high gate-drain capacitance, and excessive reverse recovery charge, further limit its application scope. To address the aforementioned performance deficiencies of CON-DTMOS, this paper proposes two novel device structures, namely LV-DTMOS and RC-ASMOS. By introducing planar channels and reverse channels, the proposed structures reduce the on-resistance and reverse turn-on voltage and effectively suppress the bipolar degradation induced by the body diode. Meanwhile, the optimized trench structure shields the gate-drain capacitance and improves the dynamic characteristics of the device. Consequently, the optimized 4H-SiC MOSFET with comprehensive and balanced performance is realized, which completely eliminates the bipolar degradation effect and lays a solid foundation for the reliable application of 4H-SiC MOSFETs in high-frequency power electronic scenarios. Methods This paper carries out device structure design and parameter optimization based on the Sentaurus TCAD simulation platform. Taking the conventional double-trench 4H-SiC MOSFET (CON-DTMOS) as the research baseline, a novel LV-DTMOS structure with integrated Lateral-Vertical channels is firstly designed. On this basis, an asymmetric stepped RC-ASMOS with embedded reverse channels is further proposed through structural optimization.The forward channels CH1 and CH2 can significantly reduce the on-resistance and improve the chip integration level. Moreover, the CH2 integrated at the trench bottom can shield the gate-drain coupling capacitance, thereby enhancing the dynamic performance of the device. Meanwhile, the embedded reverse channel CH3 can be turned on at a low reverse turn-on voltage through an ultra-thin oxide layer prior to the intrinsic body diode. This avoids the electron-hole recombination process and fundamentally eliminates the bipolar degradation effect.To improve the accuracy and reliability of simulation results, multiple critical physical models are adopted in the simulation process, including the incomplete ionization model, bandgap narrowing model, avalanche effect model, and high-field mobility saturation model. Furthermore, an interface charge model is introduced to account for the influence of interface state charges at the SiC/SiO2 interface on channel mobility. The comprehensive application of the above multi-physical models enables the establishment of a simulation model that is highly consistent with the physical behaviors of practical devices, which provides a reliable support for subsequent device characteristic analysis and parameter optimization. Results and Discussions Simulation results indicate that the RC-ASMOS achieves optimal comprehensive performance under the structural and doping parameters of a CSL layer doping concentration of 4.5×1017 cm–3, a reverse channel width of 0.35 μm, a source oxide thickness of 0.012 μm, and a P-body doping concentration of 2.0×1019 cm–3. Compared with the conventional CON-DTMOS, the optimized RC-ASMOS exhibits substantial improvements in overall performance. The specific on-resistance of the RC-ASMOS is reduced by 29.1%, and the forward voltage drop is decreased from 2.8 V to 1.6 V, corresponding to a reduction of 42.9%. In addition, the proposed device effectively optimizes the internal electric field distribution and alleviates the electric field crowding effect, leading to a 36.5% enhancement in breakdown voltage (Fig. 11 , Fig. 12 ). Benefiting from the CH2 integrated at the trench bottom, the gate-drain coupling effect is effectively weakened, which significantly reduces the gate-drain capacitance and lowers the gate charge by 53.9% (Fig. 13 , Fig. 14 . Furthermore, the RC-ASMOS realizes unipolar reverse freewheeling through the embedded reverse channel CH3, which suppresses the minority carrier storage effect and greatly improves the reverse recovery characteristics. The reverse recovery charge is reduced by 46.2%, and the peak reverse current is remarkably decreased (Fig. 15 ). Consequently, the proposed RC-ASMOS possesses excellent static conduction performance and high-frequency dynamic switching performance simultaneously. Conclusions Based on the conventional double-trench 4H-SiC MOSFET (CON-DTMOS), this paper conducts device design, parameter optimization, and performance verification using the Sentaurus TCAD simulation platform. Firstly, an LV-DTMOS structure with integrated planar-vertical channels is simulated. On this basis, structural optimization is further performed, and a novel RC-ASMOS device with embedded reverse channels is proposed. By optimizing the internal carrier transport paths, the RC-ASMOS achieves significant improvements in both static and dynamic device performance.Thanks to the unique reverse channel, the RC-ASMOS can be turned on for reverse freewheeling prior to the parasitic body diode, which avoids the internal electron-hole recombination process and fundamentally eliminates the bipolar degradation effect. Meanwhile, the RC-ASMOS remarkably reduces the gate-drain overlapping area, yielding a substantial improvement in dynamic characteristics. Overall, the proposed RC-ASMOS realizes comprehensive performance optimization and effectively breaks through the performance limitations of traditional 4H-SiC MOSFETs.
Available online , doi: 10.11999/JEIT260860
Abstract:
Objective Magnetoencephalography (MEG) is a non-invasive technique for detecting weak magnetic fields generated by neuronal activity. Compared with conventional superconducting quantum interference device-based MEG systems, optically pumped magnetometer-based magnetoencephalography (OPM-MEG) allows sensors to operate at room temperature and to be placed closer to the scalp, thereby improving signal detection efficiency and supporting wearable system design. In practical OPM-MEG systems, rigid helmets can provide stable sensor positions and improve repeated measurement consistency. However, because head size varies across subjects, a fixed-size rigid helmet may increase the sensor-to-scalp distance for smaller heads, resulting in reduced forward-field strength and limited detectable coverage. Existing OPM-MEG helmet studies have mainly focused on system implementation and experimental validation, whereas quantitative methods for evaluating the size-adaptation range of fixed-size rigid helmets remain insufficient. This study proposes a standard-head-model-based design and multidimensional quantitative evaluation method for an OPM-MEG rigid helmet, aiming to provide guidance for helmet standardization, size planning, and engineering application. Methods The ICBM152 standard head model was used as the geometric reference for helmet design. The scalp surface was extracted and truncated according to the brain coverage requirement, and a smooth inner-wall surface was generated. Following the principles of maintaining basic wearing clearance, reducing the sensor-to-cortex distance, and ensuring whole-brain coverage, a 64-channel OPM-MEG sensor array was arranged to cover the frontal, parietal, temporal, and occipital regions. Based on the inner-wall surface and sensor positions, a rigid helmet structure with sensor slots and ear openings was constructed, considering structural stability, 3D-printing feasibility, and wearing comfort. To evaluate the size-adaptation range, the standard head model was defined as the 1.0-scale model, and six head-size models with scale factors of 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0 were generated. During scaling, the helmet and sensor array remained fixed, while the head model was proportionally resized with the top of the head kept unchanged. A 306-channel SQUID-MEG system was used as the reference baseline. The Frobenius norm of the lead field was used to evaluate forward-field detection capability, while undetectable voxels were used to quantify effective detectable coverage. Mean ratio, median ratio, spatial coverage rate, undetectable-voxel ratio, Jaccard similarity, newly undetectable voxels, and recovered detectable voxels were calculated for comparison. Results and Discussions The simulation results showed that head-size mismatch had a clear influence on the detection capability of the fixed-size OPM-MEG rigid helmet. As the head scale increased from 0.5 to 1.0, the Frobenius norm over the cortical surface increased continuously, indicating enhanced forward-field strength when cortical sources became closer to the fixed sensor array. The mean Frobenius norm increased from 1.81×10−9 at the 0.5-scale model to 2.82×10−9 at the 1.0-scale model, corresponding to 88.9% and 139.0% of the SQUID-MEG reference baseline, respectively. The median ratio showed a similar increase from 87.1% to 139.1%. Based on linear interpolation, the mean and median Frobenius norm ratios reached the SQUID-MEG reference level at head-scale factors of approximately 0.66 and 0.69, respectively. However, when spatial coverage was considered, a stricter threshold was obtained: a head-scale factor of approximately 0.89 was required for 95% of cortical source points to reach at least 95% of the reference detection capability. This indicates that global statistical indicators may reflect the average detection capability, whereas spatial coverage can more sensitively reveal local performance degradation. The undetectable-voxel analysis provided complementary evidence for detectable coverage. As the head scale decreased from 1.0 to 0.5, undetectable regions gradually expanded, mainly in lower and peripheral areas of the region of interest. The number of undetectable voxels increased from 1809 at the 1.0-scale model to 4396 at the 0.5-scale model, corresponding to 11.56% and 28.10% of all voxels, respectively. In contrast, the SQUID-MEG reference system had 5063 undetectable voxels, accounting for 32.36%. Therefore, within the tested range of 0.5–1.0, the designed OPM-MEG rigid helmet maintained fewer undetectable voxels than the SQUID-MEG baseline. Nevertheless, newly undetectable voxels appeared when the head scale decreased to 0.6 and 0.5, indicating local degradation of detectable coverage. These results suggest that the size-adaptation range of a fixed-size rigid helmet is not a single fixed value, but depends on the selected evaluation criterion and application scenario. Conclusions This study proposes a standard-head-model-based design and quantitative evaluation method for the size-adaptation range of an OPM-MEG rigid helmet. A 64-channel rigid helmet was designed using the ICBM152 standard head model, and its performance was evaluated under different head-size conditions by combining forward-field strength and detectable-coverage metrics. When the mean and median Frobenius norm ratios were used as overall performance criteria, the lower adaptation limits were approximately 0.66 and 0.69 times the standard head model. When spatial coverage was used as a stricter criterion, the lower limit increased to approximately 0.89. Based on the undetectable-voxel analysis, the helmet maintained better overall detectable coverage than the SQUID-MEG reference within the tested range, while the preferred engineering adaptation limit was approximately 0.7. These findings provide quantitative guidance for OPM-MEG rigid helmet standardization, size planning, applicable population analysis, and engineering application.
Available online , doi: 10.11999/JEIT260845
Abstract:
Objective With the continuous advancement of 6G and integrated sensing and communication (ISAC) technologies, wireless communications are progressively moving towards high-mobility scenarios. In such environments, traditional orthogonal frequency division multiplexing (OFDM) signals are susceptible to Doppler shifts, which compromise the orthogonality of subcarriers and introduce inter-carrier interference, thereby degrading communication performance. OTFS signals, constructed on a discrete delay-Doppler two-dimensional grid, exhibit robust resilience to time-varying channels because each symbol experiences nearly the same channel effect during transmission. However, the limited delay and Doppler resolutions of OTFS waveforms inevitably lead to fractional delays and fractional Doppler shifts in practical applications, which cause dispersion effects that impair both communication capabilities and target detection performance in ISAC systems. This paper aims to mitigate the dispersion effects in OTFS signals and enhance the stability and reliability of ISAC systems through waveform and signal processing design. Methods This paper employs a low-sidelobe prototype filter from the filter bank multicarrier (FBMC) family to modulate the OTFS signal. The prototype filter effectively suppresses Doppler sidelobes and mitigates Doppler dispersion effects, while introducing only marginal mainlobe widening compared with an OTFS waveform carrying the same number of data symbols. A guard interval is established between communication data symbols and radar pilots in the delay-Doppler domain to prevent mutual interference, thereby establishing the FBMC-OTFS signal model. In channel modeling, fractional delays and fractional Doppler shifts are jointly considered to accurately characterize the practical channel. Furthermore, drawing on the constant false alarm rate (CFAR) algorithm widely used in radar detection, channel estimation is performed for the FBMC-OTFS system in the delay-Doppler domain, improving target detection accuracy while reducing computational complexity. Finally, the maximum ratio combining (MRC) algorithm is applied to signal detection, iteratively eliminating inter-symbol interference to reduce the bit error rate. Results and Discussions This paper presents both simulation and field experiment results in multi-path channel scenarios, comparing the proposed FBMC-OTFS waveform with traditional OTFS waveforms. The results show that when the power difference between the strong and weak targets is 18.5 dB, the FBMC-OTFS signal can still detect the weaker target, whereas the OTFS waveform fails to do so. When both waveforms successfully detect all targets (at a 9 dB strong-to-weak target power difference), FBMC-OTFS achieves a significantly lower bit error rate (BER) than OTFS at an SNR of 20 dB. Moreover, the Doppler sidelobes of the FBMC-OTFS signal are nearly imperceptible, while those of the OTFS signal are distinctly visible: the peak sidelobe ratio (PSLR) and integrated sidelobe ratio (ISLR) of FBMC-OTFS are reduced to −29.82 dB and −22.15 dB, respectively. This sidelobe suppression is achieved at the cost of a reduced spectral efficiency, which is approximately one eighth of that of conventional OTFS. Field experiments further validate the feasibility of the FBMC-OTFS waveform in practical ISAC scenarios. Conclusions This paper proposes an FBMC-OTFS waveform tailored for ISAC applications, along with its corresponding signal processing methods. By modulating the OTFS waveform with the PHYDYAS prototype filter, the proposed approach effectively suppresses Doppler dispersion effects and mitigates the mutual interference between communication data and radar pilots, enabling reliable weak-target detection. The established channel model jointly accounts for fractional delays and fractional Doppler shifts. The CFAR algorithm, originally developed for radar detection, is applied to channel estimation, achieving robust path detection while reducing computational complexity, and is complemented by an MRC-based signal detector. Simulation and field experiments demonstrate that, compared with traditional OTFS, FBMC-OTFS significantly reduces Doppler sidelobes and can detect weak targets that OTFS fails to resolve, at the cost of a reduced spectral efficiency of about one eighth. The results provide a viable waveform solution for ISAC applications in 6G high-mobility scenarios.
Available online , doi: 10.11999/JEIT260349
Abstract:
Objective Specific Emitter Identification (SEI) distinguishes individual emitters by exploiting Radio Frequency Fingerprints (RFFs) and has important applications in military reconnaissance and Internet of Things (IoT) authentication. However, conventional SEI methods generally assume fixed modulation parameters, whereas practical communication systems often employ dynamically varying modulation schemes. Existing Domain Adaptation (DA) methods for variable-modulation SEI require target-domain data retraining and still suffer from insufficient robustness of extracted features under unknown modulation variations. To address these problems, this paper introduces Domain Generalization (DG) into variable-modulation SEI and proposes a Modulation-Information-Guided Dual-branch Training Framework (MIG-DTF) with a Testing-stage Model Adaptation strategy (TMA). The proposed framework aims to learn modulation-invariant RFFs from source-domain data and improve SEI performance under unknown modulation parameter variations without large-scale target-domain retraining. Methods The proposed method consists of training and testing stages. During the training stage, MIG-DTF employs an emitter identification main branch and a modulation classification auxiliary branch for collaborative optimization. The main branch utilizes cosine aggregation loss to enhance the intra-class compactness and inter-class discriminability of RFFs. Meanwhile, the auxiliary branch maximizes cross-prediction entropy to suppress modulation-related information contained in the RFFs extracted by the main branch, thereby encouraging the model to learn modulation-invariant fingerprint features. During the testing stage, the proposed TMA strategy performs one-step fine-tuning of the Batch Normalization (BN) layer parameters using test samples under an entropy minimization objective. By adapting BN statistics and affine parameters to the testing distribution, TMA alleviates feature distribution drift caused by modulation mismatch and further improves identification accuracy. Results and Discussions Experiments are conducted on a real-world variable-modulation dataset containing three typical variation scenarios: modulation type, carrier frequency, and symbol rate. The proposed method achieves identification accuracies of 79.31% (Table 1 ), 87.39% (Table 2 ), and 88.33% (Table 3 ) in the three scenarios, respectively, outperforming mainstream DG methods. Ablation experiments demonstrate that both MIG-DTF and TMA independently contribute to performance improvement, and their combination provides further gains (Fig. 4 ). Feature visualization results show that the proposed method obtains more compact intra-class distributions and clearer inter-class boundaries under different modulation conditions (Fig. 5 ). Correlation analysis between extracted features and modulation labels further verifies that the proposed method reduces the dependence of RFFs on modulation information while preserving emitter-discriminative characteristics (Table 5 ). In anti-noise experiments with Signal-to-Noise Ratio (SNR) values from 10 dB to 30 dB, the proposed method consistently outperforms the baseline method and achieves an accuracy improvement of approximately 10% at 10 dB SNR (Fig. 7 ), demonstrating stronger robustness under noisy conditions. In addition, overhead analysis shows that TMA requires only a small number of parameter updates and eliminates the need for target-domain training data, reducing adaptation costs for practical deployment. Conclusions This paper proposes a robust SEI framework for variable modulation scenarios by combining the training-stage MIG-DTF and testing-stage TMA strategies. MIG-DTF enables the extraction of modulation-invariant RFFs by separating emitter-related fingerprint information from modulation-related information, whereas TMA provides efficient adaptation to unknown testing distributions with low computational overhead. Experimental results on a real-world dataset demonstrate that: (1) the proposed method achieves higher identification accuracy than mainstream DG methods under multiple modulation parameter variations; (2) the extracted features exhibit stronger modulation invariance and emitter discrimination capability; (3) the proposed method maintains better performance under noise interference; and (4) TMA enables efficient adaptation without target-domain retraining. Future work will investigate more complex scenarios, including joint variations of multiple modulation parameters and time-varying channel conditions.
Available online , doi: 10.11999/JEIT260760
Abstract:
Objective Non-Orthogonal Multiple Access (NOMA) and Integrated Sensing And Communication (ISAC) are regarded as promising techniques for future wireless networks. Pinching Antenna (PA) systems provide a new reconfigurable transmission architecture by generating controllable radiation points along a dielectric waveguide, offering a new approach for NOMA-ISAC system design. However, existing studies mainly focus on PA-assisted NOMA, PA-assisted ISAC, or fixed-position antenna activation in PA-assisted NOMA-ISAC systems, whereas the joint design of power allocation and continuous PA positions in PA-assisted downlink NOMA-ISAC systems remains insufficiently investigated. The coupling among NOMA interference, sensing-signal detection, successive interference cancellation, and PA position design has not been fully considered. Therefore, a PA-assisted downlink NOMA-ISAC system is investigated, where the system sum rate is maximized under communication Quality-of-Service (QoS) requirements and sensing performance constraints. Methods For the PA-assisted downlink NOMA-ISAC system, a joint sum-rate maximization problem is formulated under communication QoS requirements and sensing performance constraints. The power allocation coefficients and PA positions are jointly optimized to maximize the system sum rate. Since the optimization variables are highly coupled and the PA minimum-spacing constraint results in a discontinuous feasible region, the formulated problem is non-convex. To solve this problem, an Alternating Optimization (AO) scheme is proposed. The original problem is decomposed into two subproblems: power allocation optimization and PA position optimization. The power allocation subproblem is solved using Successive Convex Approximation (SCA), whereas the PA position subproblem is solved using a partitioned-interval search method. Results and Discussions The proposed method converges within a few iterations, and its sum-rate performance is close to that of the exhaustive-search benchmark while clearly outperforming the bisection-based benchmark (Fig. 2 ). The system sum rate increases with the number of PAs and gradually approaches saturation (Fig. 3 ). This is because additional PAs provide more spatial degrees of freedom, whereas the power allocated to each PA decreases under a fixed total transmit power. Compared with the baseline schemes, the proposed scheme achieves a higher sum rate, demonstrating that the performance improvement results from the joint optimization of power allocation and PA positions rather than single-variable optimization. Under different power budgets, the proposed method maintains performance close to exhaustive search and achieves stable advantages over other algorithms (Fig. 4 ). These results demonstrate that the proposed method provides an effective performance-complexity tradeoff for PA-assisted downlink NOMA-ISAC transmission. The sensitivity analysis further verifies that the proposed method remains close to exhaustive search under different sensing-signal cancellation SINR thresholds (Fig. 5 ). Conclusions A PA-assisted downlink NOMA-ISAC system is investigated in this paper. By jointly optimizing the power allocation coefficients and PA positions, a system sum-rate maximization problem is formulated under communication QoS requirements and sensing performance constraints. To address the non-convexity and discontinuous feasible region caused by the strong coupling among optimization variables and the PA minimum-spacing constraint, an AO-based solution framework is proposed. In this framework, the power allocation subproblem is handled by SCA, whereas the PA position optimization subproblem is solved using a partitioned-interval search method. Simulation results show that the proposed method converges within a small number of iterations, achieves performance close to exhaustive search with lower computational complexity, and outperforms the baseline schemes. These results verify the effectiveness of the joint design of power allocation and PA positions. Future work will consider multi-user and multi-target scenarios, imperfect channel state information, and more accurate sensing echo models to improve the applicability of PA-assisted NOMA-ISAC systems in complex practical environments.
Available online , doi: 10.11999/JEIT260679
Abstract:
Objective The new-generation anti-jamming satellite communication systems in China require frequency-hopping signal generation technologies that simultaneously provide ultra-wide bandwidth, fast hopping rates, and high stability. Conventional electronic schemes are restricted by the “electronic bottleneck,” which limits the realization of wideband coverage and high-precision frequency control. Existing photonic frequency-hopping methods still face challenges in frequency stability, hopping frequency number, and spurious suppression. To overcome these limitations, a microwave photonic ultra-wideband high-precision high-stability frequency-hopping signal generation method based on acousto-optic frequency shifting is proposed. Methods A continuous optical wave generated by a narrow-linewidth laser is used as the optical carrier and equally divided into two paths by an optical splitter. In the upper path, a Dual-Parallel Mach-Zehnder Modulator (DPMZM) performs Carrier-Suppressed Single-SideBand (CS-SSB) modulation on the initial Frequency-Hopping (FH) signal, converting the Radio Frequency (RF) signal into the optical domain for subsequent transmission and processing. The generated single-sideband signal is injected into an Optical Frequency Comb (OFC) generation module with eight lines for signal replication. In the lower path, the optical carrier is frequency-shifted by an Acousto-Optic Modulator (AOM) to adjust the center frequency and then introduced into an identical OFC generation module to generate an eight-line Local Oscillator (LO) optical comb. The two optical combs are combined and separated into eight independent channels through a Dense Wavelength Division Multiplexer (DWDM). Each channel is connected to an optical switch and a photodetector for channel selection and signal detection. By jointly controlling the optical switching states and signal phase selection, the input narrowband FH signal is expanded to 16 times its original bandwidth. Results and Discussions An experimental link is established to verify the proposed method. The generation performance of the dual optical combs is first evaluated. The results shown in Fig. 2 demonstrate that the Free Spectral Ranges (FSRs) of OFC1 and OFC2 are 40 GHz and 36 GHz, respectively, and the flatness of both combs is below 3 dB. Subsequently, a phase modulation experiment is performed using a single-tone signal with a center frequency of 800 MHz. The results in Fig. 3 show that frequency-hopped signals are generated at 200 MHz and 1 800 MHz, and the Crosstalk Suppression Ratio between the two channels exceeds 21 dB, indicating effective channel isolation. Fig. 4 further presents the frequency-hopping performance of all channels using optical switches. The results demonstrate that an input FH signal with a bandwidth of only 1 GHz can achieve a 16-fold expansion of hopping bandwidth. Finally, a wideband vector signal is applied to verify the frequency-hopping capability of all channels. The experimental results show that the Adjacent Channel Power Ratio (ACPR) of each channel is approximately 39.57 dB, and the Error Vector Magnitude (EVM) of the generated signal is about 6.72%. Although residual signals remain detectable at the original spectral positions after frequency hopping, they have a negligible influence on effective communication. Conclusions The proposed method achieves large-bandwidth frequency hopping while maintaining high frequency stability. Compared with existing photonic frequency-hopping approaches based on optical injection locking or Fourier Domain Mode Locking, the proposed method enables flexible frequency hopping without directly driving the laser using an electrical signal. The frequency stability mainly depends on the initial FH signal, reducing frequency instability caused by laser parameter control. Unlike photonic frequency-multiplication-based schemes, the proposed method does not require a high-frequency hopping signal as a reference. In addition, the hopping bandwidth can be further expanded by increasing the number of optical frequency comb lines. Compared with other optical frequency comb-based photonic frequency-hopping schemes, the proposed method doubles the utilization efficiency of each comb line for frequency hopping and supports flexible frequency switching within the hopping bandwidth without being restricted by fixed frequency points. The proposed method provides an effective solution for future anti-jamming communications, reconnaissance, and electronic countermeasure systems.
Available online , doi: 10.11999/JEIT260610
Abstract:
Objective With the large-scale deployment of large model inference in cloud and edge scenarios, FPGA has emerged as an important heterogeneous acceleration platform that complements GPU due to its reconfigurable computing capability, high resource efficiency, and low inference latency. However, efficient Softmax implementation on FPGAs still faces three major challenges: (1) rounding-error accumulation and numerical drift caused by long-sequence denominator summation, (2) significant utilization degradation of fixed-parallelism pipelines when processing misaligned tails of variable-length vectors, and (3) throughput limitations caused by the imbalance between memory bandwidth and computational density. Methods This paper proposes QuadPipe, a high-throughput Softmax hardware pipeline that integrates three complementary design strategies. (1) A Q48.47-format 96-bit extended-precision accumulator is introduced into the denominator summation path. This design reduces the upper bound of per-addition rounding error from O(2–24) under FP32 to O(2–47), thereby suppressing error propagation during long-sequence denominator accumulation. (2) A hardware-level boundary-agnostic processing mechanism based on negative-infinity padding is developed. By exploiting the property \begin{document}$ \mathrm{e}^{-\mathrm{\infty}}=0 $\end{document} Results and Discussions The proposed QuadPipe accelerator is implemented on a Xilinx Kintex-7 325T development board. At 150 MHz, it achieves a 59-cycle pipeline depth, a steady-state throughput of 4 FP32 results per cycle, and a peak throughput of 600 M elements/s, while consuming only 4.72% of the device LUT resources and 8 DSPs (Table 1). The Figure Of Merit (FOM) reaches 1.14×106 bit·s–1·(LUT+FF)–1 ( Table 1). Board-level measurements show that the Mean Absolute Error (MAE) remains within the 10–11～10–10 range across all tested sequence lengths (Table 2). At N=2048, the Q48.47 extended-precision accumulation reduces the denominator error by 87.59% compared with standard FP32 accumulation (Table 2). Compared with the software pre-padding scheme, the hardware boundary-agnostic processing mechanism improves the effective element throughput of misaligned vectors by up to 6.49% at N=1 025 and reduces batch-processing latency by up to 5.98% (Table 3). Conclusions The proposed QuadPipe accelerator integrates extended-precision accumulation, boundary-agnostic processing, and dual-dimensional batch scheduling within a unified pipeline framework to address three major challenges in FPGA-based Softmax acceleration: long-sequence numerical stability, variable-length input processing, and system throughput matching. The experimental results demonstrate that QuadPipe achieves low numerical error, efficient resource utilization, and reliable system-level integration, providing a feasible hardware solution for large model inference on FPGA platforms.
Available online , doi: 10.11999/JEIT260580
Abstract:
Objective In practical multi-antenna receiving systems, antenna elements usually exhibit anisotropic radiation patterns. However, the effect of such pattern characteristics on direction sensing remains insufficiently explored in both academia and industry. Particularly, when the target direction significantly deviates from the array boresight or approaches a null region of the antenna radiation pattern, the received signal power at the array can be severely attenuated. Consequently, traditional fixed antenna arrays face difficulties in meeting the direction-sensing requirements of low-altitude communication networks under extreme incident conditions. To address this issue, a rotatable antenna array system is developed by considering the directional radiation pattern of each antenna element. This study provides a solution for high-precision Direction Of Arrival (DOA) estimation of Unmanned Aerial Vehicles (UAVs). Methods To achieve high-precision direction sensing, a Recursive Rotation Root-MUltiple SIgnal Classification (RR-Root-MUSIC) algorithm based on a rotatable antenna array architecture is proposed. Specifically, the Root-MUSIC method is first applied to obtain an initial DOA estimate of the target. Subsequently, the array is rotated toward the estimated direction, and the Root-MUSIC algorithm is applied again to obtain an updated estimate. This sensing-rotation-resensing process is iteratively performed until the difference between adjacent estimates satisfies a predefined convergence threshold. Furthermore, a closed-form Cramér-Rao Lower Bound (CRLB) considering the directional antenna gain is derived as a theoretical performance benchmark. Results and Discussions The performance of the proposed RR-Root-MUSIC algorithm is evaluated by comparing its estimation results with the derived CRLB under different Signal-to-Noise Ratio (SNR) conditions and incident directions. Simulation results demonstrate that the proposed method achieves a convergence probability of at least 99.8%, and the required number of iterations decreases with increasing SNR (Figs. 4 ～6 ). When the target direction significantly deviates from the array boresight, the proposed method effectively improves the observation conditions through iterative array rotation, thereby reducing the DOA estimation error. Additional simulations using a non-ideal microstrip patch antenna radiation pattern show that the proposed method maintains stable estimation performance under antenna-pattern variations (Fig. 7 ). Along the considered UAV flight trajectories, the estimation error remains low and varies smoothly, indicating that array-orientation adjustment can reduce the influence of UAV position variations on direction-sensing performance (Fig. 8 ). Conclusions This study develops a rotatable antenna array system for direction sensing of a single UAV in low-altitude communication networks. Based on a sensing-rotation-resensing process, the proposed RR-Root-MUSIC method iteratively adjusts the array orientation according to the current DOA estimate, thereby improving subsequent observation conditions. Simulation results show that the proposed method alleviates DOA estimation performance degradation caused by directional-gain attenuation, especially when the target direction significantly deviates from the array boresight. Additional simulations using a non-ideal microstrip patch antenna radiation pattern further verify the effectiveness of the proposed method under non-ideal antenna-pattern conditions. The proposed method also exhibits stable convergence performance and maintains low estimation errors along the considered UAV trajectories. These results indicate that rotatable antenna arrays provide a potential solution for low-altitude UAV direction sensing.
Available online , doi: 10.11999/JEIT260591
Abstract:
Objective The rapid development of low-altitude logistics, urban air mobility, and security monitoring requires Integrated Sensing And Communication (ISAC) systems to detect weak uncooperative “Low, Slow, and Small” (LSS) targets while reusing communication waveforms and hardware. In practical low-altitude environments, echoes from buildings, towers, and ground facilities can be much stronger than target echoes and may mask targets during range, angle, and Doppler processing. Static clutter exhibits stable directions of arrival and near-zero-Doppler slow-time components. Conventional Moving Target Indication (MTI) suppresses zero-frequency clutter by exploiting Doppler differences, but this strategy may attenuate low-speed targets whose Doppler components overlap with the clutter region and distort micro-Doppler modulation. Spatial nulling based on clutter-angle information preserves low-Doppler components, but it may simultaneously suppress targets when target and clutter directions are closely aligned. Neither approach alone can simultaneously achieve strong clutter suppression, low-speed target preservation, and robustness to angular overlap. Therefore, this study proposes a joint spatial-slow-time clutter suppression method that exploits environmental angular prior information while preserving weak low-speed target signatures. Methods After reception of the MIMO-OFDM sensing echo, down-conversion, digitization, cyclic-prefix removal, and FFT processing are performed. Element-wise division by transmitted communication symbols removes data modulation, and measurements over array elements, subcarriers, and slow-time symbols are stacked into a three-dimensional sensing data matrix. The proposed method consists of offline clutter angle map construction and online two-stage suppression. During the offline stage, inverse FFT along the subcarrier dimension separates range cells from pure clutter measurements. Strong clutter cells are selected, and the corresponding slow-time snapshots are used to construct spatial covariance matrices, from which two-dimensional MUSIC estimates the azimuth and elevation angles of dominant static scatterers. These estimated angles are stored in a CLutter Angle Map (CLAM), which serves as reusable environmental prior information for subsequent clutter suppression. During online sensing, steering vectors associated with the stored angles form a clutter spatial manifold matrix, from which an orthogonal projection operator is derived to remove spatially separable static clutter. To maintain model consistency, steering vectors for subsequent angle estimation are projected using the same operator. When a target direction overlaps or approaches a stored clutter direction, each CLAM angle is individually examined. For each candidate direction, the remaining CLAM directions are treated as interference and suppressed using a normalized local orthogonal projection, while the current direction is retained. This traversal avoids prematurely suppressing an angularly overlapped target and provides candidate slow-time sequences for further discrimination. The residual sequence of each subcarrier is mapped onto the complex I/Q plane, where residual static clutter appears as an approximately fixed DC offset and moving targets form circular arcs due to slow-time phase evolution. Because low target speeds and finite observation intervals may produce only short arcs with limited geometric constraints, Taubin circle fitting estimates the circle center through a normalized generalized eigenvalue problem. The estimated center represents the residual clutter bias and is subtracted from the complex slow-time samples. Thus, CLAM-guided global projection removes spatially separable clutter, while local angular traversal and circle fitting address angular overlap and residual slow-time clutter without modifying the target phase trajectory. Results and Discussions MIMO-OFDM simulations under the ISAC architecture verify the proposed method in several representative scenarios. When target and clutter angles are separated, the proposed method preserves distinct target peaks in the two-dimensional angular spectrum, and the Root Mean Square Error (RMSE) of azimuth and elevation estimation is substantially lower than that of conventional MTI (Fig. 4 ). When target and clutter angles are close, spatial filtering alone may suppress the target together with the clutter and cause missed detection (Fig. 5 ). The proposed local traversal and circle-fitting stages effectively mitigate this limitation. At a low target speed of 2 m/s, the proposed method reduces the angle-estimation RMSE to a level close to the ideal clutter-free result, outperforming both MTI and CLAM+FFT (Fig. 6 ). It also exhibits lower sensitivity to changes in subcarrier spacing than MTI, indicating improved robustness under different communication-system bandwidth settings (Fig. 7 ). Time-frequency results further show that MTI causes nonlinear distortion and energy loss in target echoes, whereas the proposed method retains the principal translational component and periodic rotor-induced micro-Doppler structures required for fine-grained target characterization (Fig. 8 ). Conclusions The proposed joint spatial-slow-time method provides a target-preserving clutter suppression strategy for ISAC low-altitude sensing by combining environmental angular priors with slow-time bias estimation. It achieves accurate angle estimation in both angularly separated and closely spaced target–clutter scenarios, with a clear advantage for low-speed targets. By estimating and removing residual DC bias through circle fitting instead of Doppler-domain high-pass filtering, it avoids unnecessary target-energy attenuation and preserves micro-Doppler structures required for subsequent classification and recognition. Therefore, the proposed method provides high-fidelity echoes for subsequent low-altitude target perception. Future work will investigate dynamic clutter broadening in measured environments and intelligent target classification using real-world echo data.
Available online , doi: 10.11999/JEIT260611
Abstract:
Objective The rapid development of quantum computing poses a serious threat to traditional public-key cryptosystems, making Post-Quantum Cryptography (PQC) a critical research area. The Module-Lattice-based Key-Encapsulation Mechanism (ML-KEM) and Module-Lattice-based Digital Signature Algorithm (ML-DSA), standardized by the National Institute of Standards and Technology (NIST), rely on polynomial multiplication using the Number Theoretic Transform (NTT) and Inverse Number Theoretic Transform (INTT). The efficiency of NTT/INTT hardware accelerators directly affects the real-time performance and power consumption of these algorithms. Diverse application scenarios, ranging from resource-constrained to resource-rich environments, require different trade-offs between hardware resources and computational throughput. Furthermore, efficiently supporting both ML-KEM and ML-DSA on a single hardware platform remains a significant engineering challenge. To address these issues, this paper proposes a configurable NTT/INTT hardware accelerator design methodology that systematically explores the design space across three dimensions. Methods The design space is modeled according to three dimensions: modular multiplication implementation strategy (Digital Signal Processing (DSP)-intensive vs. Look-Up Table (LUT)-intensive), memory interface strategy (two-read-one-write vs. two-read-two-write), and pipeline depth (2-stage to 4-stage). For modular multiplication, two variants of the Barrett reduction are implemented. The DSP-intensive variant uses dedicated DSP slices for multiplication and accumulation, whereas the LUT-intensive variant uses only LUT resources and shift-add operations without consuming DSP resources. For the memory interface, the Two-Read-Two-Write (R2W2) mode fully utilizes the dual-write-port capability of Block RAM (BRAM) to eliminate write conflicts, achieving one butterfly operation per cycle. The Two-Read-One-Write (R2W1) mode requires two write cycles per butterfly operation, reducing throughput by 50%. For pipelining, four pipeline stages are inserted into the LUT-intensive Barrett modular multiplier to shorten the critical path: stage 1 performs multiplication, stage 2 calculates the intermediate sum, stage 3 computes the difference, and stage 4 completes the reduction through a correction multiplexer. A write-valid auto-delay mechanism is designed to align write-enable signals with the pipeline depth, simplifying control logic. The unified architecture supports algorithm configurability through an Alg_MODE parameter, which selects ML-KEM or ML-DSA, and dynamic NTT/INTT switching through an NTT_MODE signal. The proposed design is synthesized on a Xilinx Artix-7 FPGA (xc7a200t) using Vivado 2017.4 and ISE 14.4 for comparison. Results and Discussions Comprehensive evaluations are conducted on the modular multiplier, memory interface, pipeline stages, and the complete NTT/INTT accelerator. For the modular multiplier (Table 2 ), the LUT-intensive Barrett implementation reduces Equivalent LUT (ELUT) area by 55.3% for ML-KEM (from 928 to 415) and by 75.1% for ML-DSA (from 3 105 to 773) compared with the DSP-intensive implementation. The Equivalent Area-Delay Product (EADP) is reduced by 41.89% for ML-KEM and by 83.23% for ML-DSA, demonstrating improved resource efficiency, especially for large bit widths. For the memory interface (Table 3 ), the R2W2 mode achieves 100% higher NTT throughput than the R2W1 mode, with slightly lower LUT and register counts due to simplified control logic. For pipeline optimization (Table 4 ), the 4-stage pipeline reduces the critical path delay of the ML-DSA modular multiplier from 15.114 ns to 4.359 ns (71% reduction) and that of ML-KEM from 17.252 ns to 4.052 ns (76.5% reduction). The Area-Delay Product (ADP) decreases by 73.4% for ML-KEM and 71.3% for ML-DSA. At the complete accelerator level (Table 8 ), the 4-stage pipeline achieves an area-time product (ATP) of 1.782 kLUT·μs for ML-KEM and 3.073 kLUT·μs for ML-DSA. Compared with the 3-stage pipeline, the 4-stage pipeline reduces power consumption by 4.8% for ML-KEM and 12.9% for ML-DSA (Table 9 ). The power breakdown (Fig. 5 ) shows that the modular reduction unit power share decreases from 47% (3-stage) to 38% (4-stage) for ML-DSA. Compared with state-of-the-art single-butterfly designs (Table 10 ), the proposed accelerator reduces ATP by 25.15% for ML-KEM and by 1.44% for ML-DSA while using fewer LUTs and BRAMs. Conclusions This paper presents a systematic three-dimensional design space exploration method for configurable NTT/INTT hardware accelerators targeting ML-KEM and ML-DSA. By abstracting the design space into modular multiplication strategy, memory interface strategy, and pipeline depth, the quantitative trade-offs among resources (LUT, DSP, BRAM), timing (frequency and delay), and power consumption are analyzed. The optimal configuration is determined as LUT-intensive Barrett modular multiplication, a two-read-two-write memory interface, and a 4-stage pipeline. Experimental results on FPGA demonstrate that this configuration achieves a balance between resource efficiency and performance, with ATP values of 1.782 kLUT·μs for ML-KEM and 3.073 kLUT·μs for ML-DSA. The proposed methodology can be extended to other NTT-based PQC algorithms and provides data-driven guidelines for flexible hardware accelerator design targeting both edge devices and cloud servers. Future work will explore efficient polynomial multiplication based on this NTT/INTT accelerator and extend the method to additional domestic and international standardization algorithms.
Available online , doi: 10.11999/JEIT260621
Abstract:
Objective The Compressed Sparse Row (CSR) format is widely used in sparse neural network accelerators to reduce weight storage and support row-wise computation. In addition to nonzero weights, column indices and row boundaries are stored as metadata. These fields perform different functions during sparse computation and may lead to different fault consequences after a Single-Event Upset (SEU). Errors in weights directly perturb arithmetic operands, whereas errors in metadata may redirect activation accesses or alter row boundaries. Existing uniform protection schemes do not distinguish these field-dependent fault characteristics and introduce redundant protection for all stored data. Therefore, a field-specific fault analysis method and a selective protection strategy are developed to establish the relationship between field functions and fault consequences while reducing unnecessary storage and decoding overhead. Methods A software fault-injection framework is established based on CSR execution rules and validity constraints. The values, col_indices, and row_ptr fields store nonzero weights, column indices for activation access, and row boundaries, respectively (Fig. 1 ). For each fault case, one logical storage bit is flipped in a selected field. A corrupted column index is classified as an erroneous connection when it remains within the valid input range; otherwise, it is classified as an address failure. A corrupted internal row boundary is classified as an erroneous boundary when the range and nondecreasing-order constraints remain satisfied; otherwise, it is classified as a structural failure. Executable cases are processed through inference, and the accuracy decrease is calculated as the difference between fault-free accuracy and post-fault accuracy. The original stored value is restored after each trial to ensure independent single-bit fault injections (Fig. 2 ). The first and last row_ptr endpoints of each layer are treated as configuration constants, and only internal boundaries are considered in the row_ptr fault and protection spaces. LeNet-5/MNIST and VGG16/CIFAR-10 models are trained, pruned, and quantized using 8-bit signed INTeger (INT8) weights. For LeNet-5, all 22 7264 logical storage bits in five target layers are exhaustively evaluated, including 111784 bit in values, 111 784 bit in col_indices, and 3696 bit in row_ptr. For VGG16, layer-stratified quota sampling is adopted for 16 target layers to evaluate larger sparse matrices and wider metadata fields. The sampled fault space contains 4 096, 16 384, and 30 944 cases for values, col_indices, and row_ptr, respectively (Table 2 ). Each model is evaluated with and without injected faults using the same fixed set of 1 000 test images. Field-level failure rates are calculated based on the corresponding fault cases. Single-Error Correction and Double-Error Detection (SECDED) codes are applied for selective protection evaluation. P0 denotes the unprotected baseline; P1 protects internal row_ptr boundaries; P2 additionally protects col_indices; and P3 independently protects all three fields. PT jointly encodes values and col_indices while separately protecting row_ptr. Protection effectiveness is evaluated using the original LeNet-5 CSR data-bit fault space under the condition that each SECDED codeword contains at most one erroneous bit. Accuracy comparisons are performed on the fixed set of 218 487 executable cases under P0. Element-level encoding converts 8-bit and 16-bit data into 13-bit and 22-bit codewords, respectively, whereas 64-bit memory-word encoding converts 64-bit data into 72-bit codewords. Storage overhead is calculated relative to the unprotected data capacity of each encoding organization. Codewords are generated during model loading or memory writes, and decoding and correction are performed on the inference read path. Decoder functions are verified through C and Register-Transfer Level (RTL) co-simulation with an independent bit-level reference model. Module-level Out-Of-Context (OOC) placement and routing are performed on a Xilinx XC7Z020 device under a 10 ns reference clock period. Results and Discussions In the LeNet-5/MNIST experiments, all 111 784 values fault cases remain executable, whereas address failures in col_indices and structural failures in row_ptr account for 5.83% and 61.20% of their corresponding fault cases, respectively. The maximum accuracy decrease reaches 0.70 percentage points (pp) for executable weight faults and 10.50 pp for executable boundary faults (Table 3 ). The larger influence of row boundary errors results from the fact that one incorrect boundary may redistribute multiple consecutive nonzero entries between adjacent output rows. For col_indices, the aggregate address-failure rate reaches 31.52% at bit 7 across the five evaluated layers. No address failure occurs in the first fully connected layer because its 256-dimensional input range covers all possible unsigned 8-bit index values (Fig. 3 ). For row_ptr, bit flips at positions 8～15 always violate structural constraints and contribute 81.70% of all structural failures (Fig. 4 ). These results demonstrate that fault consequences are determined by the interaction between bit significance, valid index ranges, and CSR structural constraints. In the VGG16 samples, address and structural failures account for 36.89% and 75.15% of the corresponding col_indices and row_ptr cases, respectively, confirming that the field-dependent failure characteristics remain under larger sparse models and wider metadata fields. In the LeNet-5 protection evaluation, P1 eliminates all 2 262 structural failures caused by row_ptr faults, and P2 further eliminates all 6 515 address failures caused by col_indices faults. Among the evaluated configurations that remove both failure types, P2 achieves the lowest storage overhead, with additional costs of 31.35% for element-level encoding and 6.35% for 64-bit memory-word encoding. P3 and PT additionally protect values and restore baseline accuracy for all cases in the fixed executable-case set. Joint encoding reduces the element-level overhead from 62.09% for P3 to 37.50% for PT. However, under the 64-bit memory-word organization, both configurations require 12.50% overhead because each 64-bit data word is expanded with eight check bits (Fig. 5 ). Post-route implementation results show that P1, P2, P3, and PT require 72, 119, 166, and 147 LookUp Tables (LUTs), respectively (Table 5 ). The protected element-level read-path modules achieve longest internal data paths of 6.50～6.86 ns and Worst Negative Slack (WNS) values of 3.11～3.44 ns, satisfying the setup-time constraint under the 10 ns reference period. Since decoder branches operate in parallel, their delays do not accumulate along a single critical path. For the 64-bit decoder, pipelining reduces the longest internal data path from 9.38 ns to 6.49 ns, corresponding to a 30.8% reduction, while increasing the flip-flop count from 140 to 379. The pipelined decoder core has a High-Level Synthesis (HLS) scheduling latency of three cycles and an initiation interval of one cycle, allowing one new codeword to be accepted per cycle. Conclusions The fault consequences of CSR-based sparse neural networks depend on field function, bit position, and CSR validity constraints. Field-specific analysis shows that protecting row_ptr and col_indices is sufficient to eliminate structural and address failures within the evaluated LeNet-5 fault space, whereas protecting values is required to suppress accuracy degradation caused by weight errors. Full-field SECDED protection restores baseline accuracy for the fixed executable-case set. Joint codewords reduce parity overhead for element-level encoding but provide no additional storage benefit under the evaluated 64-bit memory-word organization. Selective protection reduces decoding resources compared with full-field protection, whereas pipelined decoding improves timing performance at the cost of additional registers. These results provide a basis for designing reliability-aware CSR storage protection by matching field coverage and codeword organization with application requirements and hardware constraints.
Available online , doi: 10.11999/JEIT260213
Abstract:
Objective In wireless communication environments under low Signal-to-Noise Ratio (SNR) conditions, signal feature separability deteriorates, and nonlinear distortion caused by multipath effects poses challenges for Automatic Modulation Classification (AMC). Existing deep learning models suffer from insufficient robustness under low SNR conditions, while Transformer-based models have high computational complexity. To address these issues, a gated recurrent network for modulation recognition integrating adaptive denoising and aggregation attention is proposed to improve recognition accuracy under low-SNR conditions while maintaining low model complexity. Methods The proposed AMC method consists of three key components. First, an adaptive denoising unit based on a Convolutional Kolmogorov-Arnold Network (CKAN) is designed to decouple the nonlinear superposition relationship between signals and noise, providing high-purity inputs for subsequent feature extraction. The CKAN combines the nonlinear function approximation capability of the Kolmogorov-Arnold representation theorem with the local feature extraction capability of convolutional operations. Then, an aggregation attention mechanism is introduced to reduce the computational complexity of conventional attention by employing an aggregation matrix. Based on this mechanism, a Talk-Attention module and an Attention-Gated Recurrent Unit (Att-GRU) are constructed to enhance cross-head information interaction and long-term temporal dependency modeling, thereby mitigating the influence of multipath interference. Finally, an orthogonal dual-channel strategy and a multi-scale feature fusion architecture are adopted to integrate global denoised features with local time-frequency features of I/Q components, improving the discrimination of intrinsic signal characteristics. Results and Discussions The proposed method achieves average classification accuracies of 64.36% and 71.52% on the RadioML 2016.10a and RML22 datasets, respectively, with only 0.23M parameters and an average inference time of 9.1 ms (Table 2 ). Comparative experiments with existing deep learning models demonstrate the effectiveness of the proposed method. On the RadioML 2016.10a dataset, the proposed method improves the average classification accuracy by 2.08%, 1.83%, 1.03%, 1.96%, and 2.15% compared with AWN, MCDformer, FE-SKVIT, AMC-NET, and MCLDNN, respectively (Table 2 ). When SNR > 8 dB, the accuracy exceeds 90% on both datasets, and approaches 100% on the RML22 dataset when SNR > 16 dB. Even at a low SNR of –10 dB, the model maintains accuracies of 48.5% and 33.0% on the two datasets, respectively (Fig. 5 ). The confusion matrix at SNR=0 dB indicates that classification difficulties mainly occur among QPSK and 8PSK, AM-DSB and WBFM, and high-order QAM modulations, due to phase distortion, spectral similarity during silent carrier periods, and reduced symbol spacing (Fig. 6 ). Significant improvements are also observed under low-SNR conditions (Fig. 7 ). Ablation experiments verify the contribution of the key components. Removing the adaptive denoising unit reduces the accuracy by 2.15%～4.30% on RadioML 2016.10a and 3.46%～6.02% on RML22, while only slightly increasing the inference time (Table 3 ). The CKAN configuration with 16/4 channels achieves the best balance between accuracy and efficiency (Fig. 8 ). Conclusions A gated recurrent network integrating adaptive denoising and aggregation attention is proposed for AMC. The proposed method effectively improves recognition performance under low-SNR wireless communication environments, enhances adaptability to complex channel conditions, and reduces deployment requirements. Experimental results on two public datasets demonstrate the effectiveness of the proposed method in terms of accuracy, model size, and inference efficiency, especially under challenging low-SNR conditions. Future work will focus on improving the computational efficiency of CKAN-based components to further accelerate inference.
Available online , doi: 10.11999/JEIT260555
Abstract:
Objective The rapid growth of global data imposes significant challenges on computing infrastructure. However, Computing Power Network (CPN) still suffers from spatiotemporal resource mismatch caused by the separation of network routing and computing task scheduling, as computing resource states and network states are represented in different dimensions. This mismatch leads to suboptimal scheduling decisions and increases task failures by nearly 20% under imbalanced task arrival scenarios compared with uniform task distributions. Moreover, existing Deep Reinforcement Learning (DRL)-based scheduling methods have limited topology adaptability and inflexible action spaces when handling arbitrary computing nodes in CPN. This paper aims to develop an integrated computing-network routing mechanism to address resource mismatch, jointly optimize task failure rates and completion times, and improve the dynamic adaptability and service quality of CPN. Methods First, a centralized software-defined CPN architecture is constructed, and a mathematical model is established for task arrivals, network bandwidth, computing resources, and service times. A multi-objective optimization function is formulated to minimize task failures and total service time. Then, the Double Deep Q-Network (DDQN)-based COMPASS algorithm is proposed with three key designs: (1) a three-dimensional state representation integrating network topology, real-time computing resource status, and task attributes; (2) a joint and truncated action space that couples computing node selection and routing path decisions to reduce exploration complexity; and (3) a multi-objective weighted reward mechanism with higher penalties for task failures to avoid suboptimal strategy convergence. An asynchronous training workflow is further designed to address delayed feedback caused by task execution latency, integrating routing decisions with DRL training. To evaluate the performance, the CNRSim platform is developed based on EdgeCloudSim, incorporating three real backbone network topologies, four task types, and representative benchmark algorithms. Results and Discussions Parameter tuning experiments identify that an action space dimension of 20 (Fig. 5 ) and a learning rate of 5.0E–5 achieve a balance between convergence speed and final performance (Fig. 6 ). Evaluations on three real backbone network topologies demonstrate that COMPASS outperforms benchmark methods, improving the average reward by 48% (Fig. 7 ) and achieving the lowest cumulative task failure rate (Fig. 8 ). In particular, under imbalanced task distributions, COMPASS reduces the task failure rate from 27.0% to 9.5% (Fig. 9 ). Further analysis on the GEANT2 topology shows that COMPASS achieves better computing node load balancing (Fig. 10 ) and handles more tasks with lower bandwidth utilization (Fig. 11 ). Failure analysis indicates that joint optimization of computing and network resources effectively reduces failures caused by insufficient bandwidth and overloaded computing nodes (Fig. 9 ). Moreover, COMPASS maintains the highest average rewards under dynamic workloads (Table 3 ). Conclusions This paper proposes COMPASS, a DRL-based integrated computing-network routing algorithm, to address performance bottlenecks in CPN caused by spatiotemporal mismatches between resource distribution and task arrivals. By adopting a three-dimensional state representation, joint action design, and a multi-objective reward mechanism, COMPASS enables global collaborative scheduling of computing and network resources. Extensive experiments on the CNRSim platform verify that COMPASS achieves better performance than benchmark methods in terms of reward improvement, task failure reduction, and load balancing, particularly under imbalanced scenarios. This work provides an effective approach for CPN optimization and large-scale distributed resource scheduling. Future work will investigate task fault tolerance and retransmission mechanisms, as well as the combination of COMPASS with emerging intelligent technologies for further optimization.
Available online , doi: 10.11999/JEIT260558
Abstract:
Objective To address the privacy leakage risks of medical images during transmission and storage, and to overcome the shortcomings of existing encryption methods (such as insufficient cropping resistance of cellular automata and limited key space of low-dimensional chaotic systems), this study constructs a medical image encryption scheme based on a five-dimensional hyperchaotic system. The proposed scheme adopts a "scrambling-diffusion-scrambling-diffusion" encryption framework and performs block processing on cellular automata to enhance the security and robustness of the encryption algorithm, thereby providing technical support for data protection in smart healthcare scenarios. Methods First, a magnetron memristor is introduced into the classical Chen chaotic system to construct a five-dimensional dual-memristive hidden hyperchaotic system, and its dynamical behavior is verified through Lyapunov exponent spectra, bifurcation diagrams, and phase space trajectory analysis. Second, based on this system, a "scrambling-diffusion-scrambling-diffusion" encryption framework is designed: the SHA-256 hash algorithm combined with preset key parameters is used to generate the initial values of the chaotic system; global scrambling and block cat mapping are employed to achieve the first round of pixel position scrambling; DNA encoding, operation, and decoding, together with a block cellular automata diffusion mechanism, are used to perform the first round of pixel value transformation; then, block affine transformation is applied for the second round of pixel position scrambling, and diffusion based on the finite field GF(257) is used for the second round of pixel value transformation. Finally, the security performance of the algorithm is evaluated through multiple metrics, including information entropy, chi-square test, correlation analysis, key sensitivity analysis, differential attack resistance testing, as well as cropping and noise attack tests. Results and Discussions Analysis shows that the proposed five-dimensional dual-memristive system exhibits two positive Lyapunov exponents, demonstrating typical hyperchaotic behavior. The encryption algorithm designed based on this system achieves a key space of 21495 and is highly sensitive to tiny key perturbations (NPCR of 99.612% and UACI of 33.475% under a 10–15 perturbation), enabling it to effectively resist brute-force. Meanwhile, the average information entropy (7.99933 ), average chi-square value (246.2866 ), and adjacent pixel correlation coefficients (close to 0) of the ciphertext images indicate that the statistical features have been concealed, allowing resistance against statistical attacks. Furthermore, the algorithm possesses outstanding differential attack resistance (average NPCR: 99.6096 %, average UACI: 33.479%) along with good robustness against cropping and noise. Conclusions Based on a five-dimensional hyperchaotic system, we construct a medical image encryption scheme that integrates DNA coding and block cellular automata under a dual scrambling-diffusion architecture, and verify its security performance against statistical attacks, differential attacks, cropping, and noise. This study provides a reference for the application of high-dimensional hyperchaos in medical image encryption, and provides practical cases for enhancing the resistance of cellular automata against cropping and noise attacks.
Available online , doi: 10.11999/JEIT260109
Abstract:
Objective Deep Reinforcement Learning (DRL) uniquely integrates the high-dimensional perception capabilities of deep learning with the autonomous decision-making optimization of reinforcement learning, demonstrating substantial promise for developing autonomous driving decision-making systems. Nevertheless, as a safety-critical system, autonomous driving mandates rigorous verifiability of its decision models prior to real-world deployment. Current DRL-based autonomous driving faces two principal challenges: first, the black-box nature of neural networks results in a lack of interpretability in the decision-making process; second, the continuous and infinite state space presents a state explosion problem for conventional formal verification methods. These challenges complicate ensuring the safety of DRL models in complex and dynamic traffic environments, thereby limiting their reliable application in practical scenarios. Prior approaches often decouple training from verification, or are tailored to white-box systems, inadequately addressing the verification demands of black-box continuous dynamical systems. To address these shortcomings, this paper aims to propose an innovative Counterexample-Guided Abstraction Refinement (CEGAR)-driven framework that achieves profound integration of training and formal verification for autonomous driving decision models. By establishing a mapping mechanism from black-box continuous systems to finite-state abstract models, coupled with an iterative refinement and retraining strategy, this framework aims to yield an autonomous driving system that possesses both intelligent decision-making capabilities and formal safety guarantees, thereby enhancing the trustworthiness of autonomous driving in complex urban settings. Methods This study adopts a systematic approach to develop and verify safe autonomous driving decision models. First, a reinforcement learning environment named CarlaEnv, compliant with the OpenAI Gym interface, was constructed based on the CARLA simulation platform, providing support for model training and testing with diverse urban scenarios. In terms of model architecture, a dual-module structure comprising driving decision and braking decision modules was designed using the Double Deep Q-Network (DDQN) algorithm. For the path-following task, a four-dimensional state space (velocity, lateral offset, relative yaw angle, trajectory completion rate) and an eight-dimensional discrete action space were meticulously designed, along with a multi-objective composite reward function. Guiding the agent to learn safe and efficient driving strategies.To address the challenge of formal verification for black-box systems, a trajectory generation-based abstraction modeling method is proposed: by collecting the interaction trajectories of the DRL agent in the simulation environment, the continuous state space is discretized using an interval abstraction domain. An efficient mapping from concrete states to abstract states is established via an R-tree data structure, ultimately constructing a formally verifiable Kripke structure. During the verification phase, ACTL temporal logic is employed to define safety and liveness properties, ensuring the eventual completion of tasks. The core innovation of the CEGAR mechanism lies in the authenticity assessment and classification of counterexamples: when a property violation is detected during model checking, an algorithm analyzes the realizability of the abstract counterexample in the concrete domain—if concrete states violating the atomic propositions exist, it is identified as a genuine counterexample, and its corresponding failure trajectories are added to the experience replay buffer to guide targeted model retraining. If the counterexample is spurious, a dimension-selective refinement strategy is adopted, subdividing only the state dimensions involved in the verified property through interval partitioning to avoid state space explosion. This closed-loop iteration of training-verification-refinement continues until all predefined properties are successfully verified. Results and Discussions Experiments conducted on the Town02 map of the CARLA simulation platform demonstrate that after 300 training episodes, the cumulative reward per episode of the driving decision model steadily increased from initially negative values to a positive range. Among the 300 episodes, 274 episodes achieved high cumulative rewards, with the average reward converging after approximately 150 episodes. The trajectory completion rate reached the 90-100% range in 269 out of 300 episodes (Fig. 8 ), confirming that the model has learned effective driving strategies. For the braking decision model, both episode rewards and average rewards in the reward curve gradually increased and stabilized after around 30 episodes (Fig. 9 ). In the successful braking curve, as training progressed, the model learned to effectively brake and maintain a safe distance within a range of 2 to 10 meters from the preceding vehicle. Formal verification results indicate that the driving decision model simultaneously satisfies both safety and liveness properties. Through the counterexample-guided refinement process, the number of abstract states required for verifying the safety property was refined from an initial 64 to 701, while the number required for the liveness property was refined to 225 (Table 6 ), effectively eliminating spurious counterexamples. During the CEGAR iteration process, the verification time overhead grows linearly with the state space size (Fig. 10 ), demonstrating the scalability of the proposed method. Horizontal performance comparison with two baseline methods, DQN and DDPG, shows that the proposed method outperforms the compared methods in terms of task completion rate (approximately 95%), mean lateral offset (approximately 0.4 m), and collision rate (remaining at 0% after 65 episodes) (Fig. 13 ). Reliability analysis indicates that under two complex scenarios (rainy weather and nighttime driving), the proposed method achieves mean cumulative episode rewards that are 4.6 to 10.8 points higher than those of DDQN, DQN, and DDPG, corresponding to improvements of 4.9% to 12.4% (Fig. 14 ). Ablation experiments compare three groups: Group A (no verification), Group B (single-shot verification), and Group C (full CEGAR method). The results show that Group C achieves the best performance across all three metrics: collision rate (3.1%), task completion rate (94.8%), and mean lateral offset (0.214 m) (Fig. 15 ). Complex scenario testing further validates the robustness of the proposed method under rainy weather and nighttime driving (Fig. 16 ), with a standard deviation of cumulative episode rewards of only 17.75 points. In summary, compared to traditional deep reinforcement learning methods, the proposed approach maintains decision-making performance while significantly enhancing model reliability and safety through the counterexample-guided refinement and retraining mechanism. Conclusions This paper proposes a comprehensive and integrated CEGAR-driven training and verification framework for autonomous driving decision models. The framework successfully bridges the gap between data-driven, learning-based autonomous systems and rigorous, mathematically grounded safety verification. By innovatively integrating trajectory-based abstraction methods for black-box systems, an efficient R-tree mapping structure, ACTL-based property specification, and an advanced CEGAR architecture equipped with counterexample authenticity assessment and dimension-selective refinement mechanisms, this approach constructs decision models that combine high performance with formal verification guarantees. Experimental results in a realistic simulation environment demonstrate that the framework effectively learns driving policies, achieves formal verification of key safety and liveness properties, and enhances the overall reliability of the model compared to traditional methods.
Available online , doi: 10.11999/JEIT260606
Abstract:
Objective In the Internet of Vehicles (IoV), vehicles often transmit Basic Safety Messages through open wireless links to support road safety and traffic management. However, these messages may expose vehicle identities and travel information. Attackers may collect messages over time and associate them with the same vehicle. Therefore, vehicle authentication should protect both network security and user privacy. Many existing anonymous authentication schemes require complex cryptographic operations, resulting in high computation and communication costs for resource-constrained On-Board Units (OBUs). Some schemes also store long-term secret keys in vehicle devices, which may be exposed through physical extraction or side-channel attacks. A Physical Unclonable Function (PUF) uses random process variations generated during chip fabrication to construct unique hardware features. It can generate device-related responses without directly storing secret keys. Based on this property, this paper proposes a PUF-driven secure anonymous authentication protocol for IoV. The proposed protocol aims to reduce the risk of long-term key storage, protect vehicle identities, and reduce authentication overhead for vehicles. Methods The proposed protocol adopts a response-Feedback-based Lightweight Anti-Machine-learning-attack PUF (FLAM-PUF) as the hardware root of trust. FLAM-PUF combines an Arbiter PUF (APUF) with a reconfigurable Linear Feedback Shift Register (LFSR). The PUF response is fed back to update the LFSR state, which hides the relationship between Challenge–Response Pairs (CRPs) and increases the difficulty of machine-learning modeling attacks (Fig. 1 ). The system consists of a Center Management (CM), Local Management (LM) entities, Roadside Units (RSUs), and vehicle On-Board Units (OBUs) (Fig. 2 ). During initialization, reliable CRPs are collected in a trusted environment and transmitted to the CM through a secure channel. The main protocol parameters are listed in Table 1 . A PUF challenge is used to generate a vehicle pseudonym, while the corresponding PUF response is used as authentication parameters and session key data. The OBU can recover the required PUF response from its local PUF when needed, eliminating the need to store long-term secret keys. The initialization process is shown in Fig. 3 . The protocol adopts two types of pseudonyms. Long-term pseudonyms are used for network access and identity management, while temporary pseudonyms are used for communication within an LM area. The local access process is shown in Fig. 4 . Temporary pseudonyms and session keys are updated using unused CRPs (Fig. 5 ), and long-term pseudonyms are renewed before expiration (Fig. 6 ). Security is analyzed under the Dolev–Yao model using Burrows–Abadi–Needham (BAN) logic, ProVerif formal verification, and informal security analysis. Results and Discussions The proposed protocol combines PUF-based key generation with dynamic pseudonym management. First, FLAM-PUF uses response feedback and a reconfigurable LFSR to hide the internal CRP relationship, increasing the difficulty of machine-learning modeling attacks (Fig. 1 ). Second, the use of long-term and temporary pseudonyms separates network access from local vehicle communication. CRPs, pseudonyms, and session keys are updated synchronously, reducing the possibility that attackers associate vehicle identities across different authentication periods (Fig. 4 , Fig. 5 , and Fig. 6 ). BAN logic analysis shows that the vehicle can verify the binding between the temporary pseudonym and the PUF-based session key, while the LM can verify that this binding is authorized by the CM. ProVerif verification results show that the secrecy queries for vehicle identities and PUF responses are satisfied, and all three injective authentication correspondences hold. Under the assumed freshness and one-time CRP usage conditions, the results support the protocol's resistance to replay and impersonation attacks. Informal security analysis demonstrates that the protocol provides mutual authentication, anonymity, traceability, revocability, and protocol-level identity unlinkability. It also resists replay, impersonation, man-in-the-middle, and false message attacks (Table 3 ). In addition, the OBU does not need to store long-term secret keys in nonvolatile memory because the required PUF responses can be recovered when needed, reducing the risk of direct key extraction. For performance evaluation, the execution times of the main cryptographic operations are measured (Table 4 ), and the computation costs of the OBU, CM, and LM at each protocol stage are summarized in Table 5. For one complete online authentication and session key agreement process, the computation time of the proposed scheme on the OBU is approximately 1 365.99 ms, and the communication cost is 992 bits (Table 6 ). Under the same evaluation conditions, its computation time is close to that of Ref. [14 ] and lower than those of Refs. [15 ]–[17 ]. Its communication cost is lower than those of all four compared schemes. These results demonstrate that the proposed protocol provides more security functions while maintaining low vehicle-side costs. Conclusions This paper proposes a PUF-driven secure anonymous authentication protocol for IoV. FLAM-PUF is used as a hardware root of trust to recover authentication parameters and session key data when needed, avoiding direct storage of long-term secret keys in OBUs. The combination of long-term and temporary pseudonyms supports anonymous access, identity tracing, revocation, and protocol-level identity unlinkability. BAN logic, ProVerif formal verification, and informal security analysis show that the proposed protocol satisfies its main security requirements and resists common network attacks. Performance evaluation results show that one online authentication and session key agreement process requires approximately 1 365.99 ms of OBU computation and 992 bits of communication (Table 6 ). The proposed protocol achieves a balance among security, privacy protection, and low computational cost, making it suitable for resource-constrained IoV devices.
Available online , doi: 10.11999/JEIT260338
Abstract:
Significance Because optical signals are broadcast, information security is a key research issue in Visible Light Communications (VLC). Conventional encryption techniques address network security at the upper layers of the protocol stack through access control, cryptographic protection, and end-to-end encryption. However, their security depends on the assumption that eavesdroppers have limited computational capability. This assumption is increasingly challenged by advances in high-performance and quantum computing. In recent years, Physical-Layer Security (PLS) has emerged as a new secure transmission paradigm and has drawn wide research attention. PLS uses the randomness, heterogeneity, and differences between the main channel and the eavesdropping channel to support secure information transmission at the physical layer. Substantial progress has been made in PLS for Radio Frequency Wireless Communication (RFWC). However, because VLC differs from RFWC in frequency band, transmitted signal form, power representation, and channel characteristics, PLS results obtained for RFWC systems cannot be directly applied to VLC. Although many studies have examined VLC PLS technologies, their fundamental theories, key techniques, and future challenges have not yet been systematically reviewed. To fill this gap, this paper presents a comprehensive survey of VLC PLS. Progress To evaluate and improve system performance, a classic VLC PLS system model is first established. This model consists of the received signal model, input constraint model, and channel gain model. A comprehensive theoretical framework for performance evaluation is then developed. The framework covers instantaneous, statistical, and asymptotic performance metrics. For instantaneous performance, existing studies on instantaneous secrecy capacity and instantaneous secrecy rate under different scenarios are summarized. For statistical performance, average secrecy capacity, average secrecy rate, secrecy outage probability, probability of strictly positive secrecy capacity, and intercept probability are analyzed. For asymptotic performance, secrecy diversity order and secrecy degrees of freedom are discussed. In addition, key techniques for improving PLS performance are reviewed, including secure beamforming, artificial-aided techniques, physical region protection, secure coding, and secure diversity. Prospects Despite recent progress, several challenges remain in VLC PLS. Four major challenges are identified in this paper. (1) Accurate PLS performance limits. Deriving exact expressions for secrecy capacity under the unique physical constraints of VLC remains difficult. (2) Incomplete evaluation framework. Some key metrics widely used in RFWC have not been fully studied in VLC, and a comprehensive VLC PLS performance evaluation framework remains to be developed. (3) Limitations of existing methods. Conventional PLS performance enhancement methods typically follow a separated research paradigm of modeling, optimization, and verification. This paradigm may lead to a cycle of inaccurate modeling, suboptimal solutions, and limited performance gains. Therefore, new technologies, such as deep learning, reinforcement learning, and digital twins, should be integrated to build a data- and model-driven framework for VLC PLS performance enhancement. (4) Hardware platform gap. The lack of dedicated hardware platforms with adversarial topologies and real-time processing capability hinders the practical deployment of VLC PLS technologies. Addressing these challenges is essential for moving VLC PLS from theoretical research toward commercial applications. Conclusions The broadcast nature of optical signals makes VLC systems vulnerable to eavesdropping attacks. This paper presents a comprehensive survey of PLS in VLC. The survey covers system models, performance metrics, and key performance enhancement technologies. The reviewed performance metrics include instantaneous, statistical, and asymptotic metrics. The reviewed enhancement techniques include secure beamforming, artificial-aided techniques, physical region protection, secure coding, and secure diversity. Although substantial progress has been achieved, challenges remain in accurate performance limits, complete evaluation frameworks, new enhancement techniques, and practical hardware implementation. By exploiting channel differences at the physical layer without relying on complex encryption, PLS provides a promising security mechanism for next-generation secure and reliable VLC networks.
Available online , doi: 10.11999/JEIT260354
Abstract:
Objective Sneak Path Interference (SPI) in Resistive Random-Access Memory (ReRAM) introduces unpredictable inter-cell correlations and increases the complexity of signal detection. Conventional detection methods usually rely on known channel noise states, which limits their generalization in practical applications. To address this issue, three data detection methods based on Convolutional Neural Networks (CNNs) are proposed to model and suppress interference without prior channel information. The first method combines constrained coding with a multilayer CNN to determine the SPI state and recover data. The second method uses a dual-CNN framework, in which a lightweight CNN first identifies SPI and a multilayer CNN then performs refined detection. The third method incorporates Transfer Learning (TL) to maintain detection accuracy while reducing the required training sample size to one thousandth of that required by conventional methods. Simulation results show that the proposed methods achieve improved Bit Error Rate (BER) performance under unknown channel conditions. Compared with existing algorithms, the BER is reduced by at least half and approaches the theoretical performance limit. In addition, TL reduces the required number of training samples from\begin{document}$ {10}^{6} $\end{document} \begin{document}$ 1\;000 $\end{document} Methods To address different challenges in SPI detection, this paper proposes three methods. First, the Constrained Coding-Aided CNN (CC-CNN) detection framework is developed to handle complex inter-cell correlations caused by SPI. This method first uses constrained coding to detect the presence of interference. It then uses a CNN to learn random correlations under interference and achieve accurate signal recovery. Second, the dual-CNN-based detection method is proposed to avoid the code-rate loss caused by conventional constrained coding. By directly using a CNN to learn and identify SPI patterns from raw data, this method removes the need for redundant coding or extra overhead. It achieves high-precision interference detection while preserving the overall code-rate performance of the system. Third, the TL-based CNN (TL-CNN) detection method is designed to reduce the dependence of high-performance CNNs on large training datasets. By reusing knowledge from pretrained models, this method enables rapid adaptation to ReRAM signal detection tasks. It reduces the required number of training samples while maintaining high detection accuracy and resource efficiency, thereby improving its feasibility in practical scenarios. Results and Discussions Simulation results show that the three proposed methods consistently approach the theoretical lower bound (Fig. 4 ). They outperform baseline methods, including the Belief Propagation (BP) detector, Deep Neural Network (DNN) detector, and Elementary Signal Estimator (ESE) detector. The two-step network achieves performance comparable to that of the single-step network while avoiding code-rate loss. The TL-CNN detector attains near-optimal BER with only 1,000 target-domain samples, and its performance stabilizes when the sample size exceeds 1,000 (Fig. 6 ), confirming its data efficiency. The integration of the Selective Kernel (SK) module enables the models to capture SPI-induced spatial correlations effectively (Fig. 5 ). The TL strategy also supports robust performance under different noise conditions. Conclusions The crossbar array architecture of ReRAM is susceptible to SPI during storage operations, reducing data reliability. To address this issue, this paper proposes three deep learning-based detection methods. Type-I integrates constrained coding with a CNN for efficient and fast interference detection. Type-II uses a two-stage processing strategy. It first classifies interference patterns in the memory array and then detects affected units, thereby maintaining high detection accuracy while minimizing code-rate loss. Type-III introduces a TL framework that uses a pretrained source-domain model to reduce the number of target-domain training samples and lower training overhead. Experimental results show that, under different noise conditions, all three proposed methods achieve performance close to the theoretical lower bound. These methods provide an effective solution for improving the reliability of ReRAM storage systems.
Available online , doi: 10.11999/JEIT260676
Abstract:
Objective Low Earth Orbit Communication Satellite (LEO-COS) constellations provide globally distributed and continuously available communication signals, which have the potential to serve as illuminators of opportunity for passive Synthetic Aperture Radar (SAR) imaging. However, when LEO-COS signals are exploited for SAR imaging, conventional single-receiver systems suffer from limited imaging performance due to restricted effective bandwidth and low Signal-to-Noise Ratio (SNR). To address these limitations, this paper proposes a novel “constellation–swarm” cooperative SAR imaging mode by integrating LEO-COS illuminators with an Unmanned Aerial Vehicle (UAV) swarm. In this mode, multiple UAV receivers cooperatively acquire echo signals to extend effective spectrum coverage and enhance coherent accumulation gain. Since the spatial configuration of the UAV swarm directly determines the wavenumber spectrum distribution and further affects imaging quality, the configuration optimization problem under dynamic and heterogeneous bistatic observation geometries is investigated. This study aims to develop an effective configuration optimization method for UAV swarm SAR using LEO-COS opportunistic illumination, improving spectrum utilization, imaging resolution, and system robustness in complex observation scenarios. Methods The relationship between UAV swarm spatial configuration and SAR imaging performance is analyzed based on wavenumber spectrum geometry. According to the echo model of the “constellation–swarm” SAR system, the effects of UAV positions and motion parameters on wavenumber spectrum distribution are derived. The UAV swarm configuration design problem is formulated as a multi-objective optimization problem by considering spectrum orthogonality, spectrum parallelism, spectrum misalignment suppression, and minimum imaging SNR. To solve the high-dimensional and strongly coupled optimization problem, an Associated Structure encoding Nondominated Sorting Genetic Algorithm II (AS-NSGA-II) is proposed. The associated structure encoding mechanism is designed to preserve the geometric coupling relationships among configuration variables during evolutionary operations, thereby reducing infeasible solutions caused by independent variable optimization. Layered Cubic chaotic mapping is introduced for population initialization to improve the distribution uniformity and diversity of candidate solutions. In addition, an adaptive mutation strategy based on population fitness is developed to dynamically adjust the search intensity during evolution, achieving a balance between global exploration and local refinement while satisfying configuration constraints. Results and Discussions Simulation results demonstrate that the optimized UAV swarm configurations effectively rearrange the wavenumber spectrum support regions of different receiving nodes, reduce spectrum overlap and misalignment, and improve spectrum stitching performance (Fig. 5 ). Compared with conventional Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) methods, the proposed AS-NSGA-II algorithm achieves faster convergence and better optimization stability. The convergence curves of the four objective functions show that AS-NSGA-II reaches a stable optimization state after approximately 60 generations, while GA and PSO exhibit slower convergence and larger fluctuations during evolution (Fig. 6 ). The area target imaging results further indicate that the proposed method suppresses false grating lobes more effectively and preserves target edge information and structural characteristics with higher fidelity (Fig. 7 ). Under low-SNR conditions, the optimized configuration maintains reliable imaging performance through multi-aperture coherent integration and achieves stronger robustness than the comparison methods. Conclusions This paper investigates the UAV swarm configuration optimization problem for “constellation–swarm” cooperative SAR imaging using LEO-COS signals as illuminators of opportunity. By analyzing the coupling relationship between UAV spatial configuration and wavenumber spectrum distribution, a multi-objective optimization model is established by jointly considering geometric spectrum constraints and imaging SNR requirements. The proposed AS-NSGA-II algorithm integrates associated structure encoding, Layered Cubic chaotic initialization, and adaptive mutation to improve the optimization capability for high-dimensional and strongly coupled configuration problems. Simulation results demonstrate that the proposed method outperforms conventional GA and PSO methods in convergence speed, optimization stability, spectrum stitching performance, and imaging quality. The optimized UAV swarm configuration effectively alleviates the bandwidth and SNR limitations of LEO-COS opportunistic illumination and improves the resolution and robustness of cooperative SAR imaging. This work provides a feasible approach for future wide-area, high-resolution, and robust SAR imaging based on integrated satellite–UAV swarm systems.
Available online , doi: 10.11999/JEIT260819
Abstract:
Objective In areas with weak ground-network coverage or limited Global Navigation Satellite System (GNSS) availability, multiple dual-functional Unmanned Aerial Vehicles (UAVs) can provide uplink communication and cooperative localization services. Ground terminals have different requirements for data volume, minimum communication rate, localization-accuracy threshold, service priority, and communication and localization service weights. Maximizing communication rate, localization accuracy, or a weighted sum of the two cannot directly reflect these differentiated requirements. A service below its activation threshold may be unusable, whereas resources allocated after demand saturation provide little additional value. A quasi-static service period with prior terminal positions is therefore considered. A topology-dependent Value of Service (VoS) is formulated, and Physical Resource Block (PRB) scheduling and UAV positions are jointly optimized. Methods Data or Sounding Reference Signals (SRSs) are transmitted by terminals over assigned PRBs. Time Difference of Arrival (TDoA) measurements are formed by synchronized UAVs, and a probabilistic air-to-ground channel model determines communication access and valid localization anchors. Communication VoS combines a sigmoidal rate utility with data completeness. Position Error Bound (PEB), derived from the accumulated Fisher Information Matrix (FIM), is mapped to an exponential utility to quantify localization VoS. Terminal priorities and communication and localization service weights are used to aggregate the two service values. The resulting problem P0 couples binary scheduling, non-concave utilities, time-frequency resources, service relationships, localization geometry, and UAV positions. For a fixed topology, the average number of allocated PRBs and active slots are used to construct a continuous service-level resource profile. This profile approximates the original schedule but does not provide an upper bound. Bandwidth responses are obtained by deterministic one-dimensional branch-and-bound search with damped Newton refinement and shadow-price bisection. Slot responses are obtained by deterministic comparison of a finite set of service-boundary, integer-slot, and feasible-domain-boundary candidates. Adjacent-integer recovery and deterministic slot packing are then used to construct a feasible integer schedule under resource-capacity and terminal-power constraints. If packing fails, the upward-rounded component with the smallest unit VoS loss is rolled back, and the schedule is repacked. VoS is then recomputed from the recovered integer schedule. Thus, the reported schedules remain feasible for P0 without any claim of global optimality. The fixed-topology resource-allocation procedure produces a complete resource response, including the recovered feasible schedule and its VoS, which is subsequently used to drive deployment. Coverage-repair, localization-geometry-improvement, and resource-saving candidates are generated according to service deficits, service relationships, and resource consumption. Normalized proxy scores, per-UAV candidate truncation, and beam search reduce the number of complete resource-response evaluations. For each retained deployment, service relationships are rebuilt, including communication access and localization anchors, and the FIM and resource allocation are recomputed. A candidate is accepted only when the VoS gain obtained from its complete resource response exceeds the preset threshold. The resulting outer sequence is monotonic and bounded and converges to a stable point within the generated candidate set. Resource allocation and node deployment remain coupled throughout the procedure. For each retained topology, communication access, localization anchors, the FIM, and the feasible integer schedule are recomputed before VoS is evaluated. Proxy scores are used only to rank candidates, whereas final acceptance is always based on the complete resource response. The next topology is therefore not selected from distance or localization geometry alone. The accepted update reflects terminal demand, resource scarcity, and localization geometry under the same feasible scheduling constraints used in the objective. This design keeps the optimization objective consistent with the final deployment decision and avoids a geometry-only selection rule. Results and Discussions Simulations are conducted in a 700 m × 700 m area with four UAVs at a baseline deployment height of 150 m. Communication-dominant, localization-dominant, and balanced terminals are included in equal proportions. At each load, all methods share 30 independent scenarios and identical random seeds. The deployment-height experiment reports a 95% confidence interval based on 1 000 scenario-level bootstrap resamples. For fair comparisons, all resource-allocation methods use the same topology, resource pool, and random scenarios. All deployment methods use the same proposed joint resource-allocation response, are evaluated from a cold start, and are subject to the same cap on complete resource-response evaluations, with initialization and training costs included. VoS-driven allocation yields smaller communication-rate and localization-accuracy demand deviations than the communication-priority and average-utility metrics because service saturation redirects resources from overprovisioned requests to insufficiently served requests (Fig. 2 ). In the convergence test, three feasible initializations converge to similar system VoS values. Damping suppresses oscillations near transitions between non-concave segments, and the small-scale benchmark indicates limited empirical loss from the continuous response and integer recovery (Fig. 3 ). With increasing terminal load, the proposed joint allocator outperforms a Particle Swarm Optimization (PSO)-based slot-response variant, as well as non-joint, learning-based, weight-driven, and random allocation methods (Fig. 4 ). Communication service is more sensitive to terminal load because both communication rate and data completeness must be maintained, whereas localization service accumulates information across anchors and slots. Resource-pool tests further show that additional bandwidth provides little benefit when slots are scarce, whereas increasing SRS bandwidth more directly improves localization. The structured deployment search achieves higher VoS with shorter end-to-end deployment time than PSO, Differential Evolution (DE), and Bayesian Optimization (BO) by focusing complete resource-response evaluations on service-aware candidates (Fig. 5 ). VoS first increases and then decreases with UAV deployment height because improved line-of-sight probability and multi-anchor visibility compete with increased propagation distance and path loss. Horizontal deployment updates balance communication-link quality and localization geometry according to terminal requirements (Fig. 6 ). Conclusions The proposed framework evaluates communication and localization services according to demand satisfaction and coordinates time-frequency resource allocation with iterative UAV deployment updates through complete resource responses and structured candidate search. It improves demand matching, system VoS, deployment efficiency, and interpretability while preserving feasibility under the original scheduling constraints. Prior terminal positions and a quasi-static service period are assumed. Future work will address dynamic terminal movement and changing demands over longer service periods through dynamic resource allocation and continuous UAV trajectory optimization.
Available online , doi: 10.11999/JEIT260525
Abstract:
Objective Low-bandwidth visual transmission is essential for edge intelligence applications such as disaster inspection, underwater exploration, and remote assistance. These scenarios require visual communication systems to provide a low bit rate, high structural fidelity, stable color reconstruction, and low inference latency. However, conventional image coding methods suffer severe quality degradation at extremely low bit rates because of the digital cliff effect and compression artifacts. Semantic communication offers a promising solution by transmitting task-relevant representations rather than pixel-level information. However, existing methods still struggle to balance transmission efficiency, reconstruction fidelity, and decoding complexity. Therefore, a structure-preserving semantic transmission method termed Edge-Link is proposed for low-bandwidth networks. Methods Edge-Link adopts a multimodal decoupled representation framework that separates image information into three complementary streams: a semantic stream, a structural stream, and a low-frequency appearance stream. The semantic stream is extracted using a frozen Contrastive Language-Image Pre-training (CLIP) encoder to provide global semantic guidance. The structural stream is derived from Canny edge information to explicitly preserve object boundaries. A low-resolution color map is used as the low-frequency appearance stream to preserve the global color distribution and illumination characteristics with minimal transmission overhead. Furthermore, a Fast Segment Anything Model (FastSAM)-based semantic gateway is used to distinguish foreground objects from background regions. An object-aware bit-rate allocation strategy is then applied to prioritize semantically important regions under bandwidth constraints. At the receiver, a deterministic dual-stream generative network based on SPatially ADaptive normalization (SPADE) is used for image reconstruction. Structural information provides spatial constraints, whereas semantic features guide image generation. This design avoids the high latency associated with iterative diffusion sampling. A LoRa communication prototype based on GNU Radio and a Universal Software Radio Peripheral (USRP) is also implemented to evaluate transmission feasibility and robustness under physical wireless-channel conditions. Results and Discussions Experiments are conducted on Set14 and DIV2K to evaluate transmission efficiency, reconstruction quality, perceptual naturalness, and latency. For 512 × 512 images, the proposed method achieves an average payload of 10.96 kB, corresponding to approximately 0.31 bit per pixel (bpp), and the end-to-end inference latency remains below 60 ms. Over a LoRa narrowband link of approximately 8 kbps, the airtime latency per frame is approximately 11 s, confirming the feasibility of the method under bandwidth-limited conditions (Table 1 ). The ablation study shows that the object-aware bit-rate allocation strategy improves the trade-off between bit rate and reconstruction quality. On DIV2K, the average payload is reduced from 126.09 kB to 93.55 kB while better reconstruction quality is maintained in semantically important regions (Table 2 ). Compared with JPEG, the proposed method also provides better structure preservation and perceptual quality at low bit rates. For image 0843 with a payload of 15.54 kB, the overall Structural Similarity Index Measure (SSIM) increases from 0.783 to 0.963, whereas the Learned Perceptual Image Patch Similarity (LPIPS) for the subject region of interest (Subject-ROI) decreases from 0.561 to 0.307 (Table 4 ). Under the same bit-rate constraint, the proposed method further outperforms VQGAN and ControlNet+Canny. It achieves a Peak Signal-to-Noise Ratio (PSNR) of 23.738 dB, an SSIM of 0.622, an LPIPS of 0.223, and a Natural Image Quality Evaluator (NIQE) value of 4.570. These results indicate a better balance between reconstruction fidelity and perceptual quality in low-bandwidth semantic image reconstruction (Table 5 ). Conclusions A structure-preserving semantic transmission framework for low-bandwidth networks is presented. By combining multimodal decoupled representation, object-aware bit-rate allocation, and deterministic dual-stream generative reconstruction, the framework balances transmission efficiency, structural fidelity, perceptual quality, and inference latency. The results show that the proposed method is suitable for high-reliability edge visual communication scenarios that require accurate object contours, stable color appearance, and efficient inference. Validation using the LoRa prototype further supports its practical applicability to bandwidth-constrained edge networks.
Available online , doi: 10.11999/JEIT260500
Abstract:
Objective Wide-area maritime surveillance requires satellite constellations to search for and revisit noncooperative maneuvering ships whose positional uncertainty increases during periods without successful observation. Meanwhile, multifunctional payloads are subject to strongly coupled constraints on observation timing, attitude maneuvering, payload mode, and energy consumption. To address dynamic target uncertainty while ensuring executable constellation scheduling, a cooperative search-and-tracking method based on dynamic information gain is proposed. Methods A closed-loop rolling-horizon framework inspired by model predictive control is constructed for prediction, optimization, execution, and feedback. At each rolling decision epoch, candidate atomic observation tasks are generated within the planning horizon, whereas only tasks within the current execution window are committed. Each task specifies the executing satellite, target, candidate pointing grid, start and end times, and payload mode. Target uncertainty is represented by a parameterized probability grid generated from the last confirmed state, speed and heading perturbations, and elapsed time since the last successful observation. Hit/Miss feedback updates the uncertainty baseline for the next rolling decision epoch, at which the probability grid, candidate tasks, and observation plan are regenerated. A state-driven dual-mode benefit model is established based on target information entropy and the number of consecutive Hits. In the robust tracking state, tasks in the narrow-field-of-view mode are evaluated using the prior capture probability, defined as the probability mass covered by the task footprint. In the lost-search state, tasks in the wide-field-of-view mode are evaluated using binary Hit/Miss event entropy as an approximation of the search information value. This approximation is based on the information-theoretic definition of Kullback-Leibler divergence and avoids explicit posterior reconstruction for each candidate task. A dynamic priority coefficient is introduced to increase the scheduling urgency of targets that have remained unobserved for longer periods and to moderately reduce the priority of repeatedly confirmed targets. The resulting multi-objective model maximizes weighted task benefit and information gain while minimizing energy consumption. Hard constraints are imposed on single-satellite temporal exclusivity, attitude-switching stabilization time, and energy budget. Based on Non-dominated Sorting Genetic Algorithm II (NSGA-II), the Cooperative Evolutionary Planning-Multi-Objective (CEP-MO) algorithm employs global integer-index encoding, constraint-aware Top-K heuristic initialization, satellite-group crossover, and an adaptive repair operator to improve feasible-solution generation and cooperative planning efficiency. The objective values of feasible Pareto solutions are normalized, and the solution closest to the ideal point (1, 1, 0) is selected for execution. Results and Discussions Simulations using a 48-satellite Walker Delta constellation demonstrate the effectiveness of the proposed method. In the 200-target scenario, Standard NSGA-II yields an average revisit interval of 53.5 min and a weighted coverage rate of 13.9%, whereas CEP-MO achieves 19.9 min and 41.7%, respectively. The average revisit interval is therefore reduced by 62.8% (Fig. 7 ). Removing the binary Hit/Miss event entropy benefit increases the system average uncertainty and average revisit interval. Replacing constraint-aware initialization with random initialization also degrades early-stage planning performance and the weighted coverage rate. As the number of targets increases from 100 to 200, CEP-MO maintains acceptable scalability (Fig. 8 ). At 200 targets, its weighted coverage rate is 16.6 percentage points higher than that of CEP-MO w/o Entropy. The computation time per rolling decision is approximately 22 s, which is shorter than the 10 min rolling period. In addition, the Gini coefficient of remaining satellite energy remains below 0.3, indicating that energy consumption is not excessively concentrated on a small subset of satellites. Conclusions The proposed framework integrates parameterized probability-grid uncertainty representation, state-driven dual-mode benefit evaluation for search and tracking, closed-loop rolling-horizon feedback, and constraint-aware multi-objective evolutionary planning. CEP-MO reduces the average revisit interval and improves the weighted coverage rate while satisfying single-satellite temporal exclusivity, attitude-switching stabilization, and energy-budget constraints. The method provides an effective approach for large-scale resource-constrained maritime surveillance and establishes a basis for future extensions that consider identification errors, communication delays, and constrained inter-satellite links.
Available online , doi: 10.11999/JEIT260630
Abstract:
Objective With the rapid development of edge intelligence, wearable electronics, and Internet-of-Things applications, edge devices increasingly require low-latency and low-power temporal inference under limited storage and hardware resources. Gated recurrent neural networks, such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), provide strong sequence modeling capability, but their dense hidden-to-hidden transformations introduce considerable parameter storage and vector-matrix multiplication overhead. In addition, the serial dependency between gate-related computation and candidate hidden-state computation increases the intra-step recurrent delay, which limits their efficient deployment in resource-constrained edge systems. Memristor-based in-memory computing can reduce data movement and accelerate vector-matrix operations, but directly mapping conventional gated recurrent structures onto memristor circuits still suffers from complex recurrent paths and peripheral circuit overhead. Therefore, a recurrent structure that jointly considers lightweight recurrence, intra-step parallelism, and hardware implementation efficiency is needed. To address this issue, this paper proposes a Point-Wise Single-Gate Recurrent Unit (P-SGRU) and a corresponding memristor-based in-memory computing circuit for efficient edge temporal inference. Methods P-SGRU is designed to simplify the recurrent path of conventional gated recurrent units. It replaces the fully connected hidden-to-hidden transformation with point-wise recurrent modulation, so that each hidden-state channel is recurrently modulated only by its corresponding channel at the previous time step. A single retention coefficient is introduced to control the fusion of the previous hidden state and the candidate hidden state. By removing the reset-gate-related dependency in GRU, the retention-coefficient branch and the candidate hidden-state branch can be computed in parallel within the same time step, thereby improving intra-step computational parallelism. According to the P-SGRU computation process, a memristor-based in-memory computing circuit is designed, consisting of a retention-coefficient computation unit, a candidate hidden-state computation unit, and a state update unit. Signed weights are represented by positive-negative differential memristor conductance arrays, and point-wise recurrent weights are mapped only to their corresponding channel positions. To reduce the mismatch between software training and circuit implementation, PWL activation-aware training is adopted, where PWL-Sigmoid and PWL-Tanh functions consistent with the circuit implementation are used during training. After training, the weights are quantized to 8 bits and mapped to memristor conductance states. The proposed model is evaluated on Adding Problem, Row-wise MNIST, IMDB, and UCI HAR. For UCI HAR, the original nine-channel smartphone inertial signals with 128 time steps are used to evaluate six-class human activity recognition in a realistic edge-sensing scenario. The corresponding memristor-based circuit is analyzed under a 40 nm process condition. Results and Discussions Compared with LSTM and GRU, the two-layer P-SGRU reduces the number of hidden-to-hidden recurrent parameters by 98.44% and 97.92%, respectively, and removes full hidden-to-hidden vector-matrix multiplication branches (Table 2 ). On the tested sequence tasks, P-SGRU-2L achieves 98.21%±0.20% accuracy on Row-wise MNIST and 86.85%±0.13% accuracy on IMDB, while the single-layer P-SGRU is selected for circuit implementation considering the model scale and circuit overhead of resource-constrained edge systems (Table 2 ). On UCI HAR, P-SGRU-2L achieves 91.87%±0.94% accuracy and 91.94%±0.94% macro-F1 with 10,246 parameters (Table 2 ). After being mapped to the proposed memristor-based circuit, the single-layer P-SGRU achieves circuit-level accuracies of 95.03%±0.06% and 86.76%±0.05% on Row-wise MNIST and IMDB, respectively, with a maximum accuracy loss of only 0.02% compared with the software model. Under the 40 nm process condition, the proposed P-SGRU circuit shows lower or comparable delay under different task configurations (Table 3 ). In addition, the single-channel, single-time-step update simulation shows that the critical-path delay is reduced from 524.5 ps in GRU to 177.9 ps in P-SGRU, corresponding to a 66.1% reduction, verifying the effect of removing the serial dependency between the retention-coefficient branch and the candidate hidden-state branch. Conclusions This paper addresses efficient temporal inference in resource-constrained edge systems by proposing P-SGRU and its memristor-based in-memory computing circuit implementation. The proposed P-SGRU replaces the fully connected hidden-to-hidden transformation with point-wise recurrent modulation, thereby reducing the parameter scale and computational complexity of the recurrent path. Meanwhile, by decoupling the retention-coefficient branch from the candidate hidden-state branch, P-SGRU improves intra-step computational parallelism and shortens the recurrent update path. Experimental results on Adding Problem, Row-wise MNIST, IMDB, and UCI HAR show that P-SGRU maintains favorable sequence modeling performance while significantly simplifying the recurrent structure. In particular, the UCI HAR results demonstrate that the stacked P-SGRU can effectively model real multi-channel inertial sequences while retaining a lower parameter count than conventional gated recurrent models. Circuit-level evaluation further demonstrates that the P-SGRU circuit after hardware mapping achieves inference results close to those of the software model, and the single-channel single-time-step update critical-path delay is reduced by about 66.1% compared with GRU under the 40 nm process condition. These results indicate that P-SGRU provides a lightweight and efficient algorithm-circuit co-design solution for edge temporal inference. Future work will further extend P-SGRU to more complex edge temporal tasks and explore its integration with high-density in-memory computing arrays and edge intelligence systems for low-power and low-latency temporal inference.
Available online , doi: 10.11999/JEIT260661
Abstract:
Objective Visible-infrared person re-identification (VI-ReID) aims to retrieve the same pedestrian identity across heterogeneous visible and infrared cameras, and is a key technique for all-day intelligent surveillance and cross-modal retrieval. The task remains challenging because visible images are formed by reflected light and carry rich color and texture, whereas infrared images are formed by thermal radiation and mainly describe body contours and thermal responses, which causes a large spectral gap and inconsistent local appearance. Viewpoint changes, pose variations, occlusion, and detection misalignment further weaken the spatial correspondence between body parts. Convolutional methods extract local cues effectively but describe global body topology poorly, while Transformer-based methods model long-range dependencies at a quadratic cost that limits practical deployment. More importantly, most existing methods align features only in the spatial domain and ignore the different roles of frequency components: low-frequency components preserve the body shape and structural layout shared by both modalities, whereas high-frequency components mainly carry visible texture, illumination edges, infrared thermal noise, and other modality-specific disturbances. To address these limitations, this paper proposes a frequency-decoupled state space model, termed FD-Mamba, that jointly enhances spatial structural modeling and frequency-domain cross-modal alignment. The proposed design is further motivated by the observation that cross-modal discrepancy does not affect all frequency components equally. Separating shared structural information from modality-specific responses therefore provides a more targeted way to improve modality invariance than treating all spectral components uniformly. Methods As illustrated in Fig. 1 , FD-Mamba is built upon a dual-stream ResNet-50 backbone. Shallow stages use modality-specific branches to preserve modality-dependent low-level cues, while deeper stages share weights to project both modalities into a common semantic space. Two modules are introduced. First, a Hybrid Spatial Mamba Attention (HSMA) module is designed for deep semantic features, as shown in Fig. 3 . HSMA combines multi-scale convolution with two-dimensional selective scanning: the convolution branch extracts local body-part patterns under different receptive fields, while the scanning branch propagates contextual information along multiple spatial directions, following the selective scan of Fig. 2 , so that long-range interactions among the head, torso, legs, and carried objects are captured and more complete identity-related representations are obtained without relying only on isolated local textures. Second, a Cross-Scale Frequency Mamba Enhancement (CSFME) module is constructed for shallow and intermediate features, as shown in Fig. 4 . CSFME divides feature channels into groups and projects them into the frequency domain through multi-scale two-dimensional Discrete Cosine Transform. The DCT coefficients of each group are partitioned into zigzag-ordered frequency bands and reweighted by a learnable band-gating mechanism, so that high-frequency modality noise is suppressed softly rather than truncated. The reweighted bands are then coupled by a frequency-aware state-space process, transformed back to the spatial domain by inverse DCT, and recalibrated by pixel and channel attention. Modality-specific noise is thus suppressed and low-frequency shared structural cues enhanced, providing more robust features for subsequent spatial alignment. The whole network is trained end-to-end with the identity, metric, and prototype-related constraints inherited from the baseline protocol. Both modules are inserted into the original backbone in a residual manner, so the enhanced representations retain the original identity information while introducing complementary spatial and frequency cues. This design also limits the additional computational cost and preserves the overall training protocol of the baseline. Results and Discussions Experiments are conducted on three representative VI-ReID benchmarks, SYSU-MM01, RegDB, and LLCM, which cover large-scale cross-camera retrieval, bidirectional visible-infrared matching, and low-light cross-modal scenarios, respectively. As reported in Table 1 , FD-Mamba achieves 79.71% Rank-1 and 77.52% mAP on SYSU-MM01 under the all-search mode, improving the strong PDM baseline by 2.76 and 3.27 percentage points and indicating that the spatial-frequency collaborative design benefits both top-ranked accuracy and overall ranking quality. Competitive performance is also obtained under indoor-search. On RegDB, FD-Mamba achieves 95.2% Rank-1 and 89.3% mAP in the VIS to IR direction, and 94.7% Rank-1 and 87.5% mAP in the IR to VIS direction. The close results in both directions suggest that the model learns a relatively symmetric cross-modal representation rather than relying on a fixed query modality. The retrieval examples in Fig.5 (b) further show that FD-Mamba reduces incorrect top-ranked matches compared with the baseline model. As shown in Table 2 , FD-Mamba remains stable on the more challenging LLCM dataset, obtaining 65.3% Rank-1 and 67.8% mAP for Visible-to-Infrared retrieval and 57.5% Rank-1 and 63.9% mAP for Infrared-to-Visible retrieval, indicating robustness under low illumination, background clutter, and local texture degradation. Ablation studies in Tables 4 and 5 show that HSMA improves Rank-1 and mAP by 1.15 and 2.50 percentage points and CSFME by 0.90 and 1.25 percentage points, while their combination performs best, demonstrating the complementarity between global spatial topology modeling and frequency-domain noise suppression. Spectral statistics and band-limited recognition experiments quantitatively support the frequency-decoupling assumption. Additional analyses show that the gains are not produced by a single component. The frequency-domain experiments indicate that low-frequency information is more stable across modalities, whereas high-frequency responses contain stronger modality-specific variations, which is consistent with the improvements brought by learnable frequency-band reweighting. FD-Mamba contains 42.6M parameters and 8.51G FLOPs, only 1.83% and 2.90% more than the baseline, with comparable throughput and latency on an RTX 4090 GPU. The attention maps in Fig. 5(a) show more continuous pedestrian foreground responses than the baseline, and the feature distance distributions and t-SNE visualization in Fig. 6 further reveal smaller intra-class cross-modal distances and clearer inter-class margins, indicating improved identity discrimination and modality consistency. Conclusions This paper presents FD-Mamba, a frequency-decoupled state space model for visible-infrared person re-identification. The HSMA module improves global body-structure modeling through hybrid spatial scanning, and the CSFME module introduces explicit frequency decomposition to suppress modality-specific disturbances while preserving shared structural information. Quantitative comparisons, ablation studies, and visual analyses consistently show that the proposed spatial-frequency collaborative framework improves both identity discrimination and cross-modal consistency, making it suitable for all-day cross-modal pedestrian retrieval and suggesting that frequency-aware state-space modeling is a promising direction for heterogeneous visual representation learning.The current framework is still evaluated mainly under supervised benchmark protocols. Future work will investigate unsupervised cross-modal settings and more lightweight state-space designs to further improve generalization and deployment efficiency in practical surveillance scenarios.
Available online , doi: 10.11999/JEIT260266
Abstract:
Objective Visual Simultaneous Localization and Mapping (SLAM) is a fundamental technology for autonomous navigation and environmental perception in Global Navigation Satellite System (GNSS)-denied environments. However, conventional visual SLAM systems generally rely on a static-scene assumption. In dynamic environments containing pedestrians, vehicles, and other moving objects, dynamic feature mismatches may violate geometric constraints and introduce unstable observations into pose estimation, resulting in trajectory drift or tracking degradation. Existing dynamic SLAM methods commonly incorporate semantic priors to suppress such interference. Semantic segmentation-based methods can provide fine-grained pixel-level dynamic regions, but usually require relatively high computational cost. In contrast, object detection-based methods provide higher inference efficiency, but their rectangular detection boxes inevitably contain static background regions, which may lead to the erroneous removal of valid static features and cause an over-masking problem. This issue becomes more pronounced in texture-poor or crowded scenes where sufficient static features are important for maintaining stable geometric constraints. To alleviate the trade-off between dynamic-feature filtering accuracy and computational efficiency, this paper proposes Depth-Consistency and Progressive-Distillation SLAM (DCPD-SLAM), a lightweight dynamic visual SLAM system for real-time state estimation under moderate computational resources. Methods DCPD-SLAM adopts a parallel dual-thread architecture comprising semantic perception and visual SLAM tracking. Semantic inference and feature extraction run concurrently, with frame-level synchronization before dynamic-feature filtering to align semantic results with the current Red-Green-Blue-Depth (RGB-D) frame. First, a depth-aware masking strategy refines coarse detection boxes. For selected non-rigid dynamic targets, such as pedestrians, valid depth pixels within the Region of Interest (ROI) are used to estimate a robust reference depth. The reference depth combines the median and the lower 33rd percentile of valid depth measurements to reduce the influence of background pixels and isolated depth noise. An adaptive depth-consistency threshold is then used to generate an initial foreground mask. When the valid-depth ratio is insufficient, the original detection box is retained as a fallback to maintain filtering reliability. Connected Component Analysis (CCA) and morphological dilation suppress isolated noise and compensate for boundary leakage, yielding a compact pixel-level pseudo-mask. Other predefined dynamic categories retain detection-box-based filtering to avoid additional mask-generation overhead.Second, a lightweight detector, Light-DEIM, reduces semantic perception costs. Partial Convolution (PConv) is incorporated into shallow C2f-FasterBlock stages to reduce spatial redundancy, while deeper C2f-Inverted Residual Mobile Block (iRMB) modules with window attention preserve semantic representations. Balanced width scaling further reduces model complexity while retaining sufficient high-level semantic capacity. Finally, progressive knowledge distillation compensates for representation loss caused by model lightweighting. Masked Generative Distillation (MGD) aligns intermediate features, while DETRDistill-based logits distillation aligns teacher–student predictions. During the first 70% of training epochs, Light-DEIM is trained with ground-truth supervision and feature and output distillation from a frozen DEIM-N teacher. For the remaining epochs, the distillation branches are disabled, and the student is optimized solely with ground-truth labels. Results and Discussions Experiments are conducted on the TUM RGB-D and Bonn RGB-D dynamic datasets. For the detection network, Light-DEIM reduces the parameter count from 4.0 M to 3.0 M and the computational complexity from 7.1 Giga Floating-point Operations (GFLOPs) to 5.6 GFLOPs. Without knowledge distillation, the lightweight model achieves an \begin{document}$ \text{mAP}_{50-95}^{\text{val}} $\end{document} \begin{document}$ \text{mAP}_{50-95}^{\text{val}} $\end{document} Table 1 ). Ablation experiments further show that PConv mainly contributes to computational efficiency, whereas iRMB and Window Attention improve semantic representation. The combination of feature-level and output-level distillation provides better accuracy recovery than either distillation term alone (Table 2 , Table 3 ).For SLAM evaluation, DCPD-SLAM achieves an Absolute Trajectory Error (ATE) of 0.019 \begin{document}$ \pm $\end{document} \begin{document}$ 0.019\pm 0.002 $\end{document} Table 4 ). The trajectory-error curves and error distributions further indicate relatively stable localization performance in highly dynamic scenes (Fig. 5 , Fig. 6 ). In four representative dynamic sequences, the proposed depth-aware masking strategy achieves an average dynamic-point rejection rate of 95.4% while limiting the static-point erroneous rejection rate to 12.4% (Table 6 ), indicating that it can reduce unnecessary removal of static background features while maintaining effective dynamic-feature filtering. At the complete-system level, DCPD-SLAM achieves average processing rates of 24.71 FPS and 27.11 FPS on the TUM RGB-D and Bonn RGB-D datasets, respectively (Table 7 ), demonstrating real-time processing capability on the moderate-compute Graphics Processing Unit (GPU) platform. Conclusions DCPD-SLAM balances localization accuracy and computational efficiency in dynamic RGB-D environments. Depth-aware masking mitigates over-masking from detection-box filtering, preserving static background features. Light-DEIM and progressive knowledge distillation reduce semantic perception overhead while retaining detection capability for dynamic-feature filtering. Experiments demonstrate real-time processing and stable localization on a moderate-compute GPU platform. However, pseudo-mask quality depends on depth measurements, with potential degradation under severe depth noise, strong illumination changes, reflective or transparent surfaces, distant targets, and severe occlusions. Future work will explore Light Detection and Ranging (LiDAR)-visual fusion to improve adaptability and measurement stability in complex, large-scale environments.
Available online , doi: 10.11999/JEIT260496
Abstract:
Objective Millimeter-wave (mmWave) bands offer abundant spectrum resources, making them ideal for high-speed indoor transmission and high-precision sensing applications. However, lacking sufficient empirical validation from ultra-wideband measurements, the understanding of how dense multipath components (DMCs) affect channel sparsity and multipath parameter extraction in complex indoor Non-Line-of-Sight (NLOS) environments remains limited. This paper conducts a comprehensive channel measurement and characterization campaign to investigate physical propagation mechanisms and evaluate the indoor positioning potential of ultra-wideband massive MIMO systems. Methods A massive MIMO channel measurement platform is established using a Vector Network Analyzer. The system operates at a center frequency of 32 GHz with a 16 GHz bandwidth. On the receiver side, a precision turntable constructs a 16×16 virtual Uniform Rectangular Array (URA). The measurement campaign covers a typical indoor environment featuring both Line-of-Sight (LOS) and NLOS scenarios. Key channel characteristics, including path loss, power delay profile, delay spread, and delay-wavenumber spectrum, are extracted. Furthermore, a multipath-assisted single-point positioning framework based on virtual anchors is implemented to evaluate localization performance across different measurement bandwidths. Results and Discussions Measurement results reveal significant differences in multipath structures and energy distributions between LOS and NLOS channels. In the time domain, LOS channels exhibit strong spatiotemporal sparsity. Conversely, obstructed NLOS environments are dominated by DMCs due to severe physical blockage and complex multipath effects, resulting in significant energy tails and large delay spreads (Fig. 8 , Fig. 10 ). Spatially, the delay-wavenumber spectrums confirm the high angular resolution for user separation (Fig. 11 ). Crucially, the comparative positioning experiments across different measured bandwidths demonstrate that the 16 GHz ultra-wideband effectively decouples dense adjacent multipath components, significantly improving the accuracy of multipath parameter extraction. This provides robust virtual anchors for multipath-assisted positioning. The 16 GHz system achieves a low mean positioning error of 0.052 m, showcasing significant robustness and accuracy advantages over the narrower 8 GHz and 4 GHz bandwidths (Table 5 ). Conclusions This study provides a systematic characterization of indoor mmWave massive MIMO channels. The results identify that channel sparsity is highly scenario-dependent, necessitating the inclusion of DMCs in future NLOS channel modeling. Moreover, the empirical localization analysis confirms that ultra-high spatiotemporal resolution is indispensable for robust indoor multipath extraction and positioning applications, providing a solid experimental foundation for next-generation network design.
Available online , doi: 10.11999/JEIT260534
Abstract:
Objective Long-text strong-constraint matching is required in policy-to-enterprise recommendation, contract verification, and compliance screening. Categorical conditions, numerical thresholds, eligibility clauses, temporal scopes, and exclusion rules must be satisfied simultaneously, and violation of any hard constraint disqualifies a candidate. As large language models (LLMs) are applied to these decisions, attributes, comparison operators, and thresholds need to be preserved consistently over long contexts. Prevailing pipelines, including zero-shot prompting, retrieval-augmented generation (RAG), and chain-of-thought (CoT) reasoning, pass intermediate evidence mainly through semantic similarity or free-text steps, and therefore provide limited support for item-by-item verification of hard constraints. Open matching benchmarks likewise give limited coverage to continuous thresholds, exclusion clauses, and non-compensatory rules. A factorized multi-agent associative matching method, termed Factorized Multi-Agent Associative Matching (FMAM), is proposed. Constraint conditions and factual evidence are structured as constraint factors and fact factors, which serve as the information carriers among agents. Non-compensatory decision is realized by coupling hard-constraint gating with multiplicative aggregation. Methods FMAM separates factor extraction from deterministic evaluation (Fig. 1 ). A Constraint Decomposition Agent (CDA) parses each source text once and writes every constraint into a field-level record, including the attribute, comparison operator, threshold, unit, type, and hard-constraint flag. A Feature Guidance Agent (FGA) collects attribute patterns from all source-side factors and extracts the corresponding fact factors from each candidate text once. The extraction is guided by these attribute patterns and is decoupled from pair-specific thresholds. The constraint-factor set of a source text is reused across all of its pairings, and the fact-factor set of a candidate text is reused across all of its pairings. For \begin{document}$ m $\end{document} \begin{document}$ n $\end{document} \begin{document}$ m\times n $\end{document} \begin{document}$ m+n $\end{document} Results and Discussions The method is evaluated on a constructed policy-enterprise matching set of 50,000 pairs (Table 1 ), comprising 50 policy texts and 1,000 enterprise texts written in official-document and research-report styles. Field-level ground truth is determined from policy rules and enterprise indicators, and positive pairs account for 8.83% of the set. All methods use Gemini-2.5-Flash with the temperature set to 0. As shown in Table 2 and Fig. 2 , FMAM obtains an accuracy of 0.9835 , a precision of 0.9523 , a recall of 0.8562 , an F1-score of 0.9017 , an AUC-PR of 0.8567 , and an MCC of 0.8942 , and achieves the highest accuracy, precision, F1-score, AUC-PR, and MCC among the compared methods. The F1-scores of few-shot CoT, Standard RAG, and GraphRAG are 0.4559 , 0.1825 , and 0.2452 , respectively. The NCER of FMAM is 4.87%, compared with 27.25% for the best baseline (Table 5 ). The average inference cost is 24.22 tokens per pair (Table 6 ), which is 1.27% of that of zero-shot LLM and 0.43% of that of ReAct. Standard RAG and GraphRAG consume 8.17 times and 28.83 times as many tokens as FMAM, respectively. Ablation studies show that replacing structured constraint factors with natural-language intermediate representations reduces the F1-score from 0.9017 to 0.3555 and raises the NCER from 4.87% to 77.57% (Table 4 ). Removing hard-constraint gating reduces the precision from 0.9523 to 0.7277 (Table 3 ). Conclusions Structured constraint factors combined with deterministic non-compensatory decision are effective for long-text strong-constraint matching. Field-level factors keep constraint information consistent across agents. Hard-constraint gating maintains a high precision under non-compensatory eligibility rules. Concentrating LLM calls in the extraction stage and performing pairing on structured factors reduces the inference cost from a multiplicative scale to a linear scale. Future work will extend the method to cross-institution texts and online calibration for broader compliance-screening applications.
Available online , doi: 10.11999/JEIT260368
Abstract:
Objective The rapid deployment of large-scale Low Earth Orbit (LEO) satellite constellations has intensified competition for orbital and spectrum resources. To improve spectrum utilization, different LEO satellite constellations commonly employ co-frequency reuse, which increases the risk of inter-system co-frequency interference. The high-speed motion of LEO satellites and their highly dynamic, heterogeneous topology further cause rapid variations in interference power, resulting in severe co-frequency interference. Existing interference assessment methods mainly rely on the regulations of the International Telecommunication Union (ITU), with the Interference-to-Noise Ratio (I/Noise) as a key evaluation metric. However, conventional ITU-based physical-iteration methods require repeated calculations of satellite positions, link attenuation and interference contributions for all visible satellites, causing computational cost to increase rapidly with constellation size. A complete interference assessment for a single ground station in a constellation of approximately 10 000 satellites can require more than 30 h. Recent deep-learning-based methods still commonly use all visible satellites as inputs and therefore do not adequately exploit the spatial sparsity of interference features. To address these limitations, an LEO constellation interference prediction and detection method driven by collaborative spatial feature mapping and temporal Transformer is proposed to reduce computational cost while maintaining accurate interference prediction and detection. Methods A spatiotemporal framework is developed to model dynamic co-frequency interference in large-scale heterogeneous multi-constellation LEO networks. Based on the ITU physical interference model, three key properties of the interference function are identified: permutation invariance, spatial sparsity and temporal continuity. The conventional physical-iteration process is therefore reformulated as a spatiotemporally decoupled feature-mapping problem. A spatial feature mapping module with adaptive attention and symmetric pooling is constructed to compress unordered and variable-length satellite interference-source features into fixed-dimensional, permutation-invariant spatial features. The attention mechanism adaptively focuses on dominant interference sources, whereas symmetric pooling using maximum and mean pooling captures extreme and global statistical characteristics while suppressing redundant satellite nodes. A temporal Transformer is then employed to model the long-range evolution of interference trajectories. The future I/Noise trajectory is predicted to support rapid interference detection under dynamic constellation configurations. Results and Discussions A large-scale heterogeneous LEO constellation scenario consisting of 6 800 Starlink satellites and 650 OneWeb satellites is simulated according to ITU regulations. Real Two-Line Element (TLE) data are used to propagate satellite orbits, and 100,000 time-series samples are generated at 0.5 s intervals. Parameter analysis shows that appropriate feature dimensions and encoder depths provide a favorable balance between feature extraction accuracy and computational cost (Figs. 4 and 5 ). The sampling interval and sliding-window length are further optimized to balance prediction accuracy and real-time inference performance (Figs. 6 and 7 ). Compared with the baseline methods, the proposed method produces interference trajectories that closely follow the physical-iteration ground truth (Fig. 8 ). At a cumulative probability of 90%, the absolute prediction error is maintained within 0.5 dB (Fig. 9 ). The method also maintains a high interference recall under a low false-alarm-rate constraint (Fig. 10 ). For a 20.0 s prediction horizon, the Root Mean Square Error (RMSE) remains at 0.45 dB, substantially lower than those of the baseline models. At a 20 s prediction step, the RMSE values of the Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) network increase to 5.53 dB and 4.82 dB, respectively (Fig. 11 ). In terms of computational efficiency, the proposed method requires 14.2 ms for a single inference, which is only 4.1% of the computational time required by the ITU-based physical-iteration method (Table 3 ). The computational cost is therefore effectively decoupled from constellation size. Conclusions To address the high computational cost caused by the highly dynamic topology of LEO satellite networks, a collaborative spatial feature mapping and temporal Transformer-based interference prediction and detection method is proposed. The spatial feature mapping module compresses variable-length interference-source features into fixed-dimensional representations, whereas the temporal Transformer captures the long-range evolution of interference trajectories. Simulation results show that the proposed method provides accurate long-term trajectory tracking while remaining consistent with ITU-based interference assessment. Parameter-sensitivity experiments demonstrate that the proposed method can balance feature extraction accuracy and computational cost under different configurations. With its long-range dependency modeling capability, the temporal Transformer maintains an RMSE of 0.45 dB over a 20 s prediction horizon. By filtering redundant satellite nodes and decoupling computational cost from constellation size, the method reduces single-inference time to 14.2 ms. Comprehensive evaluations of prediction error distributions, detection performance and model parameters demonstrate that the proposed method achieves high prediction accuracy and substantially reduced computational complexity, providing a flexible engineering solution for interference monitoring in large-scale LEO constellation deployments.
Available online , doi: 10.11999/JEIT260009
Abstract:
Objective .Traditional single-modal speech separation fails to fully exploit modal complementarity and performs poorly under dynamic occlusion and homophonic interference. Bimodal audio-visual models suffer severe performance drops when lip visual cues are unavailable. Current tri-modal methods mostly simply concatenate text features, lacking systematic cross-modal alignment and interaction. This work targets a tri-modal speech separation method to fully mine multi-modal complementary information and achieve stable performance in challenging real scenarios. Methods This paper proposes a tri-modal speech separation model (GA-CMA) driven by improved multi-head attention and gating-augmented cross-modal attention. First, a tri-modal feature extraction module is constructed. The convolutional neural network captures acoustic speech features (Fig. 2 ). Lip-motion visual features are extracted via optimized ResNet34 after lip-region preprocessing (Fig. 3 ). Semantic text features are obtained from the BERT model together with an adaptive semantic correction branch for ASR error tolerance. Second, a gating-augmented cross-modal attention layer is designed. It dynamically generates gating vectors relying on self-characteristics of the receiving modality to filter critical cross-modal complementary information and mitigate redundancy brought by naive feature concatenation. Improved multi-head attention is further adopted to realize deep tri-modal feature fusion (Fig. 1 ). Finally, the separation module guided by fused features completes target speech extraction. Experiments are conducted on the GRID dataset, with PESQ, STOI and SDR adopted as evaluation metrics. Results and Discussions Compared with the baseline tri-modal naive-fusion model, the proposed method achieves minimum improvements of 4.5 dB for SDR, 0.47 for PESQ and 0.081 for STOI. The final implemented model obtains SDR of 7.82 dB, PESQ of 2.08 and STOI of 0.870. The gating-augmented cross-modal attention adaptively suppresses low-quality modal signals instead of adopting fixed fusion weights. Benefiting from the semantic compensation of text modality, the system retains stable separation performance even when visual lip features suffer partial dynamic occlusion. Conclusions The presented GA-CMA method can well satisfy dynamic feature requirements under complex conditions. It provides an alternative multi-modal fusion solution for speech separation tasks. The receiving-modality-driven gating mechanism offers a feasible reference for subsequent tri-modal signal processing research.
Available online , doi: 10.11999/JEIT260777
Abstract:
Objective In modern block ciphers, the substitution box (S-box) is the only nonlinear component, and its cryptographic properties directly determine the resistance of a cryptosystem against linear and differential cryptanalysis. Hence, the efficient generation of S-boxes with strong cryptographic strength remains a central challenge, in which a trade-off among cryptographic quality, generation quantity, and generation efficiency has to be achieved. Conventional methods are classified into algebraic construction, chaos-based construction, and heuristic search. Algebraic methods yield S-boxes with excellent and predictable properties, but their design procedure is complicated and the number and diversity of the generated S-boxes are severely limited. Chaos-based methods produce numerous samples rapidly, yet their nonlinearity is usually inferior and the strict avalanche criterion (SAC) and the bit independence criterion (BIC) are difficult to satisfy simultaneously. Heuristic search methods, including simulated annealing, particle swarm optimization (PSO), and hill climbing, obtain high-performance S-boxes at excessive computational cost, so real-time dynamic application is not supported. Recently, artificial intelligence has opened a new path. Meta-heuristic algorithms still rely on iterative search, reinforcement learning (RL) optimizes masking efficiency rather than direct cryptographic metrics, and generative adversarial network (GAN) based methods, despite incorporating cryptographic properties into the loss of the Wasserstein GAN with gradient penalty (WGAN-GP), achieve a maximum nonlinearity of only about 104. Moreover, the computational overhead of metric evaluation remains the core bottleneck, and training instability, mode collapse, and the difficulty of integrating non-differentiable cryptographic properties into the optimization objective have not been resolved. To address these issues, a performance-guided intelligent S-box generation algorithm based on WGAN-GP is proposed, by which stable training is inherited, non-differentiable cryptographic metrics are directly embedded into the optimization objective, and massive, efficient, and high-strength S-box supply is realized for dynamic cipher systems. Methods The framework follows the logic of "data driving, adversarial learning, performance guidance, and batch generation" and consists of three cooperating modules. In the dataset preparation module, a large-scale dataset is constructed by applying random affine transformations to the standard AES S-box, for which random 8x8 invertible matrices and random 8-bit constant vectors are generated over GF(2) and used to pre- and post-process the original S-box; after normalization, the resulting variants are fed into the discriminator as real samples (Fig. 1 ). In the adversarial training module, a 256-dimensional noise vector is mapped by the generator to a 256-dimensional floating-point vector through fully connected layers with batch normalization and LeakyReLU, while the discriminator outputs a real-valued score through fully connected layers with LayerNorm, LeakyReLU, and Dropout (Fig. 2 ). The Wasserstein distance is combined with a gradient penalty that enforces Lipschitz continuity by encouraging the gradient norm at sampled points to approach 1, and the penalty weight is set to 10. The adversarial loss of the generator is the negative expectation of the discriminator score. The nonlinearity (NL) of an S-box is defined as the mean of the nonlinearities of its coordinate Boolean functions computed by the Walsh-Hadamard transform (WHT), and the differential uniformity (DU) is the maximum value in the difference distribution table (DDT). The cryptographic loss is the mean squared error between the estimated metrics and the ideal targets, normalized by the maximum nonlinearity of 112 and by 252, the range above the ideal differential uniformity of 4 that corresponds to almost perfect nonlinear (APN) functions; the total loss is a weighted sum with weights of 0.8 and 1.5. During the first 30 warm-up epochs, only the adversarial loss is used, after which the cryptographic loss is smoothly introduced. A post-processing step based on the argsort operation converts the continuous output into a permutation of the integers from 0 to 255, by which bijectivity is guaranteed; the resulting continuous-discrete gap is alleviated by the gradient penalty and by the cryptographic losses computed directly on discrete S-boxes. A GPU-accelerated parallel estimation method based on PyTorch tensor operations is developed to incorporate non-differentiable metrics into training: for NL, an eighth-order Hadamard matrix is precomputed, 1024 rows are randomly sampled, and all WHT coefficients are computed concurrently by torch.einsum; for DU, the DDT is constructed by broadcasting, advanced indexing, and torch.scatter_add, and its maximum is taken as the estimate. Python loops are eliminated in both procedures. After training, a two-stage strategy is adopted, in which ten million candidates are generated and rapidly screened by the GPU estimator against thresholds of NL greater than 100 and DU no more than 10, and the exact metrics of the top 1000 elite candidates are then computed by a NumPy-based CPU function (Fig. 3 ). Results and Discussions The effectiveness of the algorithm is verified in terms of efficiency, quality, and structural diversity. During training, the loss curves converge smoothly without the severe oscillation common in conventional GAN training, and the estimated cryptographic metrics increase steadily once the performance guidance mechanism takes effect after the warm-up period (Fig. 4 ). In terms of efficiency, only 382 minutes are required to generate ten million S-boxes, with an average of 0.0023 seconds per S-box, whereas 6520 minutes are needed without the estimation strategy, so an improvement of 94.2% is achieved (Table 1 ). Scalability is verified on three GPU platforms, on which the per-S-box times are 0.0023 , 0.0011 , and 0.0004 seconds and the total times are 382, 142, and 67 minutes, confirming that the efficiency originates from the algorithm rather than the hardware (Table 1 ). In terms of quality, the top 1000 candidates achieve an average NL of 104.88 with a maximum of 107.5 close to the theoretical maximum of 112, while all values are no less than 100; the mean per-S-box minimum nonlinearity is 102.10 with the worst case of 100.0, the average DU is 9.50 with a minimum of 8, the average SAC deviation is 0.0085 , the average BIC-SAC is 0.4979 , and the average BIC-nonlinearity is 102.26, indicating excellent diffusion and output bit independence (Table 2 ). The frequency distributions show that a large number of S-boxes are concentrated in high-performance regions, which verifies both effectiveness and mode coverage (Fig. 5 ). Bijectivity is satisfied by all ten million S-boxes, and the estimator is validated on 1000 samples, for which the NL error is below 0.5% and the DU estimate matches the exact value for 98% of the samples (Table 2 ). The continuous-discrete gap is examined, and the estimated NL of 104.7 differs from the exact value of 104.88 by only 0.17% while the estimated DU of 9.6 differs from the exact value of 9.50 by only 1.04%, so no inconsistency between training and inference is observed (Fig. 4 ). Unlike AES affine-equivalent S-boxes whose DU is always 4 and whose NL is always 112, the generated S-boxes exhibit an average DU of 9.50 with a minimum of 8 and an average NL of 104.88 with a minimum of 100.0, and their Walsh spectral statistics confirm that genuine exploration rather than memorization is performed by the model (Table 3 ). Existing methods fall into an objective-oriented paradigm and a data-driven paradigm, whose merits cannot be judged solely by identical metrics; the value of this work lies in the massive supply scale and the order-of-magnitude gain in efficiency obtained at the cost of only a minor concession in individual metrics (Table 2 ). Conclusions A performance-guided WGAN-GP framework for massive S-box generation is proposed, by which stable generation is realized, non-differentiable cryptographic properties are directly embedded into the optimization objective, and the bottleneck of metric evaluation is overcome by a GPU-accelerated estimation strategy that improves generation efficiency by 94.2%. A dynamic supply scheme compatible with standard key update protocols is also presented. Ten million S-boxes are generated within 382 minutes with an average of 0.0023 seconds per S-box, and the top 1000 exactly evaluated S-boxes achieve an average NL of 104.88 with a maximum of 107.5 and an average DU of 9.50 with a minimum of 8, meeting or exceeding the level of existing dynamic S-box generation algorithms. Meanwhile, training instability, mode collapse, and the difficulty of integrating non-differentiable properties are effectively alleviated. The potential of deep generative learning for cryptography is demonstrated, and the algorithm can be extended to other cryptographic primitives, with applications in lightweight cipher design for the Internet of Things and mobile platforms and in high-security systems where S-boxes are frequently updated to resist side-channel attacks.
Available online , doi: 10.11999/JEIT260859
Abstract:
Objective Semi-supervised learning effectively alleviates the reliance on scarce pixel-level annotations in medical image segmentation, yet confirmation bias and error amplification caused by boundary errors and topological distortions in pseudo-labels severely limit the performance of conventional methods. Furthermore, traditional approaches often rely on shallow confidence scores, lacking the ability to explicitly reconstruct correct anatomical structures. This study develops a Prototype Uncertainty-driven latent Diffusion (PU-Diff) framework that utilizes semantic prototypes to quantify spatial uncertainty, explicitly forces the denoising process to rectify low-confidence regions, and aims to achieve precise error localization and active correction of complex anatomical structures under limited supervision. Methods Alongside the conventional consistency constraints, PU-Diff introduces an independent “predict, correct, and feedback” reverse path for pseudo-label rectification. Within this path, a Semantic Prototype Uncertainty Quantification (SPUQ) module constructs category prototypes utilizing high-confidence predictions. It then quantifies the classification ambiguity by calculating the distances in the high-dimensional feature space between unlabeled pixels and these prototypes, thereby generating a reliable spatial uncertainty guidance map. Subsequently, an Uncertainty-Aware Diffusion Rectification (UADR) module integrates this uncertainty map and the original image as joint spatial conditions to guide a latent diffusion model. This mechanism explicitly forces the denoising process to focus on repairing boundary and topological errors in low-confidence regions, ultimately generating high-fidelity pseudo-labels to provide reliable supervision signals for the main network's optimization. Results and Discussions Extensive experiments are conducted on ACDC, MS-CMRSeg, and MSD Prostate and Kvasir-SEG datasets. PU-Diff achieves clear performance gains. On the ACDC dataset with 5% and 10% labeled data, the average Dice coefficients reach 89.43% and 90.72%, respectively. On the MS-CMRSeg dataset with 10% labeled data, the average Dice reaches 88.06%, and on the MSD Prostate dataset with 10% labeled data, the average Dice reaches 72.64%, and on the Kvasir-SEG dataset with 10% labeled data, the average Dice reaches 86.61%. Ablation studies validate that the joint mechanism of SPUQ and UADR effectively eliminates structural noise and mends topological fractures. In terms of efficiency, PU-Diff achieves the optimal accuracy with a reasonable parameter count of 19.28 million and a single inference time of 3.56 ms, demonstrating a superior balance between segmentation accuracy and computational efficiency compared to existing advanced methods. Conclusions A semi-supervised medical image segmentation framework, PU-Diff, that integrates semantic prototype uncertainty quantification and latent diffusion-based label rectification is presented. PU-Diff mitigates confirmation bias by using high-dimensional feature prototypes to precisely locate error-prone areas, and addresses error accumulation by establishing an independent closed-loop that actively repairs topological distortions and boundary errors. Experiments on multiple medical imaging datasets show that PU-Diff achieves superior segmentation precision and structural fidelity compared with state-of-the-art semi-supervised methods, successfully maximizing the potential value of difficult unlabeled samples rather than passively discarding them. PU-Diff is architecturally adaptable and can be further optimized in computational complexity, facilitating its application to other complex anatomical structures, diverse imaging modalities, and clinical segmentation tasks under strict resource-constrained environments.
Available online , doi: 10.11999/JEIT260968
Abstract:
Objective The rapid evolution of modern communication and sensing technologies, such as 6G communication-sensing integration, 5G ultra-dense networking, and indoor Internet of Things (IoT) applications, has rendered the operating environment of array antenna systems increasingly complex, and Mixed-Near-and-Far (MNF) field scenarios with the coexistence of far-field and near-field targets have become the norm. As the core technology of array signal processing, beamforming includes adaptive beamforming and static beamforming. Unlike static beamforming, adaptive beamforming can automatically adjust the weight vector according to changes in the actual environment, allowing the desired signal to pass without distortion while suppressing interference and noise. Existing adaptive beamforming algorithms tailored for either single far-field or single near-field scenarios are limited to signal processing in a single-distance scenario and suffer from performance limitations when addressing the filtering problem in MNF field scenarios. Therefore, it is imperative to carry out research on adaptive beamforming algorithms for the MNF field. Current studies on the MNF field mainly focus on source localization or communication precoding, while research on MNF-field beamforming for interference suppression remains insufficient. Moreover, the source position parameters are generally unknown and coupled, i.e., the response vector depends simultaneously on the angle and the distance, which causes traditional angle-only methods to suffer from significant model mismatch when processing near-field sources. Aiming at the degradation of beamforming performance caused by this coupling problem, this paper proposes a Robust Adaptive Beamforming Using Eigenspace Bases Transition in MNF field. Methods Consider a uniform linear array (ULA) composed of M isotropic sensor elements with a fixed spacing, which receives one desired signal, J interferences, and additive white Gaussian noise in the MNF field; the Fresnel first-order approximation is adopted for near-field sources. The algorithm reconstructs the interference-plus-noise covariance matrix (INCM) by estimating the eigenspace of the received data and directly removing the desired signal component from the Sample Covariance Matrix (SCM). Unlike the reconstruction algorithms based on subspace basis transition, the reconstruction algorithm based on eigenspace bases transition can avoid the direct estimation of the noise covariance matrix and improve the accuracy of INCM reconstruction. To tackle the issue of coupled source position parameters within the MNF field, a multi-start L-BFGS-B optimization strategy is employed to solve the space distance minimization problem, resulting in an estimated source response vector set (SRVS). Specifically, multiple initial points are randomly sampled within the search space bounded by nominal DOA prior information, and each point independently undergoes L-BFGS-B optimization to convergence; the solution with the minimum objective value among all runs is then selected as the parameter estimation result. This multi-start mechanism effectively overcomes the local minimum issue caused by permutation symmetry. Additionally, an eigenspace bases transition (EBT) method is proposed to compute the reconstructed INCM, thereby circumventing the need for power estimation of interference and noise. With the basis transition matrix, the reconstructed INCM is obtained by attenuating the desired-signal component and nulling the noise-subspace components through a diagonal attenuation matrix, and the noise-subspace term is then added back to compensate for the missing noise components. In typical scenarios with large arrays and controlled optimization parameters, the computational complexity of the proposed algorithm is on the same order as that of classical adaptive beamforming methods. Subsequently, the modified steering vector (SV) is derived through eigendecomposition of the desired signal covariance matrix, and the weighting vector for the MNF field beamformer is produced. Results and Discussions A series of simulation experiments are performed to evaluate the performance of the proposed EBT-based adaptive beamforming in the MNF field. In the simulations, an 11-element ULA with half-wavelength spacing is adopted, and the desired signal and two interferences are located at (15°, 30 m), (−25°, 1000 m), and (35°, 15 m), respectively, with an interference-to-noise ratio of 20 dB and 200 Monte Carlo trials. These experiments include the beamformer performance, the impact of L-BFGS-B parameter selection (number of starting points and maximum iterations) on the algorithm performance, and the impact of the attenuation coefficient on it. The results show that the proposed algorithm maintains stable mainlobe gain in the direction of the desired signal, while forming deep nulls in the direction of interference, which greatly suppresses interference (Fig.1 ). Under MNF field conditions, the existing near-field and far-field algorithms are mismatched with this scenario and suffer from significant performance degradation when processing the array-received data. In contrast, the proposed algorithm can effectively solve the problem of beamformer performance degradation caused by the aforementioned coupling situation. Meanwhile, under the condition of high input signal-to-noise ratio (SNR), the difference between the output signal-to-interference-plus-noise ratio (SINR) of the proposed algorithm and the theoretical optimal value is less than 1 dB (Fig.2 , Fig.3 ). Moreover, the output SINR value of the proposed algorithm is much higher than that of other algorithms under different snapshot numbers and approaches the optimal upper bound. In the simulation analysis of L-BFGS-B parameter and attenuation coefficient selection, the proposed algorithm can ensure that the output SINR value remains relatively stable under reasonable parameter selection and close to the theoretical optimal value (Fig.4 , Fig.5 ). Overall, the proposed algorithm can effectively solve the problem of beamformer performance degradation caused by the aforementioned coupling situation and outperforms the compared beamforming algorithms. Conclusions This paper addresses the performance degradation of beamformers in MNF field scenarios caused by unknown and coupled source position parameters. A robust adaptive beamforming algorithm using EBT is proposed. The method first estimates the source response vector set via the multi-start L-BFGS-B optimization strategy to handle parameter coupling. It then reconstructs the interference-plus-noise covariance matrix through the EBT method, eliminating the desired signal component and avoiding direct noise power estimation. Finally, the corrected steering vector and the optimal weight vector are derived. Simulation results demonstrate that the proposed algorithm effectively maintains mainlobe gain towards the desired signal while forming deep nulls at interference directions, significantly suppressing interference. The proposed algorithm achieves an output SINR close to the theoretical optimum, especially under high input SNR, and outperforms the compared far-field or near-field algorithms. The performance remains robust across different snapshot numbers and parameter selections. Therefore, the EBT-based algorithm provides an effective solution for robust adaptive beamforming in complex MNF field environments, enhancing system reliability in modern communication and sensing applications.
Available online , doi: 10.11999/JEIT261074
Abstract:
Objective Lightning nowcasting is essential for severe convective weather warning and the safety of aviation and power systems. However, existing methods face three major challenges. First, conventional fixed-window sampling tends to generate numerous non-lightning or weak-lightning background samples and may repeatedly divide or truncate the same convective system, resulting in incomplete representations of thunderstorm structures and temporal evolution. Second, substantial differences in sparsity, response intensity, and spatial distribution among lightning, radar, and satellite observations may cause strong-response modalities to obscure sparse lightning features during direct fusion. Third, multi-source observations may be delayed, missing, or abnormal in operational environments, leading to input distribution shifts and degraded forecasting performance. To address these problems, this paper proposes a three-dimensional Non-Maximum Suppression (3D-NMS) method for spatio-temporal sample construction and a Multi-source Grouping SimVP model, termed MG-SimVP for lightning nowcasting with missing multimodal information. Methods MG-SimVP is developed within the encoder–translator–decoder framework of SimVP. The model uses VLF-LLN lightning observations, weather radar reflectivity, and Himawari-8 satellite features as inputs. All observations are remapped to a 256×256 grid and temporally aligned at 10-minute intervals. Six consecutive historical frames are used to predict the lightning density distributions of the following six frames, corresponding to a 1-hour forecast window. construct spatio-temporal samples, the original lightning data are first organized into a 1024 ×1024 ×T spatio-temporal tensor. Candidate samples are extracted using a 256×256×12 Gaussian sliding window. Three-dimensional convolution is then employed to calculate the lightning density concentration response of each candidate. Background filtering removes samples with weak lightning activity or displaced convective centers, while non-maximum suppression eliminates highly overlapping candidates and retains samples with stronger responses. The proposed Multi-source Grouped Adaptive Fusion Encoder (MGAF-Encoder) separately encodes lightning, radar, and satellite features. Global average pooling and modality-specific Multi-Layer Perceptrons (MLPs) are used to estimate adaptive channel weights, after which the aligned modality features are dynamically fused. The shared fusion feature is fed back to each branch to enhance cross-modal information exchange. When a modality is missing, it is replaced with a zero-valued tensor. MGAF-Encoder adjusts the fusion weights according to changes in the global channel responses, reducing the contribution of the missing modality and strengthening the remaining valid observations. The fused features are further processed by Mid_Xnet for multi-scale spatio-temporal evolution modeling and decoded to generate future lightning density fields. Results and Discussions Compared with the unscreened sampling strategy, 3D-NMS increases the average CSI of the multi-source SimVP model from 0.5323 to 0.6014 , corresponding to a relative improvement of 13.0%. Meanwhile, the average FAR decreases from 0.4155 to 0.3317 , a relative reduction of 20.2%. These results indicate that 3D-NMS effectively removes low-value background and redundant samples, concentrates valid lightning information, and preserves the principal thunderstorm structures and their continuous evolution. Under complete multi-source inputs, MG-SimVP outperforms the standard SimVP model at all lead times from 10 to 60 minutes. The average CSI increases from 0.6014 to 0.6701 , representing an improvement of 11.4%, while the average FAR decreases from 0.3317 to 0.2354 , representing a reduction of 29.0%. At the 60-minute lead time, MG-SimVP maintains a CSI of 0.5570 , compared with 0.4594 for the standard SimVP. This demonstrates that grouped encoding and adaptive fusion can reduce the masking effect of strong radar and satellite responses on sparse lightning features. Incomplete-input experiments further demonstrate the robustness of MG-SimVP. When satellite observations are missing, the model retains 88.24% of the complete-input CSI performance. When radar observations are missing, the retention rate is 84.10%. Even when only historical lightning observations are available, the model retains 79.49% of the complete-input performance. The model therefore avoids severe forecasting failure under missing-data conditions and maintains the identification of the principal lightning activity regions. Conclusions This paper proposes a 3D-NMS-based spatio-temporal sample construction method and a multi-source lightning nowcasting model, MG-SimVP, for missing multimodal information. The 3D-NMS method improves the concentration of valid lightning information and preserves complete convective structures and continuous evolution while removing low-density and highly redundant samples. MGAF-Encoder suppresses the masking effect of strong-response modalities on sparse lightning features through grouped independent encoding, global channel response perception, adaptive weighted fusion, and shared-feature feedback. Under missing observations, the encoder dynamically adjusts modality contributions according to changes in channel responses, thereby reducing the influence of missing data and strengthening the contribution of the remaining valid observations. Experimental results show that MG-SimVP achieves an average CSI of 0.6701 within the 1-hour forecast window, improving the standard multi-source SimVP baseline by 11.4%, while reducing the average FAR by 29.0%. Under missing-input conditions, the model retains at least 79.49% of its complete-input CSI performance. The proposed method therefore exhibits satisfactory lightning-region identification accuracy, false-alarm control, robustness, and operational adaptability. Future work will incorporate richer dynamic and thermodynamic variables, including three-dimensional radar echoes and numerical weather prediction fields, and will further optimize the sample construction strategy and loss function to improve forecasting performance for sudden, small-scale, and locally high-density lightning events.
Available online , doi: 10.11999/JEIT260616
Abstract:
Objective Small sea-surface targets have low Radar Cross Sections (RCSs), and their echoes are easily masked by intense sea clutter, resulting in extremely low Signal-to-Clutter Ratios (SCRs). Single-radar systems are constrained by a single operating frequency band and fixed observation angles, which limits their detection performance. Multi-radar collaborative detection provides an effective approach to overcoming this limitation. Existing multi-radar fusion detection methods are developed for different information fusion levels. However, feature-level and signal-level fusion methods generally assume identical clutter distributions, whereas decision-level fusion is less dependent on the specific clutter distribution. In practical multi-radar detection, differences in radar parameters, including frequency band, range resolution, and grazing angle, result in non-identical statistical characteristics of sea clutter. Furthermore, target RCS varies with observation azimuth and operating frequency, resulting in different SCRs for the same target observed by different radars and causing model mismatch in conventional fusion detectors. To address the coexistence of non-identical clutter distributions and different SCRs, a Neyman-Pearson (NP) criterion-based dual-radar Adaptive Fusion Detection method, termed NP-AFD, is proposed for Rayleigh and Weibull sea clutter. Methods Amplitude distribution fitting is performed on measured S-band and X-band sea clutter data using five commonly used models: Rayleigh, lognormal, Weibull, Gamma, and K distributions. Fitting accuracy is evaluated using the Mean Square Error (MSE) of the Probability Density Function (PDF) and Complementary Cumulative Distribution Function (CCDF), as listed in Table 1 . Based on the fitting results, local optimal test statistics are derived for Rayleigh and Weibull sea clutter. Because the two radar observations are independent, the joint likelihood ratio is given by the product of their individual likelihood ratios. The fusion test statistic is formulated as a weighted sum of the two local test statistics. The fusion weights are adaptively determined from the estimated SCRs, with larger weights assigned to radars with higher estimated SCRs. Closed-form analytical expressions relating the false alarm probability and detection probability to the decision threshold are also derived. Results and Discussions Monte Carlo simulations with 105 independent trials verify the derived closed-form expressions. Under identical SCRs for the two radars, the simulated detection curves of the individual radars and NP-AFD agree well with the theoretical curves (Fig. 3 ). Compared with decision-level OR and AND fusion, NP-AFD consistently achieves the highest detection probability, followed by OR fusion, whereas AND fusion performs worst. Performance gain analysis shows that NP-AFD maintains a positive performance gain over the better-performing single radar across all tested Weibull shape parameters. In contrast, OR fusion exhibits negative performance gain at low SCRs when the Weibull shape parameter is small, whereas AND fusion maintains a negative performance gain across the full SCR range (Fig. 4 ). The simulated false alarm probability is well controlled around the preset value of 10–3 (Fig. 5 ). Furthermore, a two-dimensional joint evaluation is performed by independently varying the SCRs of the two radars, and detection probability contours are plotted with the SCRs of the two radars as the coordinate axes (Fig. 6 ). The results show that NP-AFD requires lower SCR combinations than OR and AND fusion to achieve the same detection probability. Experiments with measured sea clutter data further demonstrate that NP-AFD achieves the highest detection probability among all compared methods (Figs. 7 and 8 ). Its detection performance agrees well with the theoretical results (Figs. 9 and 10 ). Conclusions The challenges posed by non-identical clutter distributions and different SCRs in multi-radar collaborative detection are addressed. Based on the Neyman-Pearson criterion, a dual-radar adaptive fusion detection method, NP-AFD, is derived for Rayleigh and Weibull sea clutter. Closed-form expressions are obtained for the fusion weights, decision threshold, and detection probability. Theoretical derivations, simulations, and experiments with measured sea clutter data consistently demonstrate that NP-AFD outperforms OR fusion and AND fusion under arbitrary SCR combinations, providing superior detection performance and robustness to different SCR combinations. The method, however, remains sensitive to non-uniform sea clutter, as reflected by degraded false alarm control in the presence of sea spikes.
Available online , doi: 10.11999/JEIT260582
Abstract:
Objective Infrared dim and small target detection remains challenging because targets are small, have low contrast, and are easily obscured by complex background clutter. Although existing methods improve detection performance, several limitations remain, including feature loss during downsampling, insufficient use of physical priors, and contamination of semantic features by background noise. These limitations can result in missed detections and false alarms. To address these problems, a detection network combining dynamic frequency guidance and semantic purification is proposed. Deep semantic features are dynamically coupled with adaptive frequency-domain priors to strengthen target and edge representations, improve detection accuracy, reduce false alarms, and enhance robustness in complex scenes. Methods A framework consisting of a shared encoder and a dual-branch decoder is constructed for infrared dim and small target detection (Fig. 1 ). First, a Local Target Enhancement (LTA) module is used to preprocess the original infrared image, strengthen small-target boundaries, and suppress background noise. The Sobel operator is then applied to extract edge information and provide a structural prior for subsequent feature learning. During feature extraction, a Learnable Downsampling (LD) module replaces conventional max pooling. Stride-2 convolution adaptively compresses the feature maps while preserving target details and reducing information loss. Residual blocks and progressive multiscale feature fusion are also used to strengthen semantic representations at different levels. A Dynamic Frequency Guidance Module (DFGM) is then introduced to enhance high-frequency details and target-edge information. Deep semantic features are used to dynamically predict the cutoff radius of a high-pass filter. A content-adaptive high-pass filter mask is constructed in the frequency domain to suppress low-frequency components. The filtered features are subsequently transformed back to the spatial domain to generate a frequency guidance map. To suppress spurious background responses introduced during frequency-domain enhancement, a Semantic Purification Module (SPM) is further designed. The deepest semantic features are used to generate a spatial attention map for top-down unidirectional filtering of the frequency guidance features. Background interference is thereby suppressed without contaminating high-level semantic representations with low-level noise. Finally, the dual-branch decoder reconstructs the target region and edge structure, thereby improving segmentation consistency, target boundary localization, and target-shape recovery. Results and Discussions Experimental results on the SIRST-Aug and IRSTD-1k datasets show that the proposed method outperforms mainstream comparison methods. On SIRST-Aug (Table 1 ), the method achieves the best Iintersection over Union (IoU), normalized Intersection over Union (nIoU), area under the receiver operating characteristic curve (AUC), and probability of detection (Pd), reaching 76.66%, 73.20%, 94.45%, and 99.17%, respectively. The false-alarm rate (Fa) is maintained at 33.96 × 10–6. On IRSTD-1k (Table 1 ), the method achieves the best IoU, AUC, and Fa values of 67.17%, 89.52%, and 10.62 × 10–6, respectively, while obtaining an nIoU of 67.67% and a Pd of 92.59%. Qualitative comparisons further show that the proposed method provides more accurate target localization and more complete target-shape segmentation under complex background interference (Figs. 4 and 5 ). The model contains 2.94 M parameters, requires 14.68 G FLoating-point OPerations (FLOPs), and has an average inference time of 7.54 ms, providing a favorable balance between detection performance and computational cost (Table 2 ). Ablation experiments further confirm the contributions of the key modules (Table 3 ). Adding DFGM increases AUC from 93.17% to 94.60%. Introducing LD increases nIoU to 73.32%, and adding SPM further increases IoU to 76.66%. The loss-function ablation results also show that the combined mask and edge losses provide a better balance among segmentation accuracy, target boundary localization, and false-alarm suppression (Table 4 ). Overall, the proposed method provides strong detection performance and effective false-alarm suppression. Conclusions An infrared dim and small target detection network based on dynamic frequency guidance and semantic purification is proposed. Through the coordinated use of LD, DFGM, and SPM, target details are effectively preserved, high-frequency structural information is strengthened, and semantic representations are purified. Background interference is consequently suppressed, and target-boundary recovery is improved. Experiments on SIRST-Aug and IRSTD-1k demonstrate that the proposed method outperforms existing mainstream methods across multiple evaluation metrics and maintains strong robustness in complex scenes. Future work can incorporate richer physical priors and self-supervised learning strategies to further improve model generalization and robustness under challenging conditions.
Available online , doi: 10.11999/JEIT260847
Abstract:
Objective Infrared dim and small target detection exploits thermal-radiation differences between targets and backgrounds for passive sensing in air-ground inspection, maritime surveillance, and wide-area warning. Under long-range imaging, low signal-to-noise ratios, and complex backgrounds, targets occupy few pixels and exhibit weak textures, blurred contours, and low contrast. Bounding-box detection directly predicts class confidence and target locations, making it suitable for rapid edge deployment and subsequent tracking initialization. The Real-Time DEtection TRansformer (RT-DETR) provides an efficient end-to-end framework. However, its application to infrared dim and small targets has three limitations. Repeated downsampling weakens local details and initial localization cues. Intra-scale global interaction does not sufficiently distinguish high-frequency target responses from low-frequency background context. In addition, cross-scale fusion may propagate homogeneous clutter because of insufficient explicit spatial constraints. Therefore, a Frequency-domain decoupling and Spatial-Prior-constrained DETR (FSP-DETR) is proposed to coordinate shallow detail preservation, target-background decoupling, and localization constraints. Methods FSP-DETR is built on RT-DETR-R18 and follows a pipeline of backbone feature extraction, intra-scale interaction, cross-scale fusion, and query-based decoding (Fig. 1 ). Three modules are introduced at successive stages to address the identified limitations. First, a Subspace Progressive Attention Modulated Inverted Residual Block (SPA-MIRB) is embedded in the Residual Network 18 (ResNet18) backbone to preserve shallow details (Fig. 2 ). Its inverted-residual branch combines pointwise and depthwise convolutions with a residual connection to retain target edges, spots, and weak textures. Its attention branch partitions the channels into multiple subspaces and progressively transfers interaction results across them to strengthen weak target responses. Second, a Frequency-Domain asymmetrically decoupled Attention-based Intra-scale Feature Interaction module (FD-AIFI) replaces the Attention-based Intra-scale Feature Interaction module (AIFI) (Fig. 3 ). The original unified global interaction is divided into a high-pass local attention path and a low-pass global attention path. The high-pass path applies attention within local windows to preserve compact target responses and limit interference from distant clutter. The low-pass path retains the original spatial resolution of the query features but pools the key and value features. This design compresses low-frequency context while preserving spatial queries and target-position indices. The outputs of the two paths are concatenated along the channel dimension and then linearly projected. Third, a Spatial-Prior-Constrained CNN-based Crossscale Feature-fusion Module (SPC-CCFM) retains the original upsampling, concatenation, and convolution path and introduces a High-order Spatial Representation Module (HSRM) and an Overall-level Mapping Injection Mechanism (OMIM). HSRM aligns multilevel features and constructs spatial passthrough, high-order spatial relation, and low-order fidelity branches (Fig. 4 ). These branches preserve the original spatial responses, model spatially separated but response-similar clutter through hypergraph aggregation, and supplement local edge and weak-texture information, respectively. OMIM performs scale mapping and channel adaptation according to each fusion node and injects the resulting spatial prior into the corresponding node to constrain cross-scale feature refinement. Results and Discussions Experiments are performed on IRSTD-1k and NUAA-SIRST, with pixel-level masks converted into single-class minimum bounding rectangles. Component ablation shows that SPA-MIRB, FD-AIFI, and SPC-CCFM each improve Mean average precision at an Intersection over Union threshold of 0.5 (mAP@0.5) on both datasets (Table 1 ). FSP-DETR achieves mAP@0.5 values of 89.03% and 98.75% on IRSTD-1k and NUAA-SIRST, respectively, exceeding RT-DETR-R18 by 3.39 and 3.08 percentage points. The parameter count decreases from 19.87 million to 17.24 million, and the computational cost decreases from 56.9 to 50.4 billion floating-point operations, corresponding to reductions of 13.2% and 11.4%, respectively. Although some variants achieve higher precision or recall at a single operating point, the complete model obtains the highest mAP@0.5 with lower complexity. Replacement experiments further support the structural choices. SPA-MIRB and FD-AIFI achieve the highest mAP@0.5 among their respective alternatives. SPC-CCFM exceeds the two alternative fusion modules by 0.48 and 0.26 percentage points, respectively, indicating that its main benefit is stronger spatial constraint rather than complexity reduction (Table 2 ). SPA-MIRB reaches accuracy-complexity trade-off. A computation-allocation factor a = 0.50 yields the best performance, with mAP@0.5 and mAP@0.5:0.95 reaching 97.27% and 53.04%, respectively, on NUAA-SIRST (Table 3 ). Attention maps show that the high-pass path concentrates its responses around target neighborhoods, whereas the low-pass path produces broader and smoother responses that cover more surrounding background regions (Fig. 5 ). Under unified experimental settings, FSP-DETR achieves the highest mAP@0.5 on both datasets, with a frame rate of 172 frame/s and an average per-image forward time of 5.81 ms (Table 4 ). It also outperforms deeper RT-DETR variants with fewer parameters, lower computational cost, and shorter forward time. These results demonstrate the advantage of task-oriented structural adaptation over simply enlarging the backbone. Detection visualizations show better agreement between predicted bounding boxes and target regions in cluttered scenes (Fig. 6 ). Instance-level error analysis shows that the missed-detection rate decreases from 15.70% to 13.22% on IRSTD-1k and from 10.71% to 5.36% on NUAA-SIRST. The average number of false positives per image decreases from 0.490 to 0.380 and from 0.209 to 0.070, respectively (Table 5 ). Normalized center-offset errors also decrease on both datasets. However, the scale error on NUAA-SIRST increases slightly from 10.63% to 11.05%, indicating that scale regression for extremely small targets still requires improvement. Conclusions FSP-DETR coordinates shallow detail preservation, frequency-domain decoupled intra-scale interaction, and spatial-prior-constrained cross-scale fusion for end-to-end bounding-box detection. The model improves detection accuracy, reduces missed detections and false positives, and decreases normalized center-offset error. At the same time, model complexity is reduced and efficient inference is maintained. A favorable balance between detection accuracy and efficiency is therefore achieved in complex infrared scenes. Future work will investigate multi-frame spatiotemporal modeling for thermal crossover backgrounds and dense dim and small target scenes.
Available online , doi: 10.11999/JEIT260694
Abstract:
Objective HyperGraph Neural Networks (HGNNs) have received increasing attention because of their ability to model higher-order relationships among entities. However, their high computational cost and low inference efficiency limit their deployment in large-scale applications. Existing knowledge distillation methods that transfer knowledge from HGNNs to Multi-Layer Perceptrons (MLPs) have limited interpretability and accuracy. They also suffer from information loss in softmax-based soft labels and do not account for differences in node reliability. To address these limitations, a hypergraph knowledge distillation framework, Distill Hypergraph neural network to Kolmogorov-Arnold Network (DH2KAN), is proposed for fast inference. Methods DH2KAN consists of three main modules (Fig. 2 ). First, a Kolmogorov-Arnold Network (KAN) is adopted as the student model instead of an MLP. Learnable spline-based univariate functions replace fixed activation functions and linear weights to improve the fitting ability and interpretability of the student model. Second, a representation similarity distillation mechanism is introduced to directly align the pre-logit representations of the HGNN and KAN. This alignment reduces the information loss caused by softmax normalization and preserves higher-order structural knowledge in the hypergraph. Third, a high-reliability node-aware distillation method is introduced (Fig. 3 ). Node reliability is quantified from changes in information entropy under noise perturbations. High-reliability nodes that are less sensitive to noise are selected, and their soft labels are used as additional supervision to improve the quality of distilled knowledge. Results and Discussions DH2KAN is evaluated under both transductive learning (Table 2 ) and production learning (Table 3 ) settings. Under transductive learning, its average accuracy is 10.53 percentage points higher than that of the baseline KAN student model, 1.73 percentage points higher than that of the teacher HGNN, and 1.20 percentage points higher than that of the MLP-based distillation baseline. These results confirm the effectiveness of knowledge transfer from the HGNN to the KAN and show that the proposed method achieves higher inference accuracy than conventional MLP-based distillation methods. In addition, DH2KAN is particularly well suited to feature-dominated hypergraphs and remains robust on hypergraphs with sparse structures and noisy nodes. Conclusions DH2KAN is proposed to accelerate HGNN inference in large-scale, low-latency applications. Knowledge distillation narrows the performance gap between the KAN and HGNN and removes dependence on the hypergraph structure during inference, thereby enabling fast, low-complexity prediction. Representation similarity distillation and high-reliability node-aware distillation use the teacher model’s pre-logit representations and the soft labels of high-reliability nodes as additional supervision, respectively. These mechanisms enable effective transfer of task-relevant knowledge to the student model and support the practical deployment of DH2KAN in large-scale, low-latency scenarios.
Available online , doi: 10.11999/JEIT260702
Abstract:
Objective Ground unmanned platforms operating in urban streets, industrial parks, and underground passages require reliable short-burst command-and-control communication. In these complex near-ground environments, reception is affected by strong multipath fading, large Carrier Frequency Offset (CFO), residual Timing Offset (TO), Sampling Frequency Offset (SFO), and in-band interference from coexisting wireless systems. Conventional Chirp Spread Spectrum (CSS) receivers based on single-peak decisions in the dechirp-Discrete Fourier Transform (DFT) domain are vulnerable to multipath-induced spectral splitting and interference-induced bin masking. Direct-Sequence Spread Spectrum (DSSS) receivers can also be sensitive to synchronization errors under combined offsets and show reduced energy efficiency under multipath. Robust handling of these impairments is therefore required for low-latency control links and for the coexistence of unmanned ground platforms with legacy wireless infrastructure in dense deployments. Methods A Preamble-Referenced Cyclic Cross-correlation CSS (PRCC-CSS) scheme is proposed in which frame structure, synchronization estimation, and payload detection are jointly designed. Each frame consists of multiple identical up-chirp preamble symbols, a down-chirp Start-of-Frame Delimiter (SFD), and CSS-modulated payload symbols. Integer CFO and TO are jointly estimated from closed-form linear combinations of the complementary frequency-domain indices at the dechirp-DFT outputs of the up-chirp preamble and down-chirp SFD. Fractional CFO is estimated from inter-symbol phase differences between adjacent preamble symbols, whereas fractional TO is estimated by interpolation around the main peak of the averaged preamble spectrum. Under a common oscillator-reference assumption, SFO-induced bin drift is compensated using the clock-offset relation derived from CFO. After offset compensation, the averaged preamble spectrum is used to construct a frame-specific reference spectrum that represents the current multipath frequency-domain structure. Each payload symbol is then detected by cyclically cross-correlating its spectrum with the reference spectrum over candidate cyclic shifts. The shift corresponding to the maximum cyclic cross-correlation is selected as the demodulated symbol index. In this way, the multipath-induced frequency-domain structure is converted from an adverse perturbation into a matchable intra-frame reference feature. Multipath-robust structure-matched payload detection is therefore achieved without explicit path-by-path channel estimation. Results and Discussions PRCC-CSS is evaluated under identical bandwidth, sampling rate, and processing gain at a target Bit Error Rate (BER) of 10–4. First, it is compared with three DSSS baselines using Binary Phase-Shift Keying (BPSK), Quadrature Phase-Shift Keying (QPSK), and 16-ary Quadrature Amplitude Modulation (16QAM) under three channel conditions. Under Additive White Gaussian Noise (AWGN, Fig. 1), PRCC-CSS reaches the target BER at approximately 4.9 dB Eb/N0, compared with 8.5～9.0 dB for the best DSSS baseline. This result indicates that chirp index modulation combined with preamble-referenced correlation provides frequency-domain energy aggregation independent of a specific multipath profile. Under the Extended Typical Urban (ETU) channel without interference (Fig. 2 ), PRCC-CSS requires approximately 6.5 dB Eb/N0, compared with approximately 10.0 dB for the best DSSS baseline. The reference spectrum captures and reuses the per-frame multipath frequency-domain structure, whereas finite-finger DSSS-RAKE is more strongly affected by path-capture and code-synchronization errors. Under ETU with in-band interference at a Jamming-to-Signal Ratio (JSR) of 5 dB (Fig. 3 ), PRCC-CSS requires approximately 7.8～8.0 dB Eb/N0, compared with 10.8～11.0 dB for the best DSSS baseline. The reference and payload spectra experience nearly the same multipath frequency-domain structure within a frame, which helps preserve decision separability during cyclic cross-correlation. By contrast, DSSS-RAKE accumulates residual interference across its combining fingers. Second, PRCC-CSS is compared with Dechirp Non-Coherent (DNC) and coherent peak detection at Spreading Factor (SF)=7 and SF=9 under ETU with interference (Fig. 4 ). At SF=7, DNC does not reach the target BER within the tested Signal-to-Interference-plus-Noise Ratio (SINR) range. At SF=9, DNC requires approximately 4 dB SINR. PRCC-CSS reaches the same target at approximately –5 dB SINR at SF=7 and –11 dB at SF=9, corresponding to an SINR threshold improvement of approximately 9～15 dB. These results show that the gain is mainly attributable to frame-wide cyclic matching rather than phase compensation alone. Third, a Software-Defined Radio (SDR) prototype is evaluated on an ETU-emulated channel at JSR = 5 dB (Figs. 5 and 6 ). Of the 332 received frames, 300 are retained as reliable, corresponding to a frame retention rate of 90.4%. Among the retained frames, no symbol errors are observed in 6 000 payload symbols, and no bit errors are observed in 54 000 payload bits. The retained-frame conditional BER therefore has a one-sided 95% upper confidence bound of 5.56 × 10-5. Conclusions The proposed PRCC-CSS scheme jointly integrates integer and fractional CFO and TO estimation, SFO estimation, and cyclic cross-correlation payload detection. At BER=10–4, PRCC-CSS reduces the required Eb/N0 by approximately 2.8～4.1 dB relative to the best DSSS baseline using BPSK, QPSK, or 16QAM across AWGN, ETU, and ETU with JSR=5 dB. Under ETU with interference, PRCC-CSS reduces the required SINR by approximately 9～15 dB relative to DNC at SF=7 and SF=9 and maintains a further consistent margin over coherent peak detection. In the SDR prototype, 90.4% of the received frames are retained, and the retained-frame conditional BER has a one-sided 95% upper confidence bound of 5.56×10–5. By using the multipath-induced frequency-domain structure as a matchable intra-frame reference feature rather than as a perturbation, PRCC-CSS provides a candidate physical-layer solution for reliable short-burst communication in complex near-ground environments. Future work will extend the scheme to higher-order CSS modulation, multi-antenna diversity reception, and adaptive reference-spectrum updating for time-varying channels.
Available online , doi: 10.11999/JEIT260452
Abstract:
Spacecraft Six-Dimensional (6D) pose estimation aims to determine the relative pose of a target spacecraft with respect to a servicing spacecraft in a spatial coordinate system. It is a key step in close-range operations, including defunct satellite removal, space debris capture, on-orbit spacecraft manipulation, and space station rendezvous and docking. In recent years, Convolutional Neural Network (CNN)-based 6D pose estimation methods have received considerable attention. However, their strong dependence on convolutional architectures makes them sensitive to image textures and limits their ability to model long-range contextual information effectively. Moreover, mainstream methods typically adopt a pipeline consisting of object detection followed by pose estimation. Such methods provide limited diversity in feature extraction and are sensitive to low-light conditions and complex background interference. To address these problems, a spacecraft pose estimation network based on multi-frequency feature interaction and dynamic fusion is proposed. Backbone networks with different receptive fields are first used to separately extract spatial high-frequency features, such as semantic and edge details, and spatial low-frequency features, such as global structural information. A Transformer-based feature interaction mechanism is then used to perform self-attention and cross-attention between the two types of features, thereby aggregating long-range contextual information. To further exploit rich frequency-domain representations, a frequency-guided feature module is introduced to dynamically fuse multi-frequency features. Extensive experiments on spacecraft pose estimation datasets demonstrate that the proposed method achieves competitive performance and strong generalization capability compared with existing methods. Objective The local receptive fields of CNN operators limit their ability to capture long-range dependencies. By contrast, Transformers with multi-head self-attention are more effective in globally modeling long-range contextual information and are well suited to multiscale targets and small spacecraft at long distances. Therefore, improving the extraction and interaction of spatial high- and low-frequency features is essential for increasing the accuracy of spacecraft pose estimation. Methods A 6D pose estimation network based on Transformer-based multi-frequency feature interaction and dynamic fusion is proposed. The framework consists of two branches. A ResNet-18 branch is used to extract spatial high-frequency features, including semantic and edge information, which are also used for spacecraft semantic segmentation. A DarkNet-53 branch is used to extract spatial low-frequency features that emphasize global structural information. A Transformer-based feature interaction module is then designed to model interactions between the high- and low-frequency features through self-attention and cross-attention. This mechanism aggregates long-range contextual information, increases the diversity of feature representations for texture-poor spacecraft images, and facilitates network optimization. A frequency-guided dynamic feature fusion module is further introduced to integrate the multi-frequency features. Results and Discussions The proposed method estimates spacecraft 6D pose through multi-frequency feature interaction and dynamic fusion. On the SwissCube dataset, it achieves an overall ADI-0.1d accuracy of 82.31%, exceeding those of CA-SpaceNet (79.39%) and WDR* (78.78%). On the SPEED dataset, the combined error eq + et is 0.0299, which is 22.3% lower than that of CA-SpaceNet (0.0385) and 25.2% lower than that of WDR* (0.0400). The qualitative results in Fig. 5 and Fig. 7 show that the predicted keypoints are closer to the ground truth, particularly for challenging images dominated by low-frequency content and affected by complex backgrounds. Ablation experiments further show that the feature interaction module alone increases the overall accuracy from 79.39% to 80.67%, whereas the addition of frequency-guided dynamic fusion further increases it to 82.31%. These results demonstrate that the proposed framework improves the robustness of spacecraft pose estimation and provides more reliable pose predictions for on-orbit servicing and spacecraft rendezvous and docking. Conclusions To improve the accuracy of spacecraft 6D pose estimation under challenging space imaging conditions, a multi-frequency feature interaction mechanism and a semantic segmentation branch using semantic and edge information are introduced. Long-range contextual information from spatial high- and low-frequency features is aggregated to improve feature representation. In addition, a frequency-guided dynamic feature fusion module is introduced to exploit rich frequency-domain representations. Comprehensive comparisons with mainstream methods on the SwissCube and SPEED datasets demonstrate that the proposed method improves feature representation and increases the robustness of spacecraft pose estimation under complex interference.
Available online , doi: 10.11999/JEIT260727
Abstract:
Objective Space-Air-Ground Integrated Networks (SAGINs) combine satellites, aerial platforms, terrestrial access points, and edge nodes to support heterogeneous services in highly dynamic environments. Collaborative decision-making is challenging because observations are local and delayed, network topology varies over time, raw-data exchange is restricted, and spectrum, power, and computing resources are coupled. Existing Multi-Agent Reinforcement Learning (MARL) and game-theoretic methods do not jointly address uncertainty-aware semantic consensus, hierarchical resource allocation, cross-domain adaptation, and lightweight deployment. This study develops a Collective Active Inference (CAI)-based integrated intelligence generation and transmission loop. Agents exchange inferred beliefs rather than raw observations, consensus beliefs guide resource decisions, and learned models are adapted and compressed for resource-constrained satellite and edge nodes. The method targets time-varying, partially observable SAGINs and is evaluated through an Open Radio Access Network (O-RAN) system-level simulation rather than an operational deployment. Methods The problem is formulated as a time-varying decentralized partially observable Markov decision process over channel, queue, congestion, interference, and node states. This formulation captures decentralized decisions under incomplete information and intermittent connectivity. Multi-agent cognition is modeled by minimizing Collective Variational Free Energy (CVFE). A likelihood term preserves local observation fidelity, while a Kullback-Leibler divergence term penalizes disagreement between neighboring beliefs. A Gaussian precision-weighted belief update assigns greater weight to more reliable beliefs. Symmetric Metropolis weights produce a doubly stochastic mixing matrix. For a fixed connected graph, the consensus error decreases geometrically according to the second-largest eigenvalue modulus. This result extends to intermittent links under B-joint connectivity. Agents exchange only belief means and covariance matrices and act on converged beliefs (Algorithm 1). A vertical-horizontal hierarchical game maps consensus beliefs to resource allocations. Vertically, the operator acts as the Stackelberg leader, and service slices respond to prices. Horizontally, the slices reach a Generalized Nash Equilibrium (GNE) under shared spectrum, power, and Service Level Agreement (SLA) constraints. Cross-domain adaptation combines Differentially Private Federated Learning (DP-FL), top-10% gradient sparsification, Model-Agnostic Meta-Learning (MAML), and knowledge distillation. The distillation temperature is 4, and the supervised and distillation loss weights are 0.3 and 0.7, respectively. Evaluation is conducted using an OMNeT++ O-RAN simulator with 12 coverage nodes, 19 underlay nodes, 72 links, a 28 GHz carrier, and 100 MHz bandwidth. Concurrent Ultra-Reliable and Low-Latency Communications (URLLC), enhanced Mobile BroadBand (eMBB), and massive Machine-Type Communications (mMTC) traffic is simulated for 10 000 control periods. The network load factor increases from 0.2 to 1.0. Frequency-sweeping interference and traffic bursts are introduced to test Out-Of-Distribution (OOD) recovery. All baselines use the same observations, action space, constraints, training budget, and random-seed set. Reported values are averaged over at least 10 independent runs with 95% confidence intervals. Results and Discussions At full load, the proposed method achieves a throughput of 93.2 Mb/s, compared with 81.8, 88.5, and 78.4 Mb/s for Symmetry-Informed MARL (SI-MARL), Quantum MARL (QMARL), and Graph Attention Network-based MARL (GAT-MARL), respectively. Its task-scheduling success rate is 0.92. The 95th-percentile latency is 20.5 ms, compared with 31.2, 26.4, and 33.0 ms for the three baselines, and the SLA violation rate is approximately 0.081 (Fig. 3 ). After an OOD disturbance, the Key Performance Indicator (KPI) retention rate is approximately 0.83, and recovery to the 95% threshold requires 23 control periods, compared with 35 and 41 periods for QMARL and GAT-MARL, respectively (Fig. 4 ). In the resource-competition scenario, the mean task-scheduling success rate is 0.92, with a 95% confidence interval of [0.90, 0.94]. The corresponding values for three hierarchical-game or federated-game baselines range from 0.80 to 0.85 (Fig. 5). With 50 agents and an update dimension of 80, top-10% gradient sparsification reduces the number of transmitted scalars from 321.79 × 103 to 34.58 × 103, an 89.3% reduction. The number of convergence rounds increases from 58.84 to 71.26, or by 21.1% (Table 1 ). For few-shot relation inference, the F1 score is 0.71 with 20 labeled samples and 0.98 with 80 labeled samples. Approximately 48 labeled samples are required to reach an F1 score of 0.90, compared with 75, 90, and 120 samples for the baselines (Fig. 6 ). Knowledge distillation reduces the model size from 1 050 KB to 160 KB and the inference latency from 21 ms to 15 ms. The student model retains a task success rate of 91% and a path efficiency of 93%, compared with 95% and 97%, respectively, for the teacher model. Removing environment-domain information causes the largest loss in path efficiency (Table 2 ). Conclusions The proposed CAI-based method forms a closed loop that combines precision-weighted semantic consensus, hierarchical resource coordination, cross-domain adaptation, and lightweight deployment. The convergence analysis relates consensus speed to graph connectivity. Under the stated simulation settings, the O-RAN evaluation shows higher throughput and task-scheduling success rates, lower 95th-percentile latency and SLA violation rates, and faster disturbance recovery. Top-10% gradient sparsification reduces communication overhead, whereas knowledge distillation reduces model size and inference latency. The current evaluation is limited by simulated channels, idealized observation quality, synchronous updates, and modeled prior distributions that may differ from operational conditions. These results therefore do not guarantee performance across arbitrary operational networks. Further validation using measured channels, hardware-in-the-loop tests, and field experiments is required.
Available online , doi: 10.11999/JEIT260482
Abstract:
Objective Marine Heterogeneous Wireless Networks (MHWNs) deeply integrate multidimensional resources across space, air, and sea. Multiple communication systems coexist in these networks, creating challenges in both interconnection and resource coordination. Existing routing algorithms based on Deep Reinforcement Learning (DRL) have limited capability to represent dynamic network topologies and therefore cannot make efficient routing decisions when the topology changes frequently. This limitation mainly arises because mainstream frameworks typically use standard Graph Neural Networks (GNNs) or fully connected networks for state encoding, which cannot effectively capture the structural features of highly dynamic topologies. Methods A modular Multi-Radio Access Technology (Multi-RAT) gateway supporting the fusion of 5G and mesh ad hoc networks is designed, and an intelligent routing method integrating Contrastive Message Passing for Graph Neural Networks (CMPGNN) with DRL, termed CMPGNN-DRL, is proposed. The method is designed to improve Quality of Service (QoS), forwarding efficiency, and resource allocation. In the gateway, service data undergo IP encapsulation, protocol identification, and semantic conversion through communication interface modules before being forwarded through the target interface. The gateway periodically collects node features and link states to construct the input graph for routing decisions. Key indicators, including link bandwidth utilization, queue depth, packet loss rate, and end-to-end delay, are monitored in real time. To compensate for information delays caused by periodic reporting, short-term trends in key indicators are incorporated into the state vector, and an asynchronous decision-execution architecture is adopted. Under a centralized Software-Defined Networking (SDN) control plane, the network is modeled as a graph with continuously monitored node and link features. For each node, CMPGNN synchronously constructs homophily and heterophily views through two message-passing paths. The resulting embeddings are constrained by a contrastive loss to obtain discriminative node representations that are robust to edge perturbations. These representations are then input into a Double Deep Q-network (DDQN) agent, which makes hop-by-hop routing decisions using an ε-greedy exploration strategy. For each neighboring node, CMPGNN predicts a topology prior value. This value is combined with the corresponding DDQN Q-value estimate by weighted summation to obtain a joint action value. The topology prior can correct inaccurate Q-value estimates when model training is insufficient or observations are noisy. A normalized multiobjective reward function assigns negative contributions to end-to-end delay, packet loss rate, and link load and a positive contribution to throughput, while explicitly penalizing routing loops. Results and Discussions The proposed solution is evaluated through hardware prototype measurements and extensive simulations. Prototype tests show average CPU utilization rates of 14%, 25%, and 37% under mesh-only, 5G-only, and dual-mode operation, respectively. The aggregate throughput under dual-mode operation reaches 108 Mbps, compared with 32 Mbps under mesh-only operation and 84 Mbps under 5G-only uplink transmission. The average ping delays between the gateway and the application server are 6 ms for mesh and 16 ms for 5G. CMPGNN-DRL is compared with six baseline methods, namely OSPF, AODV, GNN, DQN, MPNN-DQN, and DAR-DRL, on the GEANT2, GBN, Germany, and Synth50 topologies. The evaluation covers dynamic traffic, random link failures at rates of 3%～24%, and large-scale topology changes. The training reward increases rapidly and then stabilizes, and ablation experiments confirm the effectiveness of the contrastive learning mechanism. Compared with the best-performing baseline, CMPGNN-DRL reduces the average end-to-end delay by 20.8%～47.7% and the packet loss rate by 0.3%～5.3%, while increasing the average throughput by 5.2%～14.2%. Additional MHWN scenarios are configured according to the environmental constraints of the Maritime Internet of Things (MIoT). A 1 500 m × 1 500 m maritime area containing 50～100 randomly deployed nodes is simulated, and each configuration is evaluated through repeated Monte Carlo simulations. CMPGNN-DRL maintains stable performance as the network scale and node mobility vary and outperforms MPNN-DQN and DAR-DRL in terms of Packet Delivery Ratio (PDR) and packet loss rate. When the network load increases from 20% to 50%, CMPGNN-DRL improves PDR by 4.2%～17.3% over MPNN-DQN and DAR-DRL while maintaining lower end-to-end delay, higher bandwidth utilization, and a lower packet retransmission rate under medium-to-high network loads. Conclusions For sea-air cross-domain heterogeneous networks, a Multi-RAT fusion gateway supporting 5G and mesh ad hoc networks is designed, and a CMPGNN-DRL intelligent multipath routing method is proposed. The contrastive learning mechanism improves topology representation and enhances the robustness of routing policies. Experimental results show that CMPGNN-DRL outperforms existing mainstream algorithms in key performance indicators, including PDR, end-to-end delay, and throughput, and achieves good generalization across different network topologies. Future work will focus on validation in real maritime environments and improvement of training efficiency to support the practical deployment of integrated sea-air communication systems.
Available online , doi: 10.11999/JEIT260857
Abstract:
Objective Image steganography embeds secret information into a cover image for secure transmission and is an important technique for confidential communication and privacy protection. With the development of deep learning, learning-based steganography has substantially improved embedding capacity and reconstruction quality. However, high steganographic performance, strong resistance to steganalysis, and lightweight design remain difficult to achieve simultaneously. This trade-off has become a major obstacle to practical deployment. Single-domain methods have difficulty balancing these requirements, whereas existing high-performance models are often structurally complex and difficult to deploy in resource-constrained environments. Moreover, mainstream dual-domain schemes generally assume that embedding distortion follows a continuous distribution in the spatial and frequency domains. In practice, however, steganographic modifications tend to concentrate in discontinuous, complex-texture regions of the cover image. To address these problems, a lightweight image steganography network is designed to jointly optimize the embedding and extraction paths. The aim is to improve stego image quality and resistance to steganalysis while maintaining low computational cost and inference latency. Methods A lightweight steganographic network named GISNet is proposed. An encoder-decoder architecture with skip connections is adopted, and the hiding and extraction networks are structurally symmetric but do not share weights. First, Improved Bidimensional Empirical Mode Decomposition (IBEMD) is used to adaptively decompose the secret image into multiple intrinsic mode functions and a residual component. The secret information is thereby hierarchically dispersed among components with different morphological characteristics. Gaussian blur replaces extremum interpolation to estimate the local mean envelope, and the separability of the Gaussian kernel is used to reduce computational complexity. Fixed Gaussian parameters and ensemble averaging are further used to support deterministic and reversible reconstruction. Next, the Innovative Multispatial FreqBlock (IMFB) is designed by integrating multibranch dilated convolutions, an attention mechanism, dynamic gating, and frequency-domain processing. Spatial- and frequency-domain features are jointly processed to improve resistance to detection while maintaining high-fidelity extraction of secret information. Finally, the GraphFusion Neural Network (GFNN) is used as the bottleneck layer. A sparse graph is constructed in the feature space using the K-nearest-neighbor algorithm, and messages are propagated only among highly similar nonlocal nodes. Long-range pixel associations are therefore modeled explicitly, modification patterns in discontinuous texture regions are represented more effectively, and model complexity is substantially reduced. The hiding and extraction networks are jointly optimized using a four-term loss function consisting of the hiding loss, restriction loss, Laplacian pyramid loss, and perceptual loss. The terminology and module functions follow those defined in the main manuscript. Results and Discussions GISNet achieves the best overall hiding and recovery performance on DIV2K, COCO, and ImageNet, demonstrating high reconstruction quality and strong cross-dataset generalization (Table 1 ). On DIV2K, the Peak Signal-to-Noise Ratio (PSNR) reaches 58.07 dB for cover/stego image pairs and 60.10 dB for secret/recovered secret image pairs. The cover and stego images are visually indistinguishable, and the residual maps remain nearly black even after 30-fold magnification (Fig. 5 ). Ablation experiments show that replacing the Discrete Wavelet Transform (DWT) with IBEMD increases the PSNR by 8.04 dB and 5.57 dB for the two types of image pairs, respectively (Table 2 ). The complete combination of IBEMD, IMFB, and GFNN produces the best results. These findings confirm the complementary effects of hierarchical information dispersion, multiscale spatial-frequency mapping, and nonlocal feature association (Table 3 ). Multi-dilation-rate branches and joint spatial-frequency processing further improve hiding quality and recovery accuracy (Tables 4 and 5 ). Under four steganalysis methods, the detection accuracies range from 49.35% to 49.65%, indicating that the stego images are difficult to distinguish from natural cover images (Table 6 ). GISNet contains only 8.00 M parameters and requires 7.88 GFLOPs, with an inference time of 76 ms (Table 7 ). These results demonstrate that GISNet effectively balances visual quality, recovery accuracy, resistance to steganalysis, and computational efficiency. The manuscript reports the same comparative and computational results. Conclusions A lightweight dual-domain graph convolutional network for image steganography, named GISNet, is proposed. The experimental results support the following conclusions. (1) IBEMD hierarchically disperses secret information and supports deterministic and reversible reconstruction, providing a reliable basis for high-quality hiding and recovery. (2) IMFB and GFNN jointly improve stego image quality and resistance to steganalysis, enabling the stego images to resist detection by multiple steganalysis tools. (3) The lightweight design based on depthwise separable convolutions and sparse graph construction significantly reduces model complexity and computational cost, making GISNet suitable for deployment in resource-constrained environments. Future work will focus on robustness against channel interference and extensions to multi-image and cross-modal steganography to further improve the security and applicability of the proposed scheme. The lightweight mechanisms are consistent with the manuscript, which attributes the low computational cost to depthwise separable convolutions in IMFB and KNN-based sparse graph construction in GFNN.
Available online , doi: 10.11999/JEIT260636
Abstract:
Objective Traditional Auto-Regressive (AR) Kalman speech enhancement algorithms face three critical challenges under non-stationary noise: limited adaptability of noise model updates, model mismatch caused by burst noise, and a lack of environment-aware covariance adjustment. These limitations degrade enhancement performance and restrict their use in practical speech communication. An improved adaptive Kalman speech enhancement method is therefore proposed to improve speech quality and intelligibility under complex noise conditions. Methods First, an environmental deviation measure based on the Energy Entropy Ratio (EER) is constructed to quantify the statistical deviation between the current frame and the background environment. A dual-time-scale EER tracking mechanism is then established to capture instantaneous variations and stable background statistics. Their difference is used to generate an adaptive deviation intensity factor for the joint adjustment of the process-noise and observation-noise covariances. Second, a two-stage burst noise discrimination scheme is developed based on the characteristics of burst noise. Frame energy changes and the Spectral Flatness Measure (SFM) are jointly used for preliminary burst noise detection. A Speech Modeling Metric (SMM) based on the linear prediction residual variance ratio is further used to distinguish speech from weak-colored burst noise and prevent contamination of the speech AR model. Results and Discussions Experiments are conducted on the NOIZEUS database. The proposed method outperforms the conventional AR-based Kalman filter and the sensitivity-based Augmented Kalman Filter (AKF) in terms of Short-Time Objective Intelligibility (STOI), Perceptual Evaluation of Speech Quality (PESQ), and Segmental Signal-to-Noise Ratio (SegSNR). It shows improved adaptability and stability under non-stationary and burst noise conditions. Dual-time-scale environmental perception and burst noise discrimination facilitate rapid adaptation to environmental changes and reduce speech distortion. Conclusions The proposed burst noise suppression and dual-time-scale adaptive Kalman method addresses the limitations of conventional AR-based Kalman speech enhancement under complex noise conditions. EER-based tracking enables environment-aware covariance adjustment, whereas the two-stage discrimination scheme improves burst noise detection and reduces contamination of the speech AR model. The experimental results demonstrate the robust performance of the proposed method and support its application to speech enhancement under non-stationary noise.
Available online , doi: 10.11999/JEIT260843
Abstract:
Objective Low-light face images commonly suffer from insufficient illumination, reduced contrast, amplified noise, and loss of local details. These degradations reduce visual quality and destabilize identity recognition, fatigue analysis, and face behavior understanding. Structural and discriminative information around the eyes and mouth is especially vulnerable to edge blurring, texture loss, and local structural shifts. Existing low-light image enhancement methods mainly focus on illumination recovery or general image reconstruction, while preservation of key face structures and identity-related representations remains limited. From a frequency-domain perspective, low-light degradation weakens low-frequency structural responses while disturbing high-frequency edges and textures, and directly enhancing high-frequency components without structural guidance may amplify noise or introduce false textures. Therefore, low-frequency face structures should guide spatially corresponding high-frequency detail restoration while maintaining structural stability during cross-frequency interaction and multi-scale reconstruction. To address these problems, a structural-prior-guided cross-frequency collaborative method is proposed for low-light face image enhancement. Methods An encoder-decoder enhancement network is constructed with structural-prior-guided cross-frequency association (CFA), high-frequency modeling (HFM), and face structural consistency modulation (FSCM) (Fig. 1 ). In CFA, input features are decomposed into low-frequency and directional high-frequency subbands using a discrete wavelet transform. Key-region heatmaps generated from facial landmarks are combined with directional structural responses and local statistical information to construct a spatial structural prior, which guides the restoration of spatially corresponding high-frequency details according to key-region locations, structural directions, and local significance, thereby reducing excessive enhancement of weak-structure or noise-dominated regions. HFM combines depthwise separable convolution with two-dimensional selective scanning. Local convolution is used to extract edge and texture details, whereas multi-directional selective scanning models long-range spatial dependencies. FSCM further regulates structurally relevant features during cross-frequency interaction and cross-scale feature fusion (Fig. 2 ). Its cross-frequency branch selectively modulates high-frequency responses using low-frequency structural statistics and key-region priors, while its cross-scale branch controls the transfer of shallow structural features through encoder-decoder skip connections. The network is jointly optimized using basic enhancement, identity consistency, key-region structure, and cross-frequency feature consistency constraints. CelebA is used for main training, and LaPa is used for pose adaptation and evaluation. Identity-level partitioning is adopted to avoid identity leakage, with 90%, 5%, and 5% of identities used for training, validation, and testing, respectively. Synthetic low-light images are generated by brightness attenuation, Gamma mapping, illumination-related noise, chromatic noise, color shifts, and black-level offsets. Dark Face is used only for real low-light evaluation. Results and Discussions The proposed model contains 1.47 M parameters and requires 11.74 G floating-point operations for a 256 × 256 input, with an average inference time of 70.48 ms and a speed of approximately 14.19 FPS. On CelebA-Test, the proposed method achieves 25.36 dB PSNR, 0.88 SSIM, 0.8934 ArcFace similarity, 0.0566 Eye-LPIPS, and 0.0753 Mouth-LPIPS, obtaining the best results among the compared methods on all five metrics (Table 1 ). On LaPa-Test, ΔEAR, Eye-LPIPS, ΔMAR, and Mouth-LPIPS reach 0.0065 , 0.0512 , 0.1219 , and 0.0714 , respectively, achieving the best results for all four key-region metrics (Table 2 ). The ArcFace similarity is 0.9398 , lower than those of Low-FaceNet and SCI. Sample-wise analysis shows that this gap is weakly related to pose magnitude and remains in many samples with improved local metrics, indicating partial decoupling between key-region structural fidelity and whole-face identity embedding fidelity. Visual comparisons show improved illumination and better preservation of eye contours, mouth boundaries, and facial textures without obvious local over-enhancement (Fig. 3 ). On Dark Face, the proposed method obtains a NIQE of 10.92±1.12 and improves face-region visibility in representative samples, while challenging cases reveal limitations under nonuniform illumination, occlusion, and extremely low signal-to-noise ratios (Table 3 , Fig. 4 ). Ablation experiments confirm the complementary effects of the three additional constraints and the joint contribution of CFA, HFM, and FSCM, with the complete model achieving the best overall results (Table 4 ). Landmark perturbation experiments further show that the continuous Gaussian heatmaps and structural modulation remain stable under spatial deviations of up to ±6 pixels (Table 5 ). Conclusions A structural-prior-guided cross-frequency collaborative method is proposed for low-light face image enhancement. Low-frequency face structures guide high-frequency detail restoration, while local convolution and two-dimensional selective scanning model fine details and long-range dependencies. FSCM further regulates structural feature transfer across frequencies and scales. Experiments on CelebA and LaPa demonstrate advantages in reconstruction quality, key-region structure preservation, and local perceptual consistency. LaPa results also show that local structural improvements do not always yield equivalent gains in whole-face identity embeddings, indicating room for further optimization of global identity consistency. Real low-light and landmark perturbation experiments indicate a certain degree of scene adaptability and robustness to small spatial deviations. However, performance remains limited under severe occlusion, small-scale faces, extreme low-light conditions, and mixed real degradations. Future work will investigate real low-light face data, denser and confidence-aware regional priors, adaptive constraints for different face regions, joint optimization of local structures and global identity representations, and collaborative optimization with downstream facial analysis tasks.
Available online , doi: 10.11999/JEIT260403
Abstract:
Objective With the rapid popularization of Intelligent Connected Vehicles (ICVs), the interaction between the inherently closed In-Vehicle Network (IVN) and external environments occurs increasingly frequently. Consequently, vehicles are exposed to open network environments, facing severe threats from external cyberattacks such as message injection and semantic manipulation. Once a vehicle is remotely compromised, devastating consequences, including core control system hijacking, sensitive data leakage, and fatal accidents, can occur. As a critical security measure, Intrusion Detection Systems (IDS) are widely recognized by both academia and industry due to their protocol-independent and non-intrusive deployment characteristics. Furthermore, the R155 regulation issued by the UN/WP.29 explicitly requires that vehicles on the market must possess IVN intrusion detection capabilities. While deep learning-based models possess powerful feature extraction capabilities and have become the mainstream direction for IVN intrusion detection, traditional deep Graph Neural Networks (GNNs) and large-parameter Transformers face severe computational and memory constraints when deployed on vehicle edge devices. To address these challenges, RepCAN, a novel lightweight intrusion detection technology for IVNs integrating temporal graph re-parameterization, is proposed in this paper. The primary objective is to balance high-generalization detection capabilities and low resource consumption through a unique "ghost node" mechanism and lossless architecture transformations, specifically targeting the Controller Area Network (CAN) protocol under a vehicle-cloud collaborative architecture. Methods The proposed RepCAN framework is constructed as an efficient lightweight intrusion detection model suitable for real-world automotive environments, comprising three core modules. First, a graph structure construction and data augmentation module is developed. To transform the unstructured raw CAN message stream into semantic tensors processable by GNNs, a hierarchical sliding window strategy is employed. Within each window, dual features encompassing statistical patterns and deep payload semantics are extracted. Statistical features include message entropy, average periodicity, and periodicity jitter to capture fuzzing noise and temporal conflicts. For payload semantics, first-order differences of the payload are calculated to capture non-continuous jumps, and discrete wavelet transforms are utilized to extract the energy spectrum of detail coefficients. This approach successfully identifies disguised attacks without relying on explicit semantic decoding. In terms of graph topology, a lightweight dual-stream graph and a "ghost node" mechanism are designed. A learnable global static adjacency matrix is constructed to learn physical topology control priors, while a dynamic adjacency matrix is generated based on communication frequencies. The ghost node mechanism redirects sudden out-of-distribution messages to a virtual node, transforming traffic bursts into structural feature evolutions. Second, a re-parameterized spatiotemporal feature extraction module is designed. In the training phase, a multi-branch Re-parameterized Graph Convolutional Network (RepGCN) and a Re-parameterized Temporal Convolutional Network (RepTCN) are utilized to capture complex spatiotemporal dependencies. RepGCN handles the spatial stream using a multi-head aggregation structure, allowing the model to capture dependencies among Electronic Control Units (ECUs) from different subspaces. RepTCN employs multi-branch temporal convolutions with varying kernel sizes to capture multi-scale long- and short-term temporal patterns. During the inference phase, structural re-parameterization techniques mathematically collapse these multi-branch architectures into highly efficient single-path structures. Third, an unsupervised anomaly detection module is implemented. Deep Support Vector Data Description (Deep SVDD) is adopted to map the extracted deep spatiotemporal features into a minimum enclosing hypersphere. Anomalies are systematically detected in a completely unsupervised manner by calculating the Euclidean distance between the test sample and the hypersphere center. Results and Discussions Extensive experiments are conducted on the widely recognized Car-Hacking dataset, which includes normal traffic and four attack types: DoS, Fuzzy, Gear, and RPM (Table 1 ). Comparative evaluations against mainstream supervised and unsupervised baseline methods demonstrate the superiority of the proposed RepCAN model. Under an unsupervised setting, RepCAN achieves an Accuracy of 0.9840 , an F1-score of 0.9798 , and an Area Under the Curve (AUC) of 0.9942 , outperforming all unsupervised baselines and rivaling classic supervised models (Table 2 ). Notably, RepCAN maintains a False Positive Rate (FPR) of 0.0117 while achieving a Recall of 0.9774 , indicating a strong ability to balance missed detection and false alarm risks in highly imbalanced vehicular network scenarios. Comprehensive ablation studies are performed to validate the contribution of specific components. The integration of statistical and payload semantic features significantly enhances defense robustness, increasing the F1-score to 0.9798 compared to relying on a single feature domain (Table 3 ). Furthermore, topological ablation experiments confirm that removing the ghost node mechanism drops the F1-score to 0.8902 , highlighting its necessity for the adaptive perception of unknown threats. The removal of the dual-stream structure or the multi-branch re-parameterization module similarly degrades overall performance (Table 4 ). To verify practical deployment feasibility, the framework is evaluated in a vehicle-cloud collaborative environment utilizing an NVIDIA Jetson Orin Nano edge computing device (Fig. 2 ). Through structural re-parameterization, the model's parameters are reduced by 64.15% and Floating-Point Operations (FLOPs) are decreased by 59.32% during inference (Table 5 ). Experimental results indicate that RepCAN achieves an overall inference latency of 0.004827 ms/message on the vehicle-end device (Fig. 3 ). This latency is significantly lower than the shortest physical transmission time limit of a CAN message (approximately 0.1 to 0.15 ms), effectively avoiding competition for computing power with core driving tasks. Additionally, the operational memory overhead of RepCAN is restricted to approximately 360 MB, substantially lower than many baseline models, proving its exceptional lightweight characteristics and adaptability to strictly constrained storage resources (Fig. 4 ). Further verification on the full Car-Hacking test set (Table 6 ) validates the lossless re-parameterization mechanism, with deep feature errors at the 10-4 level and ΔAUC below 0.0007 across all attack scenarios. Additional evaluation on ROAD dataset masquerade attacks (Table 7 ) reveals notable performance degradation, as pure semantic attacks without timing disturbances remain a limitation of this lightweight framework. Conclusions A lightweight intrusion detection technology for CAN buses, RepCAN, integrating temporal graph re-parameterization, is proposed in this study. The innovatively designed ghost node mechanism maps out-of-distribution anomalous traffic to a virtual node, enhancing structural representation and model generalization without expanding the graph scale. Moreover, structural re-parameterization resolves the conflict between high-capacity multi-scale training requirements and low-latency inference constraints. Evaluations on the benchmark dataset and hardware platform confirm that RepCAN achieves a compelling balance between state-of-the-art detection performance and minimal resource consumption. Future investigations will target stealthy camouflage attacks, such as the Masquerade Attack, by incorporating fine-grained semantic analysis mechanisms to eliminate feature perception blind spots.
Available online , doi: 10.11999/JEIT260730
Abstract:
Objective Group activity recognition aims to understand collective behavior among interacting individuals. Unlike individual action recognition, it depends on both actor states and their spatial and temporal interactions. Most methods rely on group activity annotations and individual action labels, making annotation expensive. Recent representation-learning approaches reduce reliance on group labels, but their objectives are often transferred from generic visual learning. Pixel- or low-level feature reconstruction may emphasize appearance unrelated to collective behavior, whereas discrete classification compresses evolving interactions into a few categories. This study investigates group activity representation learning without group labels during representation training. Individual action labels from the standard training split are retained as auxiliary supervision; thus, the setting is not fully label-free or purely self-supervised. Inspired by joint-embedding predictive architectures, we propose a framework that predicts masked actors’ high-level latent semantics and encourages group representations to encode relationships between visible interactions and missing actor states. Methods The framework contains three components: actor feature extraction and masking, latent semantic context reasoning, and joint alignment–contrastive optimization. An ImageNet-pretrained VGG-16 extracts multi-scale feature maps, RoIAlign obtains actor features from trajectory boxes, and a fully connected layer with LayerNorm and ReLU maps them into a 1,024-dimensional semantic space. The backbone and linear projection are initialized by Stage-1 individual action prediction on Volleyball using nine action classes without group activity labels. In Stage 2, they are jointly optimized with the context encoder and predictor. For each 10-frame clip, three trajectories are masked consistently. The complete semantic tensor provides targets, while a binary mask excludes selected actors from spatial aggregation and group pooling. A dual-path Transformer models space-to-time and time-to-space interactions, producing a 2,048-dimensional group representation. Combined with target-position encoding, a three-layer predictor recovers masked semantics. Stop-gradient is applied only to targets, without a teacher network. Cosine loss aligns semantic directions, while cross-video InfoNCE treats visible actors from other clips as unfiltered instance-level negatives. Auxiliary action classification is retained. The loss weights are 1, 0.10, and 0.10, with a temperature of 0.10. For CAD, the framework uses an Inception-v3 backbone and the standard split. Results and Discussions Experiments on Volleyball used retrieval and nearest-neighbor evaluation. In a controlled single-run comparison with the same VGG-16 backbone, Stage-1 initialization, data split, and evaluation protocol, the proposed method achieved 86.46% Hit@1 and 59.58% mAP, versus 84.82% and 57.02% for the baseline. Hit@2 increased from 89.6% to 89.9%, whereas Hit@3 decreased from 91.8% to 91.5%, indicating nonuniform gains across retrieval depths. On CAD, after dataset-specific retraining, Hit@1 increased from 94.90% to 96.47%, whereas mAP decreased from 95.32% to 87.83%, indicating improved local discrimination but weaker global ranking. Across three random seeds, the method obtained 85.49%±0.91% Hit@1 and 58.37%±1.13% mAP, with lower variation than the baseline. With class-balanced galleries of 5–208 samples per class, it exceeded the baseline in Hit@1 by 1.05–3.11 percentage points. Ablations showed that semantic prediction and cosine alignment improve nearest-neighbor discrimination, while InfoNCE improves global ranking. Three masked trajectories best balanced visible context and prediction difficulty. A post hoc analysis found that cross-video pairs sharing the same video-level activity label comprised 13.68%±0.12% of candidates. This is a video-class collision rate, not the true actor-level semantic false-negative rate. Position-only input produced 39.12% Hit@1, showing that court layout alone is insufficient. Removing positional guidance retained 73.22% Hit@1, indicating that appearance and multi-actor context remain informative, although coordinate perturbation revealed sensitivity to localization errors. The Davies–Bouldin index decreased from 3.2339 to 2.8287 , supporting improved compactness and separation. Group-DINOmics uses DINOv3 with a ViT-L backbone and different pretraining resources; therefore, its result is a cross-configuration reference rather than a controlled comparison. Conclusions Masked high-level semantic prediction provides an effective objective for learning interaction-aware group representations when group activity labels are excluded from representation training. The proposed framework uses visible actors to infer missing actor semantics, while cosine alignment and instance-level contrastive learning jointly improve local discrimination and global ranking. The results also show that individual action supervision supplies an important semantic anchor and that positional information is helpful but insufficient on its own. The CAD results further support applicability beyond sports scenes. The current method remains sensitive to localization errors and is less effective for several short-duration or weakly collaborative activities. Future work will investigate more robust spatial relation modeling, finer temporal dynamics, and stronger backbones under unified settings.
Available online , doi: 10.11999/JEIT260796
Abstract:
Objective With the rapid growth of data transmission demands in the Internet of Vehicles (IoV), the limited licensed spectrum has become a critical bottleneck for large-scale IoV deployment. Dynamic access to unlicensed spectrum offers a promising solution. However, existing spectrum sharing methods inadequately address fairness for unlicensed spectrum providers such as Wi-Fi systems, and lack a comprehensive definition of their costs and benefits. To address these issues, this paper proposes a joint power control and resource allocation scheme for IoV paid access to Wi-Fi unlicensed spectrum. A holistic utility function is designed to balance IoV throughput gains and spectrum occupancy costs against Wi-Fi performance degradation due to spectrum leasing. A fairness index and a dynamic pricing factor are introduced to ensure fair value exchange between the two systems. The resource allocation problem is formulated as a non-convex NP-hard utility maximization problem, and an associated Particle Swarm Optimization (aPSO) algorithm is developed by exploiting the correlation between consecutive time slots and a dynamic inertia weight to obtain suboptimal solutions efficiently. The proposed approach aims to enhance IoV data transmission performance while guaranteeing Wi-Fi service quality, achieving an effective balance between resource efficiency and system fairness. Methods A heterogeneous scenario consisting of an IoV system and a Wi-Fi system is considered. The IoV Roadside Unit (RSU) serves vehicular users via licensed spectrum, while multiple Wi-Fi access points operate on independent unlicensed channels. When licensed resources are insufficient, the RSU dynamically accesses multiple unlicensed channels to transmit IoV data packets. Packet arrivals follow a Poisson distribution, packet sizes follow a normal distribution, and each packet must be transmitted on a single channel within a maximum delay constraint. A holistic utility function is constructed by jointly considering IoV transmission gains and spectrum occupation costs against Wi-Fi performance degradation and compensation revenues. A fairness index and a dynamic price factor are introduced to ensure equitable spectrum sharing under Wi-Fi performance constraints. The joint optimization of RSU transmit power, channel occupancy time, and the price factor is formulated as a non-convex NP-hard utility maximization problem. To solve it, an aPSO algorithm is proposed. The algorithm exploits consecutive time-slot correlation to initialize particle positions and adopts a dynamic inertia weight to balance global exploration and local exploitation, thereby achieving an effective suboptimal solution with improved convergence performance. Results and Discussions Simulation results verify the superior performance of the proposed aPSO algorithm. Figure 3 shows that aPSO achieves the highest overall utility across different unlicensed channel numbers by jointly optimizing time occupancy and transmit power, while EPA-aPSO and ETA-aPSO require more power or time due to single-dimensional optimization. Figure 4(a) indicates that aPSO maintains the highest packet transmission success rate, even under overload conditions. In Fig. 4(b) , aPSO’s energy consumption slowly decreases as channel number grows, effectively balancing energy and transmission performance, unlike ETA-aPSO's low power consumption at the cost of a low success rate. Figure 5 demonstrates that aPSO converges at the 28th iteration, about 65% faster than standard PSO, and attains a higher utility. Figure 6 shows that aPSO significantly outperforms GA and SA under \begin{document}$ B=10\;\text{MHz} $\end{document} Figures 7 and 8 further confirm the robustness of aPSO: its utility remains clearly superior under varying channel bandwidths and numbers, whereas EPA-aPSO and ETA-aPSO exhibit limited scalability. Conclusions This paper proposes a joint power control and resource allocation scheme with an aPSO algorithm for IoV paid access to Wi-Fi unlicensed spectrum. By constructing a holistic utility function incorporating a fairness index and a dynamic price factor, the resource allocation is formulated as a utility maximization problem and solved by the proposed aPSO algorithm. The main findings are as follows: (1) The proposed aPSO algorithm converges approximately 65% faster than standard PSO while attaining a higher overall utility; (2) The joint optimization scheme maintains the highest packet transmission success rate even under overload conditions and achieves an effective balance between RSU energy consumption and transmission performance; (3) aPSO consistently outperforms EPA-aPSO, ETA-aPSO, PSO, GA, and SA across varying channel numbers and bandwidths, verifying its robustness and effectiveness in complex resource allocation problems. However, this study considers a single RSU scenario, and future work will investigate distributed cooperative resource allocation mechanisms in multi-RSU and large-scale IoV environments.
Available online , doi: 10.11999/JEIT260522
Abstract:
Objective MilliMeter-Wave Radar (MMW Radar) gait recognition has demonstrated considerable potential for non-contact biometric identification because of its inherent advantages in privacy preservation and robustness to variable lighting conditions. However, practical deployment remains challenging under complex wearing conditions, such as long coats or backpacks. These external factors introduce non-stationary high-frequency interference, resulting in spectral aliasing and masking of intrinsic micro-Doppler (m-D) features. Conventional deep learning methods generally treat m-D spectrograms as generic images and overlook the physical relationship between Doppler frequency and human motion. Therefore, clothing-induced interference is easily confused with motion-related features, resulting in reduced recognition performance. This study proposes a physics-aware framework that integrates radar signal physics with human biomechanics to achieve frequency-domain decoupling and adaptive interference suppression for robust radar-based gait recognition under complex wearing conditions. Methods To reduce clothing-induced interference, this paper proposes PRISM-Net, a physics-aware gait recognition framework for MMW Radar. The framework is built on the biomechanical characteristics of human motion, where the torso, representing the primary body mass, generates relatively stable low-frequency Doppler components, whereas limb motion produces higher-frequency periodic components. (1) Physics-aware Frequency Structural Reconstruction: Instead of uniformly processing the entire m-D spectrogram, the proposed method performs Physics-aware Frequency Structural Reconstruction by exploiting the velocity distribution associated with different body parts. The original aliased m-D spectrogram is reconstructed into torso- and limb-related frequency components. Low-frequency components preserve stable identity-discriminative features, whereas high-frequency components characterize limb motion. This frequency-domain structural reconstruction isolates spectral regions that are most susceptible to clothing-induced interference, thereby reducing interference propagation. (2) Weighted Attention Mechanism (WAM): The WAM adaptively reweights feature responses according to the reliability of different frequency components. Because clothing-induced interference predominantly affects high-frequency regions, the WAM suppresses interference-contaminated high-frequency responses while enhancing stable torso features, thereby improving feature fusion and recognition robustness. (3) Experimental Configuration: The proposed method is evaluated on the MMRGait-1.0 dataset under a subject-independent evaluation protocol. The training set contains data from 74 subjects, while the remaining 47 unseen subjects are used for testing. All m-D spectrograms are resized to 224 × 224 pixels. The network is optimized using the AdamW optimizer with a joint loss comprising Cross-Entropy Loss, Triplet Loss, and Center Loss to improve both classification performance and feature discriminability. Results and Discussions Experimental results demonstrate the effectiveness of incorporating biomechanical priors into gait recognition. As shown in Table 2, PRISM-Net achieves an average Rank-1 accuracy of 89.7% under the 90° side-view condition. In the coat (CT) scenario, the proposed method maintains a Rank-1 accuracy of 85.1%, representing an 18.1% improvement over ShuffleNetV2 and a 7.4% improvement over the ResNet-18 baseline. Model stability is verified through ten independent trials. As illustrated in Fig. 1 , PRISM-Net achieves a standard deviation of ±0.55%, compared with ±1.75% for the baseline model. An independent-samples t-test yields p<0.001, confirming that the improvement is statistically significant. Ablation results in Table 3 further verify the contribution of each component. Removing Physics-aware Frequency Structural Reconstruction reduces the CT Rank-1 accuracy to 75.5%, demonstrating the importance of physics-aware frequency-domain decoupling for preventing feature distortion. The WAM further improves CT accuracy by 4.2% through adaptive suppression of high-frequency interference. Regarding computational complexity, Table 4 shows that PRISM-Net contains 11.33 M parameters and requires only 0.75 GFLOPs. Compared with computationally intensive 3D convolution-based models requiring more than 10 GFLOPs, the proposed method achieves superior recognition performance with significantly lower computational complexity. Furthermore, the t-SNE visualization in Fig. 7 shows more compact intra-class distributions and clearer inter-class separation, demonstrating improved feature discriminability. Conclusions The proposed PRISM-Net demonstrates that incorporating biomechanical priors into deep learning improves the robustness of MMW Radar gait recognition under complex wearing conditions. By combining Physics-aware Frequency Structural Reconstruction with the Weighted Attention Mechanism, the proposed framework effectively performs frequency-domain decoupling and suppresses clothing-induced high-frequency interference. Experimental results on the MMRGait-1.0 dataset demonstrate that the proposed method achieves high Rank-1 accuracy with low computational complexity, indicating its potential for real-time security applications on edge computing devices.
Available online , doi: 10.11999/JEIT260346
Abstract:
Polarization-Adjusted Convolutional (PAC) codes, which are essentially a concatenated structure of polar codes and convolutional codes, significantly enhance error correction performance compared to polar codes, demonstrating particular advantages under short to moderate code lengths. Aimed at the highly dynamic and strongly attenuated channels of Low Earth Orbit (LEO) satellite communications, traditional neural network-based decoders often struggle to adapt to such complex and variable channel environments. In this paper, we propose an adaptive PAC code decoding framework based on a Hypernetwork-assisted Multi-Layer Perceptron (HNA-MLP). This framework generates the weights and biases of the main network through a hypernetwork branch conditioned on Channel State Information (CSI). The main network then performs feature extraction and decoding on the input signal, enabling adaptive network reconfiguration under varying channel conditions. We apply the proposed framework to optimize the Bit Error Rate (BER) of PAC codes under different levels of fading, Doppler shift, and rain attenuation. By calculating the weighted cross-entropy loss and backpropagating to optimize network parameters, the framework aims to minimize the BER. This method paves the way for PAC code decoding in satellite communications with high adaptability and low latency. Simulation results demonstrate that the proposed adaptive hypernetwork decoding outperforms successive cancellation list (SCL) decoding, two-stage adaptive CRC-aided SCL (TA-SCL) decoding, and other neural network-based decoders across a range of rainfall intensities, signal-to-noise ratios (SNRs), and Doppler shifts. Objective Targeting the characteristics of high dynamics and strong attenuation in LEO satellite communication channels, this paper aims to address the issue that traditional fixed-structure neural network decoders struggle to adapt in real time to complex and changing channel environments, resulting in limited decoding performance. The goal is to design a PAC code decoding scheme with high adaptability and low decoding latency. Methods This paper proposes a Hypernetwork-based Adaptive Multilayer Perceptron (HNA-MLP) decoding framework for Polarization Adjusted Convolutional (PAC) codes. The core of this framework is a weight generation mechanism driven by Channel State Information (CSI), which adopts a dual-branch structure: one branch is used for feature extraction, and the other is used to generate an adaptive feature matrix. By feeding real-time CSI into the hypernetwork, the decoder parameters are dynamically adjusted, enabling adaptive tuning according to different fading, Doppler shift, and rain attenuation conditions. During the optimization process, training is performed by calculating the cross-entropy loss with the goal of bit-level accuracy, thereby minimizing the system's bit error rate. Results and Discussions Simulation results demonstrate that within specific signal-to-noise ratio and Doppler shift ranges, the proposed adaptive hypernetwork decoding scheme significantly outperforms traditional SCL decoding and other non-adaptive neural network decoders in terms of Bit Error Rate (BER) performance. This indicates that the CSI-driven dynamic weight generation mechanism can effectively capture the changing characteristics of the channel, enabling the decoder to maintain stable performance in highly dynamic environments. Additionally, by avoiding complex manual feature recomputation, this method also exhibits advantages in decoding time, verifying its potential for engineering applications. Conclusions This paper addresses the challenges of highly dynamic and strongly attenuated LEO satellite communication channels by proposing a Hypernetwork-based Adaptive Multilayer Perceptron (HNA-MLP) decoding framework for Polarization Adjusted Convolutional (PAC) codes. The proposed framework achieves adaptive adjustment of decoder parameters through a weight generation mechanism driven by Channel State Information (CSI). The method is evaluated through multi-scenario simulation experiments, and the results show that: (1) Within the ranges of medium-to-high rain intensity, low signal-to-noise ratio, and high Doppler shift deviation, the proposed adaptive hypernetwork decoding scheme significantly outperforms traditional SCL decoders and TA-SCL decoders in terms of Bit Error Rate (BER) performance; (2) Through the CSI-driven dynamic weight generation mechanism, the proposed decoder effectively adapts to various conditions including different fading, Doppler shift deviations, carrier frequencies, symbol rates, signal-to-noise ratios, and rain attenuation, maintaining stable performance across diverse channel environments; (3) In terms of complexity and decoding efficiency, although the proposed decoder does not surpass some neural network decoders in parameter count and inference speed, it exhibits lower overall computational complexity, and achieves decoding speed improvements of approximately 5.37 times and 2.08 times over the SCL and TA-SCL schemes, respectively. Furthermore, this study is conducted under ideal CSI assumptions; future work will incorporate CSI estimation errors, sensor noise, and hybrid fading in real flight environments, and evaluate the feasibility of GPU platform deployment, so as to enhance the practicality and robustness of the algorithm.
Available online , doi: 10.11999/JEIT260870
Abstract:
Objective Spaceborne Dual-Beam Along-Track Interferometric Synthetic Aperture Radar (DBATI-SAR) acquires fore- and aft-looking radial velocities to support two-dimensional ocean surface current retrieval. The stability of the inversion depends on the relative directions of the two Radar Line-Of-Sight (RLOS) projections on the target local tangent plane. Conventional flat-Earth geometry models generally assume symmetric squint angles or time offsets. However, orbital curvature, Earth curvature, Earth rotation, and local projection nonlinearity produce asymmetric spaceborne geometries. Therefore, symmetric squint angles do not necessarily guarantee orthogonal ground-projected RLOS directions. A method is developed to determine the fore- and aft-looking observation times, squint angles, and down-looking angles directly under the ground-projected RLOS orthogonality constraint. Methods A complete Earth-Centered Earth-Fixed (ECEF) geometry model is established from the satellite state and target position. The RLOS is projected onto the target local tangent plane, and a signed ground-projected RLOS angle is defined relative to the zero-squint reference direction. The squint and down-looking angles are calculated from the observation times rather than prescribed independently. A two-dimensional observation geometry matrix is constructed to relate the horizontal current components to the two radial velocities. Under an ideal symmetric geometry used for the analytical derivation, the singular values and condition number show that a one-sided ground-projected RLOS angle of \begin{document}$ {45}^{{^{\circ}}} $\end{document} \begin{document}$ {-45}^{{^{\circ}}} $\end{document} \begin{document}$ +{45}^{{^{\circ}}} $\end{document} Results and Discussions Numerical experiments are conducted using 500 random orbit-parameter sets. Fifth- and sixth-order polynomials produce relatively large inverse-mapping errors, whereas a seventh-order model substantially improves the accuracy. With 20 samples, the mean ground-projected RLOS angle error is 0.0392°, and further increases in polynomial order or sample number provide limited improvement (Fig. 2 , Table 2 ). The analytical leading term agrees well with the complete ECEF geometry model. For a representative case, the angular root-mean-square error and maximum deviation are approximately \begin{document}$ {0.11}^{\circ } $\end{document} \begin{document}$ {0.31}^{\circ } $\end{document} Fig. 3 ). The conventional flat-Earth geometry model produces a squint-angle correction of up to approximately \begin{document}$ {3}^{\circ } $\end{document} \begin{document}$ {6.5}^{\circ } $\end{document} \begin{document}$ {0.04}^{\circ } $\end{document} Fig. 4 ). The semimajor axis has the strongest effect on the aft-looking observation time, which varies from approximately 34 s to 76 s, while variations caused by other orbital parameters remain within 7 s (Fig. 5 ). Under combined squint- and down-looking-angle perturbations of up to 0.2°, most samples retain ground-projected RLOS angle errors below \begin{document}$ {1}^{\circ } $\end{document} Fig. 6 ). Conclusions A method is proposed to determine key geometric parameters for spaceborne DBATI-SAR under asymmetric geometry. The analytical leading term explains the dominant observation-time scale, whereas the complete ECEF numerical inverse mapping accurately determines the fore- and aft-looking observation times and corresponding squint and down-looking angles. The proposed method provides substantially higher geometric accuracy than the conventional flat-Earth geometry approach. Practical mission design should also consider pulse repetition frequency, azimuth ambiguity, Doppler bandwidth, beam-steering range, and along-track interferometric coherence.
Available online , doi: 10.11999/JEIT260823
Abstract:
Objective Due to feedback delay, the Channel State Information (CSI) used during tag selection may become outdated before data transmission, causing a mismatch between the selected tag and the tag with the largest backscatter-link channel gain during transmission. Most existing studies of outdated CSI assume independent and identically distributed (i.i.d.) channels, which may not adequately reflect the heterogeneous characteristics of practical links caused by different propagation distances. In addition, when the eavesdropper is close to the destination, the legitimate and eavesdropping links may experience correlated fading. Neglecting these factors may cause theoretical results to deviate from actual system performance. Accordingly, under independent and non-identically distributed (i.n.i.d.) channel conditions, the secrecy performance of a multi-tag Bistatic Backscatter Communication (BBC) system is studied under the joint effects of outdated CSI and correlation between the legitimate and eavesdropping links. Methods Candidate tags are ranked according to their backscatter-link channel gains, and the tag with the largest backscatter-link channel gain is selected for transmission. An outdated CSI model is introduced to characterize the mismatch between the CSI used during tag selection and the CSI during data transmission. Because tag selection depends only on the backscatter-link CSI, CSI aging is modeled only for this link. A correlated Rayleigh fading model is used to characterize the statistical dependence between the legitimate and eavesdropping links. Under i.n.i.d. Rayleigh fading channel, order statistics are used to derive the Probability Density Function (PDF) of the selected tag’s outdated backscatter-link channel gain, whereas the joint PDF of the legitimate- and eavesdropping-link channel gains is derived by incorporating their correlated fading relationship. Based on these results, closed-form and high-transmit-power asymptotic expressions for the Secrecy Outage Probability (SOP) are derived. The secrecy performance is further analyzed in terms of the legitimate-to-eavesdropping channel-gain ratio. Results and Discussions Monte Carlo simulations validate the analytical and asymptotic results. The results show that outdated CSI significantly degrades secrecy performance because feedback delay causes the CSI used during tag selection to differ from the CSI during transmission, so the selected tag may no longer provide the largest backscatter-link channel gain. For a fixed legitimate-to-eavesdropping channel-gain ratio, a secrecy outage floor emerges at high transmit power (Fig. 2 ), indicating that increasing transmit power alone cannot eliminate this performance bottleneck. In contrast, increasing the legitimate-to-eavesdropping channel-gain ratio effectively mitigates the outage floor and yields a secrecy diversity order of 1 with respect to this ratio (Fig. 4 ). Under the considered system model, correlation between the legitimate and eavesdropping links also improves secrecy performance (Fig. 3 ) by reducing the probability that the legitimate link experiences severe fading while the eavesdropping link remains strong. Moreover, despite outdated CSI, the proposed tag selection scheme based on backscatter-link channel-gain ranking remains effective and consistently outperforms random tag selection (Fig. 3 ). Conclusions A secrecy-performance analysis framework is developed for multi-tag BBC systems with outdated CSI and correlated legitimate and eavesdropping links. Closed-form SOP and high-transmit-power asymptotic expressions characterize the effects of outdated CSI, link correlation, and the legitimate-to-eavesdropping channel-gain ratio. The results identify outdated CSI as a major source of secrecy degradation and indicate that low-latency feedback is beneficial. Increasing the legitimate-link gain advantage over the eavesdropping link and selecting tags according to backscatter-link channel-gain ranking effectively improve secrecy performance.
Available online , doi: 10.11999/JEIT260236
Abstract:
Objective As maritime ship density increases and the demand for intelligent maritime supervision rises, utilizing multi-source heterogeneous sensors to achieve continuous and robust target tracking has become a critical means of enhancing navigational perception. In nearshore ship monitoring, visual modalities are highly susceptible to environmental interference such as low visibility and occlusions. Furthermore, traditional tracking methods struggle with the severe temporal asynchrony between Automatic Identification System (AIS) data and video trajectories, as well as the instability of identity association in multi-target tracking. This research aims to address these challenges by proposing a ship tracking algorithm based on the deep fusion of AIS and video information combined with spatio-temporal contrastive learning, referred to as STAVFT. The primary goal is to leverage AIS spatial priors to guide visual detectors in locking onto targets under low-visibility conditions and to ensure stable identity consistency in complex multi-target scenarios. Methods The STAVFT algorithm integrates perception-level enhancement with trajectory-level association to create a unified tracking pipeline. At the detection stage, the algorithm employs a Soft Mask feature enhancement strategy and an ROI re-detection compensation strategy. The Soft Mask strategy maps AIS spatial priors into Gaussian-weighted guidance maps that are injected into the feature extraction and fusion stages of the YOLOv11 detector, forcing the network to focus its representational power on high-probability target regions. For targets missed by the initial detector, the ROI re-detection strategy generates high-confidence candidate windows based on AIS priors to perform secondary detection and recover lost tracks. In the association stage, the algorithm constructs a Spatio-Temporal Aware Sequence Transformer (SAST) encoder to handle the asynchronous nature of AIS and video data. By introducing a continuous time-aware self-attention mechanism, the encoder explicitly models the dynamic intensity of heterogeneous trajectories and maps them into a unified embedding space. Additionally, a Spatio-Temporal Negative-sample Contrastive Estimation (STNCE) loss function is designed to strengthen the discrimination between physically adjacent and confusing targets. This loss function incorporates physical constraints, such as spatial distance and motion similarity, to penalize hard negative samples in the feature space. Results and Discussions Experimental results on the FVessel dataset demonstrate that the STAVFT algorithm significantly improves detection performance, raising the recall rate from 65.6% to 79.4% and the mAP50 to 0.792. The SAST encoder handles asynchronous AIS data effectively, achieving a Multi-Object Fusion Accuracy (MOFA) of 0.934, while the STNCE loss successfully mitigates identity swaps in dense traffic by widening the decision boundaries between confusing targets. Overall, the algorithm maintains a stable MOFA above 95.8% and an identity consistency (IDF1) exceeding 96.8%. Although the integration of AIS guidance and ROI re-detection increases the average processing time to 0.334 seconds per video second, the system remains well within real-time requirements, and its adaptive logic provides superior robustness against clutter and track fragmentation. Conclusions This paper presents the STAVFT algorithm, which successfully fuses AIS and video information through spatio-temporal contrastive learning to achieve robust ship tracking. By addressing the issues of environmental interference, data asynchrony, and identity confusion, the algorithm significantly improves the recall rate and identity consistency compared to current mainstream methods. The experimental results on the FVessel dataset validate the scientific validity of the "perception-guided association" logic. Future research will focus on model lightweighting for edge-cloud collaboration to further optimize the deployment performance of the system on embedded terminals.
Available online , doi: 10.11999/JEIT260646
Abstract:
Objective The rapid dissemination of rumors in online social networks (OSNs) poses significant threats to societal stability. Counter-rumor strategy, which combats misinformation by proactively spreading factual information, is a promising non-intrusive countermeasure. Its essence is a resource-constrained optimization problem: selecting an optimal set of seed users to initiate the counter-rumor, aiming to simultaneously minimize the rumor's final influence and the intervention cost. However, existing research predominantly relies on a static network assumption, neglecting the intrinsic dynamic nature of OSNs where user influence, activity, and susceptibility evolve over time and are perturbed by external events. Strategies optimized for a static snapshot of the network often fail when the environment changes, highlighting a critical gap. This work addresses this gap by formally modeling the problem as a Dynamic Multi-objective Optimization Problem (DMOP). The necessity lies in developing an adaptive framework that can track the time-varying optimal trade-off between control effectiveness and resource expenditure, which is essential for robust and practical rumor governance systems. Methods First, a Dynamic Belief-driven Competitive Cascade (DyBCC) model is proposed. Its core innovation is the time-varying user belief, B(v,t), which synthesizes static structural influence with a dynamic component driven by natural decay and external event shocks (Definition 2.1), providing a realistic stochastic simulation framework for competitive propagation under dynamic environments. Second, based on DyBCC, the counter-rumor seed selection is formalized as a Dynamic Multi-objective Optimization Problem for Rumor Control (DMOP-RC), with objectives to minimize expected rumor impact f1 and seed cost f2, where both the objective function and the Pareto-optimal set change over time. Third, to solve DMOP-RC, a Prediction-Memory guided Multi-population Evolutionary Algorithm (PM3EA) is designed. It maintains three collaborative populations: a main population Pm using MOEA/D for local exploitation, a memory archive Pa storing historical elites, and an exploration population Pe using Differential Evolution for global search. A key mechanism is its adaptive response to detected environmental changes (Algorithm 1). Results and Discussions The proposed framework is evaluated in a dynamic environment constructed from a real-world OSN dataset. Algorithm Performance: PM3EA is compared against three dynamic multi-objective optimizers. In terms of the comprehensive Dynamic Inverted Generational Distance (DIGD), PM3EA achieves a significantly superior value of 0.0996 ±0.0569 (Table 2 , Fig. 3b ). Visually, the final Pareto front obtained by PM3EA is the closest to the reference front and shows the best distribution (Fig. 3a ). Its DIGD value remains consistently the lowest and most stable across all time windows, demonstrating robust tracking capability (Fig. 3c ). Model and Strategy Validation: A detailed case study provides micro- and macro-level insights. The evolution of individual node beliefs visually demonstrates the dynamic competition captured by the DyBCC model, showing patterns like belief oscillation in rumor seeds and sudden "clearance" in some nodes upon effective counter-rumor exposure (Fig. 5 ). At the macro level, the effectiveness of the rumor control method is demonstrated on three datasets of different scales (Fig. 6 ), quantitatively proving the suppression effect of the derived seed set. Conclusions This work systematically tackles adaptive rumor control in dynamic social networks. The primary contributions are: 1) The DyBCC model effectively captures the core dynamic of information competition by integrating time-varying user belief, providing a more realistic foundation than static models. 2) The DMOP-RC formulation correctly frames the seed selection as a dynamic trade-off, aligning with practical needs. 3) The PM3EA algorithm, with its tri-population synergy and adaptive change-response mechanisms, demonstrates superior performance in tracking the time-varying Pareto front, outperforming established counterparts. The experiments, from algorithm comparison to case study, holistically validate the effectiveness and superiority of the proposed framework. This work provides a complete methodology for developing intelligent, self-adaptive online rumor governance systems. Future work may explore integrating more complex user behavior models and online learning mechanisms.
Available online , doi: 10.11999/JEIT260388
Abstract:
Objective This paper investigates downlink Multi-User Multi-Input Multi-Output (MU-MIMO) precoding policy learning from the perspectives of computational complexity and energy consumption. Traditional numerical optimization algorithms achieve high performance but exhibit rapidly increasing computational complexity as the numbers of base station antennas and served users increase, leading to high inference latency and energy consumption. In recent years, deep learning has been widely adopted to reduce online computational cost; however, existing evaluations generally rely on training or inference time and FLoating-point OPerations (FLOPs), without direct measurements of energy consumption and power. More importantly, the computational cost of a deep learning model is closely related to network architecture design, which should effectively exploit the wireless prior knowledge of the precoding policy. Therefore, this paper develops a network architecture that matches the multidimensional permutation properties of the precoding policy and systematically investigates how wireless priors affect computational complexity and energy consumption through comprehensive hardware-based measurements and simulations. Methods The MU-MIMO precoding policy is formulated as a mapping from multiuser channel information to the optimal precoding matrix under a transmit power constraint. The optimal policy satisfies multidimensional joint permutation equivariance and invariance with respect to user indices, receive antenna indices, and base station antenna indices. To exploit these wireless priors, an Attention-based Graph Neural Network (AGNN) is proposed based on a hypergraph structure, in which the update and aggregation operations satisfy the required equivariance and invariance properties. An attention mechanism is incorporated to model inter-user interference and improve generalization across different numbers of users. For broadband precoding, multi-subcarrier channel information is aggregated at the input layer to construct an expanded feature representation. To quantify computational energy cost, a hardware measurement platform is developed to collect energy consumption and power for the GPU, CPU, and DRAM during both training and inference. Simulations are conducted using 3GPP TR 38.901 Urban Macrocell (UMa) channel datasets with different antenna array sizes and bandwidth configurations. The proposed AGNN is compared with a traditional numerical optimization algorithm based on Zero-Forcing Block Diagonalization (ZFBD) with greedy user pairing and two Transformer-based architectures that satisfy only one-dimensional permutation equivariance. Results and Discussions Two major findings are obtained. First, incomplete exploitation of wireless priors results in inferior performance and higher computational cost. In the MU-MISO scenario, the Transformer-based architectures achieve lower spectral efficiency than the ZFBD+Greedy baseline while requiring substantially larger models and higher inference FLOPs than AGNN. By matching the multidimensional permutation properties of the precoding policy, AGNN achieves higher spectral efficiency while reducing inference FLOPs by approximately one order of magnitude. Hardware measurements further demonstrate that AGNN reduces inference energy consumption and power on both the CPU and GPU. Second, in small- and large-scale MU-MIMO scenarios, ZFBD+Greedy increases the system sum rate by 10.9×, whereas inference FLOPs, inference time, and inference energy increase by 79.0×, 21.3×, and 38.6×, respectively. In contrast, AGNN increases the system sum rate by 11.5×, while inference FLOPs increase by only 1.89×. Meanwhile, inference time and inference energy are reduced to 0.03× and 0.20×, respectively. These results demonstrate that exploiting the multidimensional permutation properties of the precoding policy provides an effective approach for reducing computational complexity, inference latency, and energy consumption in large-scale 6G MU-MIMO systems. Conclusions This paper investigates the effect of wireless priors on the computational complexity and energy consumption of MU-MIMO precoding learning. By exploiting the multidimensional joint permutation equivariance and invariance of the optimal precoding policy, an AGNN is developed that is well matched to these properties. A hardware-aware measurement platform is established to obtain direct measurements of energy consumption and power for the GPU, CPU, and DRAM during training and inference. Simulations based on 3GPP TR 38.901 UMa channel datasets demonstrate that Transformer-based architectures satisfying only one-dimensional permutation equivariance achieve lower spectral efficiency while incurring substantially higher computational and energy costs. In contrast, AGNN achieves higher spectral efficiency while substantially reducing inference FLOPs, inference time, inference energy consumption, and training complexity. As system size increases, traditional numerical optimization algorithms exhibit much faster growth in computational and energy costs than in system sum rate, whereas the proposed learning method based on wireless priors maintains low inference latency and energy consumption. Overall, exploiting the wireless prior knowledge of the MU-MIMO precoding policy in network architecture design provides an effective solution for computationally efficient and energy-efficient high-dimensional precoding optimization in future 6G networks.
Available online , doi: 10.11999/JEIT260520
Abstract:
Objective Sixth-Generation (6G) mobile networks aim to achieve global seamless coverage through space-air-ground integrated architectures. In this context, Unmanned Aerial Vehicles (UAVs) serve as mobile aerial relay nodes to support massive ground-user access in complex environments. However, traditional data-oriented communication paradigms incur substantial bandwidth overhead, limiting their applicability in spectrum-constrained UAV networks. In addition, conventional centralized resource allocation methods are difficult to implement in real time because of highly dynamic network topologies and the strong coupling among multidimensional resources. To address these challenges, semantic communication has emerged as a communication paradigm that extracts semantic information at the transmitter and reconstructs it at the receiver, thereby reducing redundant data transmission. Existing semantic-driven resource allocation methods, however, primarily focus on static terrestrial networks or single-UAV scenarios and do not adequately address coverage limitations and co-channel interference in multi-UAV relay networks. Therefore, a joint resource optimization model and a distributed resource allocation framework are proposed to improve semantic transmission efficiency and long-term user fairness through the joint optimization of multidimensional resources, thereby supporting intelligent resource scheduling in future 6G integrated networks. Methods A joint resource optimization model is formulated for multi-UAV relay networks under the semantic communication paradigm (Fig. 1 ), jointly optimizing the number of semantic symbols, UAV trajectories, power control, and channel allocation. To evaluate semantic communication performance, a Semantic Communication Quality of Service (SC-QoS) metric is proposed by combining Semantic Quantization Efficiency (SQE) with normalized transmission delay. The optimization objective is formulated as a Mixed-Integer NonLinear Programming (MINLP) problem that maximizes the weighted sum of system-wide SC-QoS and long-term user fairness measured by Jain’s fairness index. To solve this problem, a Two-Stage Hybrid Reinforcement Learning (TS-HRL) framework is proposed (Fig. 2 ). In the first stage, a Capacity-Aware K-means (CA-K-means) algorithm performs heuristic UAV pre-deployment. By introducing a dynamic distance compensation term based on residual capacity, edge users are guided toward lightly loaded UAVs, thereby achieving load balancing while preserving spatial proximity. In the second stage, the dynamic scheduling problem is formulated as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) and solved using a Recurrent Independent Proximal Policy Optimization with Parameter Sharing (R-IPPO-PS) algorithm (Algorithm 1). The algorithm employs Long Short-Term Memory (LSTM) networks to aggregate historical observations and action trajectories, enabling the inference of hidden environmental states. Furthermore, the parameter-sharing mechanism improves training efficiency as the number of agents increases, while the association preference decoupling strategy transforms discrete channel allocation decisions into continuous association preference variables to facilitate policy optimization. Results and Discussions The proposed TS-HRL framework is evaluated in a dynamic environment with randomly moving ground users. Convergence analysis shows that the proposed method achieves a higher initial reward and converges with fewer iterations than the random deployment, memoryless resource allocation, and Enhanced Independent Soft Actor-Critic (EI-SAC) baselines (Fig. 3 ). By combining LSTM-based temporal modeling with CA-K-means pre-deployment, the proposed framework reduces ineffective exploration and improves convergence stability. Compared with conventional bit-based communication, the semantic communication framework increases Semantic Spectral Efficiency (S-SE) by 3.6-fold, reduces transmission delay by 92.1%, and improves fairness by 14.3% (Fig. 4 ). Within the semantic communication framework, compared with the memoryless and heuristic schemes, the proposed method improves S-SE by 92.7% and 51.9%, reduces transmission delay by 13.5% and 57.7%, and improves fairness by 17.3% and 39.7%, respectively. Although the EI-SAC baseline achieves a fairness index of 0.93, its S-SE remains relatively low. In contrast, the proposed TS-HRL framework maintains a high level of fairness while achieving an S-SE approximately 4.3 times that of EI-SAC. Compared with random deployment, the proposed method improves S-SE by 11.3% with only a 1.1% decrease in fairness, demonstrating a better balance between transmission efficiency and fairness. As the number of users increases from 10 to 40, most baseline methods exhibit decreases in S-SE and fairness because of intensified co-channel interference and spectrum limitations (Fig. 5 ). In contrast, the adaptive scheduling strategy and global fairness reward mechanism mitigate performance degradation and maintain the average transmission delay below 0.1 ms. These results demonstrate that the proposed method improves overall system performance while ensuring long-term user fairness. Conclusions This paper investigates joint resource allocation and trajectory optimization for dynamic 6G multi-UAV relay networks under the semantic communication paradigm. A joint optimization model that couples the number of semantic symbols, UAV trajectories, power control, and channel allocation is formulated to maximize SC-QoS and long-term user fairness. To address the high-dimensional coupling of the optimization problem, a TS-HRL framework integrating CA-K-means pre-deployment with the R-IPPO-PS algorithm is proposed for multidimensional resource scheduling under partial observability. Simulation results demonstrate the convergence, stability, and scalability of the proposed method under different user densities. Through distributed multi-agent cooperation among UAVs, the proposed method improves semantic transmission efficiency and reduces transmission delay while maintaining long-term service fairness for ground users. These findings provide an effective approach to intelligent resource allocation for UAV-assisted semantic communication in future space-air-ground integrated networks.
Available online , doi: 10.11999/JEIT260536
Abstract:
Objective To address the complexity and limited modulation capabilities of existing reconfigurable metasurfaces that integrate radiation and stealth functions, a space-time-coding metasurface is proposed for dynamic switching between beam scanning and Radar Cross Section (RCS) reduction. By periodically modulating the states of the meta-atoms in the time domain, differentiated phase distributions are generated for the incident wave and its harmonics. This enables integrated radiation and electromagnetic stealth without complex Transmit-Receive (T/R) components or multilayer structures. The proposed design provides a simple and low-cost approach for integrated radar communication and electromagnetic stealth. Methods The metasurface adopts a metal-dielectric-metal structure, with each meta-atom integrating a PIN diode for 1-bit reflection-phase modulation. Electromagnetic simulations are performed using CST Microwave Studio. At the center frequency of 10.0 GHz, the two diode states provide a reflection phase difference close to 180°, with a near-lossless co-polarized reflection amplitude. Binary Particle Swarm Optimization (BPSO) is used to optimize the space-time-coding sequences for beam scanning and RCS reduction. An 8 × 8 prototype is fabricated using Printed Circuit Board (PCB) technology. A Vector Network Analyzer (VNA) equipped with an S97082A option is used to measure the radiation and scattering characteristics of the prototype. Results and Discussions Simulations and measurements confirm that the proposed space-time-coding metasurface provides beam scanning and RCS reduction. In the radiation mode, the fundamental, +1st, +2nd, +3rd, –1st, –2nd, and –3rd harmonics are steered to +2°, +14°, +32°, +46°, –15°, –28°, and –44°, respectively. The average error between the measured and target beam directions is 1.3°. The average sidelobe levels of the ±1st, ±2nd, and fundamental harmonics are 10.59 dB lower than the corresponding main lobes. In the scattering mode, the peak echo gain at 10.0 GHz is reduced by 11.56 dB relative to a copper plate. Over 9.9～10.1 GHz, the echo gain is reduced by approximately 10 dB, except at 9.93 GHz and 10.04 GHz, where the reductions are 8.07 dB and 7.89 dB, respectively. The maximum reduction reaches 14.83 dB at 9.95 GHz. These results verify the integrated radiation and scattering-control capabilities of the proposed metasurface. Conclusions A space-time-coding metasurface is proposed for integrated radiation and electromagnetic stealth. By integrating PIN diodes into the meta-atoms, the reflection states are periodically modulated on a planar metasurface to generate differentiated equivalent phase distributions at different harmonic frequencies. This enables multi-angle beam scanning without T/R components and reduces the echo gain through spatial and temporal coding. The proposed design simplifies the system architecture and reduces hardware requirements. The demonstrated beam scanning and RCS reduction indicate its potential for integrated radar communication and electromagnetic stealth.
Available online , doi: 10.11999/JEIT260633
Abstract:
Objective Chiplet-based integration provides an important approach for constructing large-scale heterogeneous systems. In a 2.5D chiplet system, inter-chiplet packets travel from a source node to a boundary interface in the source chiplet, traverse an interposer network, and enter the destination chiplet through a destination-side interface. Because packaging resources, micro-bump count, routing resources, and power and area budgets are limited, interfaces can be deployed only at a subset of boundary nodes. Their number and spatial distribution affect access distance, service-region formation, interface-load distribution, and traffic remapping after failures. Existing studies have addressed die-to-die standards, interposer architectures, placement, routing, simulation, and fault tolerance. However, the independent structural effects of boundary-interface configuration remain insufficiently characterized. This study investigates how interface count, location, service-region partitioning, and failures affect communication performance and fault-induced degradation in 2.5D chiplet–interposer networks. Methods A graph-level structural model is developed for a 2.5D chiplet–interposer network. Each chiplet is modeled as an n × n mesh, and selected boundary nodes serve as active interfaces. Each node is mapped to its nearest active interface, and nodes mapped to the same interface form a service region. Average interface-access distance, abstract end-to-end hop count, squared coefficient of variation of interface load, maximum load ratio, and fault-induced degradation are evaluated. Three chiplet scales, n = 4, 8, and 16, are considered, with corresponding interface budgets of k = 2, 4, 6; k = 4, 8, 12; and k = 8, 16, 24, respectively. Five layouts are analyzed: Uniform-corner, Balanced-edge, Two-edge-centered, Single-edge-clustered, and Greedy-DL. Uniform, Hotspot, Transpose, Tornado, and Neighbor traffic patterns are considered. For fault analysis, each active interface is removed in turn, and affected nodes are remapped to their nearest healthy interfaces. A cycle-accurate gem5 model with Ruby/GARNET, comprising four 4 × 4 mesh chiplets and a 4 × 4 interposer mesh, is constructed. Uniform and Hotspot traffic are evaluated at injection rates from 0.01 to 0.15 flit/node/cycle. The latency saturation injection rate is defined as the first sampled rate at which the average packet latency exceeds 50 cycle. This two-level evaluation separates structural effects from cycle-level network behavior and enables interface count, placement, traffic pattern, and fault location to be compared under consistent topology and routing assumptions. Results and Discussions Graph-level results show that increasing the number of boundary interfaces reduces average interface-access distance and abstract end-to-end hop count, whereas the marginal benefit decreases as boundary coverage becomes sufficient (Fig. 3 ). Under a fixed interface budget, different interface locations induce different service-region partitions and interface-load distributions (Table 2 , Fig. 4 ). For an 8 × 8 chiplet with k = 8 under Uniform traffic, Greedy-DL achieves the smallest average interface-access distance of 3.125, whereas Balanced-edge provides a favorable compromise, with an access distance of 3.500 and a maximum load ratio of 1.250. Single-edge-clustered produces the largest access distance and abstract end-to-end hop count because its interfaces are concentrated along one boundary. Under Hotspot traffic, Two-edge-centered reduces the squared coefficient of variation of interface load to 0.016, but its access distance remains higher than those of Balanced-edge and Greedy-DL, indicating a trade-off between distance and load balance (Table 2 , Fig. 5 ). Interface failures cause only moderate increases in average access distance but substantial load concentration at the remaining healthy interfaces. Under Hotspot traffic, the worst-case maximum-load-ratio degradation is 23.94% for Balanced-edge and 93.75% for both Two-edge-centered and Single-edge-clustered (Table 3 ), indicating that fault sensitivity is mainly associated with service-region migration and load reconcentration. The gem5 results further support the graph-level observations. When the interface count increases from k = 1 to k = 4, low-load latency decreases from 24.50/24.44 cycle to 18.05/18.09 cycle under Uniform and Hotspot traffic, respectively, whereas the latency saturation injection rate increases from 0.06 to 0.14 (Fig. 6 ). Across the k = 4 layouts, Balanced-edge achieves low latency and hop count, whereas Single-edge-clustered has the highest baseline communication cost. Under Hotspot traffic, Balanced-edge reduces low-load latency by about 14.3% and average hop count by about 18.7% compared with Single-edge-clustered. Greedy-DL provides short communication paths under low load, but its latency saturation injection rate is 0.13, lower than the 0.14 achieved by Balanced-edge and several other layouts. This result indicates that minimizing access distance alone does not guarantee better medium- and high-load performance (Fig. 7 ). For fault validation, all 16 single-interface failure scenarios are evaluated for each of the three layouts, yielding 48 scenarios in total. The results show increased latency and hop count, together with an earlier onset of congestion. The worst-case latency saturation injection rate decreases to 0.10～0.11, and Balanced-edge exhibits smaller average and worst-case degradation than the more concentrated layouts (Fig. 8 , Table 5 ). Conclusions Boundary-interface configuration is a key structural parameter in 2.5D chiplet interconnect design. It affects interface-access distance, service-region formation, interface-load distribution, and fault-induced traffic remapping. Increasing the interface count improves communication efficiency, but the benefit diminishes as boundary coverage increases. Under the same interface budget, interface placement determines whether traffic remains balanced or becomes concentrated after mapping and remapping. The graph-level metrics support low-cost structural screening and mechanism analysis, whereas gem5/Ruby GARNET simulations provide cycle-accurate validation. Therefore, interface count, location, load balance, and fault degradation should be evaluated jointly in 2.5D chiplet–interposer network design.
Available online , doi: 10.11999/JEIT260556
Abstract:
Objective With the proliferation of edge computing in the Internet of Things (IoT), smart wearables, and offline terminals, low-latency and privacy-preserving Chinese text analysis on local devices has become a core requirement. Although neural network-based and Transformer-based language models achieve high accuracy, their large parameter sizes and computational demands make them difficult to deploy on power- and storage-constrained edge devices. HyperDimensional Computing (HDC), an emerging brain-inspired computing paradigm, represents text using 2 k～10 k-dimensional hypervectors and employs lightweight encoding and querying mechanisms instead of complex multilayer networks, providing a hardware-friendly approach for edge-side text processing. However, existing HDC studies have mainly focused on alphabetic writing systems such as English, in which a limited set of letters is sufficient for N-gram encoding. For Chinese, which contains thousands of commonly used characters, directly applying these methods would require a large number of base hypervectors, resulting in substantial storage overhead and weakening the lightweight advantage of HDC. This paper aims to address this limitation by proposing an efficient character encoding method and a dedicated HDC framework with hardware acceleration for Chinese text classification. Methods Based on the glyph structure of Chinese characters, a hyperdimensional encoding method based on character glyph structure is proposed. Specifically, the encoding process of the Wubi input method is reverse-engineered to decompose each Chinese character into an equivalent Wubi letter sequence for efficient hyperdimensional encoding. A retrieval library covering 3 500 commonly used Chinese characters, as defined in the General Standard Chinese Characters Table issued by the Ministry of Education of the People’s Republic of China, is constructed. For each character, the Wubi library is queried to obtain a letter sequence of length 3 or 4, with each letter mapped to a base hypervector. Cyclic shift and binding operations are then applied to the base hypervectors of consecutive letters. This produces a character-level hypervector that captures both letter identity and positional order information, thereby avoiding confusion caused by different letter permutations. Compared with directly assigning an independent hypervector to each Chinese character, the proposed method reduces the storage requirement for base hypervectors by more than 95% (Table 1). Building on this character encoding method, the HDChinese framework is developed to support inference, training, and retraining. During inference, the sentence-level hypervector obtained by bundling all character hypervectors queries the class hypervectors using cosine similarity. Because the hypervectors are binary, the inner-product operation is efficiently implemented using XNOR logic. During training, class hypervectors are generated by bundling sentence hypervectors belonging to the same class and applying sign-based binarization. During retraining, the nonbinary class hypervectors are iteratively updated using misclassified samples with a learning rate of 0.25, thereby reducing the effect of outlier hypervectors on the cluster center. Furthermore, a dedicated hardware accelerator architecture is designed for HDChinese. The architecture comprises three main modules: an encoding module, a querying module, and a class hypervector update module. These modules can be selectively activated to support inference, training, and retraining modes, respectively (Fig. 2 ). The accelerator is prototyped on an Ultra96v2 FPGA development board equipped with a Xilinx ZYNQ System on Chip (SoC). Results and Discussions Four open-source Chinese text classification datasets covering binary and multiclass tasks are used for evaluation (Table 2 ). The Wubi-based encoding method achieves higher accuracy than the Pinyin-based method on all four datasets and reduces the storage requirement by 31.46% (Table 4 ). Ignoring uncommon Chinese characters, which occur at an average frequency of less than 0.5%, has a negligible effect on accuracy (Table 4 ). A hypervector dimension of 2 k is selected because increasing the dimension beyond 2 k provides no significant accuracy improvement (Fig. 3 ). FPGA measurements show that HDChinese achieves a model size of approximately 16 kB, representing a reduction of more than 99% compared with k-Nearest Neighbor (kNN), Support Vector Machine (SVM), and random forest models (Table 7 ). Compared with SVM and random forest, the proposed accelerator reduces inference time by 6.67%～47.25% and total training time by 4.91%～77.36%. After retraining, the classification accuracy reaches levels comparable to those of the comparison methods (Fig. 4 , Table 7 ). The kB-scale model also enables operation without external memory chips, whereas the comparison methods require more than 400 MB of runtime memory. At 100 MHz, the FPGA implementation consumes 0.278 W, providing the most favorable trade-off between throughput and power consumption (Table 6 ). Conclusions A hyperdimensional encoding method based on Chinese character glyph structure is proposed to address the incompatibility between existing HDC methods and Chinese text. The HDChinese framework and its dedicated hardware accelerator reduce model complexity by more than 99% while maintaining competitive classification accuracy. Training time is reduced by 4.91%～77.36%, and inference time is reduced by 6.67%～47.25%. The proposed approach achieves a favorable balance between accuracy and computational and storage overhead, providing a hardware solution for Chinese text analysis on edge devices.
Available online , doi: 10.11999/JEIT260397
Abstract:
Objective The rapid evolution of Large Language Models (LLMs) has increased the demand for efficient design and evaluation of domain-specific accelerators. Sampling-based performance evaluation methods reduce the cost of cycle-accurate simulation but rely on Instruction Set Simulators (ISS) to execute complete workloads for profiling and checkpoint generation. In LLM inference scenarios, frequent changes in model architectures, inference frameworks, and accelerator instruction sets prevent checkpoint reuse and substantially increase the overhead of ISS-based functional simulation. This overhead has become a major bottleneck in the evaluation workflow. Existing ISS acceleration methods, such as hardware virtualization and Dynamic Binary Translation (DBT), either require target and host systems to share the same Instruction Set Architecture (ISA) or have high implementation complexity, making them difficult to apply to rapidly evolving LLM workloads. Therefore, a flexible and efficient ISS acceleration method is needed to support fast and accurate sampling-based performance evaluation of LLM accelerators. Methods An ISS acceleration method, SkipSync, is proposed for instruction sampling of LLM inference workloads. The key observation is that the control flow of most LLM operators is independent of runtime tensor values and is determined by static parameters. Based on this observation, the Skip mechanism is introduced to bypass the execution stage of accelerator instructions within selected operators while preserving instruction fetch and decode to maintain sampling accuracy. To ensure correct subsequent execution, a lightweight Host-Simulator Synchronization mechanism is further introduced to synchronize host-computed results back to the ISS. Lightweight synchronization primitives and custom instructions are designed to integrate SkipSync into existing sampling-based performance evaluation workflows with minimal implementation effort. Results and Discussions SkipSync is implemented on QEMU and supports RISC-V vector and matrix extensions as accelerator instruction sets. Experimental results show that SkipSync substantially reduces ISS functional simulation overhead while preserving sampling accuracy. Compared with the baseline, average speedups of 4.64× and 5.51× are achieved in the Prefill and Decode stages, respectively (Fig. 5 ), by eliminating the dominant execution cost of vector and matrix instructions, which accounts for 78.67% of runtime in the Prefill stage and 81.95% in the Decode stage (Table 1 ). SkipSync outperforms SIMD_DBT (Fig. 6 ) and provides better support for newly introduced accelerator instructions. The synchronization mechanism introduces limited overhead, with an average runtime increase of only 8.5% relative to ideal Skip-only execution (Fig. 8 ). Thus, synchronization does not negate the performance gains from the Skip mechanism. SkipSync also maintains high sampling accuracy, with an average sampling error of 2.55%, and the sampling results are nearly identical to those of baseline QEMU execution (Fig. 9 ). Conclusions An extensible ISS acceleration method, SkipSync, is presented for sampling-based performance evaluation of LLM inference workloads. By combining the Skip mechanism with Host-Simulator Synchronization, SkipSync effectively alleviates the ISS bottleneck in LLM sampling workflows while preserving execution correctness and sampling accuracy. Experimental results demonstrate substantial speedups, low synchronization overhead, and negligible effects on sampling accuracy. The proposed method provides a practical solution for efficient accelerator evaluation in rapidly evolving LLM inference systems. Future work will extend SkipSync to more complex LLM inference optimization scenarios and explore automatic identification of skippable operators and insertion of synchronization primitives to further improve usability.
Available online , doi: 10.11999/JEIT260414
Abstract:
Objective To address the degradation in Global Navigation Satellite System (GNSS) signal detection performance under weak-signal conditions and the limited control of the probability of false alarm (Pfa) in existing Deep Learning (DL) models, a DSSS signal detection method with a prescribed Pfa is investigated. The method enables the DL detector to be evaluated within a Constant False Alarm Rate (CFAR) framework. Methods DSSS signal detection is formulated as a binary classification problem, and a DL-based detection framework is developed. An improved one-dimensional ResNet-18 is designed for I/Q sampled time-series data. The input convolution kernel is set to 1×7, and the initial maximum pooling layer is removed to preserve weak-signal temporal features. The first residual layer is also configured without downsampling. A noise score-pool-based threshold calibration mechanism is developed to impose Pfa constraints on the detection decision. A large number of pure-noise samples are processed by the trained network to obtain the empirical distribution of confidence scores for the signal-present class. The decision threshold is then calibrated according to the quantile corresponding to the preset Pfa. In addition, the effect of signal normalization on detection performance is evaluated. The proposed method is validated using a simulated GPS L1 C/A signal dataset under different Signal-to-Noise Ratio (SNR) and Pfa settings and in non-ideal colored-noise environments. Results and Discussions The proposed method achieves high detection performance under different Pfa constraints. At a Pfa of 0.01, the detection probability reaches 100% at an SNR of –8 dB. When the Pfa decreases from 0.01 to 0.001 and 0.000 1, the detection curve shifts toward higher SNRs, but the decrease in detection performance remains limited. The unnormalized preprocessing strategy consistently outperforms Root-Mean-Square (RMS) normalization, providing a performance gain of approximately 1 dB. Compared with the traditional autocorrelation detection method, the proposed DL-based detector provides a detection performance gain of approximately 3～4 dB across the tested Pfa settings. In colored-noise environments not used for network training, the proposed method maintains effective detection performance and demonstrates robustness to noise mismatch. The structural ablation results further show that removing the maximum pooling layer and retaining the temporal resolution of the first residual layer improve detection performance under low-SNR conditions. Conclusions A DL-based DSSS signal detection method with noise score-pool-based threshold calibration is proposed. The empirical distribution of pure-noise confidence scores is used to calibrate the decision threshold, thereby incorporating the DL detector into a CFAR-based detection framework. The improved one-dimensional ResNet-18 effectively extracts features from I/Q time-series data, whereas the unnormalized preprocessing strategy preserves useful signal-amplitude information. The proposed method improves detection sensitivity while maintaining effective Pfa control and exhibits robustness under non-ideal colored-noise conditions.
Available online , doi: 10.11999/JEIT260889
Abstract:
Objective: A comprehensive multimodal benchmark is developed for Anti-Unmanned Aerial Vehicle (UAV) detection in low-altitude urban environments. Existing datasets generally provide limited sensing modalities and UAV models, with relatively coarse annotations that constrain tasks requiring spatial, motion, and cross-modal information. DroneRFc-MM addresses these limitations by providing synchronized multimodal data, broader coverage of consumer-grade DJI UAV models, and fine-grained annotations for target detection, UAV model recognition, trajectory analysis, flight-direction reasoning, and multimodal fusion evaluation. Methods: DroneRFc-MM is synchronously collected using six heterogeneous sensor types: a Pan-Tilt-Zoom (PTZ) camera, a fisheye camera, a Radio Frequency (RF) antenna, LiDAR, millimeter-wave radar, and a microphone array. Data are acquired on an open rooftop at a university in Zhejiang Province, representing a typical urban low-altitude environment. The dataset contains recordings of six consumer-grade DJI UAV models. All devices are synchronized using a common network time reference, with inter-device timestamp discrepancies of approximately 0.3 s. The UAVs fly in “H”-shaped and vertical reciprocating trajectories at distances of 20–60 m from the sensor array. Fine-grained annotations, including UAV model, position, attitude, and velocity, are derived from flight logs. For the flight-direction reasoning task, approximately 5-s multimodal clips are generated, including camera videos, RF spectrogram videos, microphone audio, and coordinate-based text representations of radar point-cloud data. Zero-shot inference is conducted using Qwen 3.6-Plus and Qwen 3.5-Omni-Plus with unified prompts. Prediction accuracy and inference time are evaluated by comparing predicted directions with ground-truth directions calculated from UAV positioning data. Results and Discussions: The DroneRFc-MM dataset provides multimodal data from six sensor types and six consumer-grade DJI UAV models, together with fine-grained annotations and sample extraction tools. In the flight-direction reasoning task, the Qwen-series multimodal large language models (MLLMs) achieve accuracies ranging from 20% to 30% across the different input modalities. The inference time is also relatively long, with the mean response time exceeding 40 s for most sensor inputs. These results indicate that current general-purpose MLLMs can capture weak motion-related information from UAV videos, audio, RF spectrograms, and point-cloud data, but their accuracy and response speed remain insufficient for practical real-time Anti-UAV detection. Conclusions: DroneRFc-MM provides a multimodal benchmark for Anti-UAV detection, UAV model recognition, flight-direction reasoning, and multimodal model evaluation. The dataset integrates six sensor types, six consumer-grade DJI UAV models, and fine-grained annotations within a common measurement framework. The experimental results show that current general-purpose MLLMs remain limited in flight-direction reasoning and real-time inference in Anti-UAV scenarios. Domain-specific pre-training, supervised fine-tuning, knowledge augmentation, and lightweight inference are therefore needed to improve their practical utility. Future work will expand the dataset scale and application scenarios to support intelligent and efficient low-altitude airspace management systems.
Available online , doi: 10.11999/JEIT260565
Abstract:
Objective The rapid progress of large-scale intelligent computing, especially Mixture-of-Experts (MoE), has positioned Three-Dimensional Integrated Circuit (3D IC) based on Through-Silicon Via (TSV) as a key solution to memory-wall bottlenecks via high bandwidth and density. As a core 3D IC technology, TSVs enable vertical inter-chip connections, reducing path length, parasitic delays, power, and boosting data rates. MoE-specific accelerators, characterized by high data density and strong fault tolerance, introduce new challenges and opportunities for TSV layout. These include aggravated signal integrity and reliability issues in dense arrays, and the inadequacy of static TSV allocation for dynamic, bursty MoE traffic. Conversely, their inherent fault tolerance permits optimization design spaces for employing fault-tolerance mechanisms. This paper exploits MoE dataflow characteristics and hardware fault tolerance to devise a data mapping strategy for high-density TSV arrays based on fault-tolerance mechanisms, targeting improved performance and reliability. Methods This paper investigates cluster partitioning schemes and data mapping strategies to enhance the reliability of TSV data transmission. To address the high complexity of global optimization in large-scale TSV arrays, a cluster size partitioning scheme is proposed. By structurally partitioning a large-scale TSV array into several small-scale TSV clusters, the global optimization problem is decomposed into local, scalable subproblems, thereby improving optimization efficiency and flexibility while ensuring optimization moderation. Through comprehensive consideration of multiple metrics and simulation-based evaluation, the cluster size is finally determined to be 6×6. In response to the varying dataflow characteristics and load distribution across different computational stages, this paper proposes a Phase- and Load-Aware Dynamic Data Mapping (PLDM) strategy. The strategy pre-partitions the TSV array into multiple fixed-size clusters and classifies them into critical clusters and general clusters based on metrics such as coupling strength, bandwidth, and latency. At runtime, the PLDM strategy dynamically adjusts data mapping according to the characteristics of different computational stages. Furthermore, this paper achieves a co-optimization design of PLDM with the encoding circuit. The load monitoring module and the error monitoring module share certain data buffers and control status registers, enabling hardware resource reuse. Meanwhile, the error monitoring results provide real-time feedback on the reliability level of each TSV cluster, based on which the mapping controller preferentially allocates data transmission to clusters with lighter loads and lower bit error rates. This approach realizes resource sharing and load balancing, thereby improving data transmission reliability and link utilization efficiency for high-density TSV arrays. Results and Discussions This paper analyzes the bandwidth utilization and load balancing performance of three mapping schemes: random, static, and dynamic. The results show that both the dynamic and random mapping schemes achieve average bandwidth utilization close to the theoretical maximum. However, the random mapping scheme maps approximately 37.52% of critical data into general clusters with relatively high bit error rates, thereby increasing unreliability. Compared with static mapping, the dynamic mapping scheme improves average bandwidth utilization from 0.7982 to 0.8984 , a relative increase of about 12.6%, reduces inter-cluster load fluctuation by 54.6%, and correspondingly improves load balancing by a factor of 2.2 (Fig. 4 ). Compared with random mapping, the dynamic mapping scheme reduces inter-cluster load fluctuation by about 8.6%, and reduces the latency of critical data and non-critical data by 36.6% and 34.7%, respectively (Table 2 ). To further evaluate the optimization effects of the proposed PLDM strategy on metrics such as load balancing and bandwidth utilization, four comparative schemes are configured: (1) Baseline scheme; (2) Static mapping scheme; (3) Sparse TSV layout using TSV-Aware Adaptive Fault-Tolerant Coding (TSV-AFTC) and PLDM; (4) High-density TSV layout based on scheme (3). Taking the Qwen3-30B-A3B model as an example, the normalized loads of 16 clusters in the Multi-Head Attention (MHA) and Feed-Forward Network (FFN) stages are compared across the four schemes. The results indicate that the proposed dynamic data mapping scheme achieves balanced load distribution across clusters in both the MHA and FFN stages, ranging from 0.48 to 0.52, while ensuring that all critical data are mapped to critical clusters. The high-density TSV scheme further reduces the load per cluster to approximately 0.34–0.37, demonstrating that dynamic mapping can effectively suppress stage-wise hot spots and improve load balancing (Fig. 7 ). Subsequently, system-level fault injection is applied to the transmitted data to simulate data reliability under extreme conditions for different schemes. The results show that for the proposed scheme (sparse), the degradation in perplexity (PPL) compared to the ideal case is controlled within 0.02, while the average bandwidth utilization is improved by approximately 15% and cluster load balancing is enhanced by a factor of 3.4. Under the high-density scheme, the PPL increase is controlled within 0.05, the average bandwidth utilization reaches about 71.7%, and the cluster load balancing is improved by a factor of 2 (Table 4 ). Conclusions This paper investigates TSV data mapping for 3D MoE accelerators and proposes a PLDM strategy based on TSV-AFTC, which allocates data from different computational stages to reliable and lightly loaded TSV clusters according to cluster-level bit error rates and load conditions. Through circuit co-design, approximately 12% of hardware resources can be saved. Compared with static mapping, the proposed scheme improves average bandwidth utilization by about 12.6% and enhances cluster load balancing by a factor of 2.2. Under system-level fault injection, the scheme limits the degradation of model inference perplexity to within 0.02, while achieving approximately 15% improvement in bandwidth utilization and a 3.4× enhancement in cluster load balancing.
Available online , doi: 10.11999/JEIT260958
Abstract:
Objective Due to their capacity approaching performance, Low Density Parity-Check (LDPC) codes have been widely applied to wireless communication and data storage systems. Compared to their binary counterparts, Non-Binary LDPC (NB-LDPC) codes with short or moderate code lengths have been demonstrated to achieve superior error performance under non-binary Belief Propagation (BP) decoding. However, the computational complexity of Check Node (CN) update of the optimal BP decoding is too complex for practical applications. Recently, many works have been presented to perform updates of CNs based on truncated messages, rather than full-length reliability messages, to significantly reduce the computational complexity of CN updates. Most of them construct the trellis of a CN based on the truncated input vectors, called truncated-trellis, such that CN updates are efficiently processed in parallel based on the selected candidate paths. These paths generally contain only a small number of deviation nodes, and such deviation nodes usually have high reliability. However, the Variable Node (VN) update in most decoding algorithms based on CN truncated-trellis still sequentially processes each element in the input vectors of each VN by the elementary steps. When the CN update is simplified, the complexity of the VN update may primarily determine the overall computational complexity. To address the above issues, this paper proposes the Dual-Trellis Min-Sum (DTMS) decoding algorithm. By further introducing truncated-trellises for VNs and updating the output messages of CNs and VNs in parallel, respectively, it further improves the decoding efficiency, while maintaining the similar decoding performance. Methods The different contributions of nodes in the CN truncated-trellis of the Pruning path Min-Sum (PMS) decoding algorithm on the selected highly reliable candidate paths are first analyzed, and it reveals that the selected highly reliable paths are primarily determined by the deviation nodes from the first few rows of the trellis of a CN, especially the second row. Thereby, it is not critical to update and sort every element of each output vector of one VN during the VN update. Next, a new trellis of one VN is constructed, and highly reliable elements over this trellis shared by all the output vectors of this VN are searched using a row-wise pruning strategy, such that the conventional element-wise VN updating procedure is transformed into a trellis-based parallel updating process based on an extra column in the trellis. In this basis, the unequal protection for the reliability values of each VN output vector is conducted, e.g., only the first few elements in each output vector of VN are updated and arranged, and the rest elements of each output vector are directly set to a compensation value. As a result, the computational complexity required for less reliable elements during each VN update can be significantly reduced, while retaining the crucial messages. Results and Discussions Experimental results show that compared with the PMS decoding algorithm using the original VN updating procedure, the proposed DTMS decoding algorithm maintains almost the same Bit Error Rate (BER) performance and convergence speed for decoding NB-LDPC codes under different finite fields, code lengths, and code construction methods (Figs. 3 –7 ). Meanwhile, the number of real-domain operations required for the proposed simplified VN updates is reduced by approximately 71.8% on average (Table 2 ). In addition, the error-correction performance and convergence speed of the proposed DTMS decoding algorithm are close to those of the sub-optimal BP decoding algorithms (Figs. 3 –7 ) with relatively low computational complexity (Table 3 ). The average performance gap of the DTMS decoding algorithm from the optimal BP decoding algorithm is only about 0.11 dB (Figs. 3 –7 ). Thus, optimizing the VN updating is an effective way to further reduce the decoding complexity of truncated-trellis-based message-passing decoding algorithms. Conclusions This paper proposes a DTMS decoding algorithm to reduce the computational complexity of VN update in truncated-trellis-based decoding algorithms for NB-LDPC codes. Based on the CN updating process of the PMS decoding algorithm, the proposed algorithm further constructs a truncated-trellis and introduces the unequal protection scheme for VN update, such that the output vectors of each VN can be efficiently updated in parallel. Experimental results show that, under the same CN trellis-based update, the proposed parallel VN updating method significantly reduces the computational complexity compared with the original VN updating method, while maintaining similar decoding performance. Moreover, the proposed DTMS decoding algorithm performs closely to the sub-optimal BP decoding algorithms with similar convergence speed and lower complexity. In future studies, it will be interesting to further exploit the adaptive pruning strategies for the VN parallel updates. Based on the distribution of field elements from different iterations, less reliable field elements can be adaptively eliminated to reduce the set of candidate field elements, which may further reduce the complexity of VN update with negligible performance loss.
Available online , doi: 10.11999/JEIT260922
Abstract:
Objective Troposcatter communication serves as a valuable complement to satellite communication and thus is still quite promising in scenarios such as military long-distance communication. However, when a troposcatter communication system is deployed in mountainous environments, it is often faced with a prevalent and challenging engineering problem known as the “Line-of-Sight (LoS) obstruction”. Traditional solutions to this issue are still confronted with engineering difficulties. Increasing the antenna elevation angle to cross obstacles makes the scattering angle increase sharply and consequently lead to transmission loss surging beyond acceptable link budget limits; alternatively, building tall towers to raise antenna height preserves low-angle transmission but introduces construction difficulties and sacrifices the advantage of terrain concealment. Reconfigurable Intelligent Surface (RIS) has emerged as a disruptive technology in wireless communications, with the capability of reconstructing the wireless environment and artificially altering channel characteristics. It has been successfully applied in various civilian mobile communication systems. Obviously, it also provides an alternative to address the problem of LoS obstruction in troposcatter communications. Unfortunately, it has never been reported that RIS had been applied in such scenarios. Herein, to solve the LoS obstruction problem in troposcatter communications, RIS is involved for the first time in this field, a conceptual architecture of RIS-assisted troposcatter communication is set up, and then the problem of optimal RIS deployment is systematically investigated. Methods Based on the proposed framework of RIS-assisted troposcatter communication system, the deployment optimization of RIS is addressed step by step:Firstly, a three-dimensional model of feasible deployment region is established under four types of practical engineering constraints, i.e., the intrinsic constraint of obstacle-crossing, the optimal constraint of engineering upper bound, the antenna radiation constraint of Fresnel near-field region, and the hardware constraint of RIS effective angle.Secondly, the optimization problem of RIS deployment is formulated as minimizing the comprehensive system gain loss. The overall loss consists of three major components, namely, troposcatter transmission loss, free space path loss, and the dynamic gain attenuation of Cassegrain antennas with respect to their elevation angles. The first two parts are easily computed according to corresponding engineering knowledge of troposcatter communication and typical antenna theory, respectively. Then to calculate the dynamic gain attenuation of Cassegrain antennas, a quantitative model is developed based on cantilever beam bending theory. The model quantifies the pointing errors of a Cassegrain antenna caused by dynamic over-compensation with its elevation angle adjustment, and then the gain loss is calculated with the assistance of Taylor radiation pattern.Thirdly, through theoretical analysis and numerical verification via sectional slicing heatmaps, a dimensionality reduction property of the objective function is observed and validated. Within the feasible region, the first-order partial derivative of the objective function with respect to deployment height is always negative, implying that the global optimal deployment position necessarily lies on the upper boundary surface of the feasible region. The dimensionality reduction property is rigorously validated through slice analysis across the entire feasible domain, with more than 46,100 verification points confirming that the optimal position always resides on the upper boundary surface. This finding reduces the intractable three-dimensional constrained optimization problem to a much simpler two-dimensional manifold optimization, which significantly reduces the computational complexity of the optimization.Finally, based on this dimensionality reduction property, an improved gradient descent algorithm with momentum and adaptive backtracking line search (IGD-M&ABLS) is put forward. The algorithm introduces momentum gradient updates to suppress zigzag oscillations and accelerate convergence; it also incorporates an adaptive backtracking line search strategy to dynamically adjust step sizes, balancing iterative stability with computational efficiency. Results and Discussions A series of simulation experiments are conducted under typical engineering parameters, i.e., a 3-meter aperture Cassegrain antenna; 5 GHz signal frequency; mountain heights of 150 m, 200 m and 250 m, representing medium-high hills, the dividing line between hills and mountains, and relatively mountainous terrain, respectively. The proposed IGD-M&ABLS algorithm is benchmarked against Grid Search (GS, a classic deterministic exhaustive-search method) and Particle Swarm Optimization (PSO, a typical efficient heuristic algorithm). The results demonstrate that IGD-M&ABLS consistently converges to the global optimal solution with less gain losses than both benchmarks. Specifically, for the typical case with a 200 m mountain height, in 50 independent runs, IGD-M&ABLS always achieves the best objective function value of 11.6094 dB, much more steadily than PSO does, and it also outperforms GS's 11.6127 dB. As far as time consumption is concerned, IGD-M&ABLS exhibits remarkable advantages of computational efficiency. Its average runtime is approximately 0.005 seconds, compared with 0.07 seconds of PSO and more than one hour of GS. This order-of-magnitude improvement in computational speed is consistent with the theoretical complexity analysis. IGD-M&ABLS optimizes two independent variables on a 2D manifold, whereas PSO and GS handle three independent variables in the 3D feasible domain. Robustness tests under varying terrain conditions (H = 150 m and H = 250 m) confirm that IGD-M&ABLS reliably obtains the best results across different scenarios. In all simulation tests, IGD-M&ABLS demonstrates excellent stability and reproducibility, producing consistent results across multiple independent runs, while PSO exhibits randomness-induced variations and GS remains limited by its discretization step size. Conclusions This paper pioneers the application of RIS technology in troposcatter communication, providing a new technical solution to address the LoS obstruction problem in mountainous environments. A conceptual framework of RIS-assisted troposcatter communication system is established, incorporating a three-dimensional feasible RIS deployment region model with four practical engineering constraints. Then the optimization problem of RIS deployment is formulated as minimizing the overall system gain loss including troposcatter transmission loss, free space path loss, and the dynamic gain attenuation of Cassegrain antennas with respect to their elevation angles, and the computation method of its third term is also developed for the first time based on cantilever beam bending theory. More interestingly, the objective function is found and verified with dimensionality reduction property, i.e., its minimum value always resides on the upper boundary manifold surface. That property effectively transforms the complex 3D optimization into an equivalent 2D manifold problem. Finally, a new algorithm IGD-M&ABLS is proposed by introducing momentum and adaptive backtracking line search into the traditional gradient descent framework. Simulation results show that compared with benchmarks, IGD-M&ABLS algorithm achieves the best deployment positions with order-of-magnitude faster computation, while maintaining excellent stability and reproducibility.
Available online , doi: 10.11999/JEIT260648
Abstract:
Objective Reliable data delivery in underwater acoustic communication networks is challenged by high packet error rates, long propagation delays, limited bandwidth, and topology variations. In single-source dual-destination multi-hop transmission, the same data generation must be reliably delivered to two destination nodes. Packet losses at individual hops can accumulate during multi-hop forwarding and joint recovery at the two destinations, further complicating reliable delivery. Existing reliability-enhancement mechanisms, including retransmission, redundant forwarding, forward error correction, and multipath redundant transmission, generally rely on predetermined forwarding structures or fixed redundancy configurations. They have limited capability to exploit complementary coded information distributed among multiple relay nodes, resulting in insufficient joint recovery capability and high redundant transmission overhead. To address these limitations, a Network-Coded Cooperative Reliable Transmission Protocol for underwater acoustic communication networks (NCCRTP) is proposed. Methods NCCRTP operates on a generation basis and employs Random Linear Network Coding (RLNC) over the Galois field \begin{document}$ \text{GF}({2}^{8}) $\end{document} Results and Discussions Simulation results show that NCCRTP achieves the highest Joint Packet Delivery Ratio (JPDR) under both regular and random topologies. In the controlled comparison with Cooperative Uncoded transmission (CU), Single-branch Uncoded transmission (SU), and Single-branch Coded transmission (SC), NCCRTP consistently outperforms schemes using only cooperative forwarding or only RLNC. This result indicates that the reliability gain is jointly provided by distributed relay cooperation and joint utilization of linearly independent coded packets (Fig. 5 ). As the packet error rate increases or the end-to-end transmission depth increases from 3 to 7 hops, NCCRTP maintains a higher JPDR, demonstrating stronger robustness under lossy multi-hop conditions (Figs. 5(a) and 5(b) ). In random topologies, Vector-Based Forwarding (VBF) and Focused Beam Routing (FBR) are separately combined with packet REPlication (REP) or RLNC to form the VBF+REP, VBF+RLNC, FBR+REP, and FBR+RLNC schemes (Figs. 6 and 7 ). Under medium-to-high packet error rate or multi-hop transmission conditions, NCCRTP improves the JPDR by up to approximately 50%, while reducing the equivalent transmission overhead per successful joint delivery by up to approximately 40% (Fig. 6 ). These results indicate that NCCRTP improves dual-destination reliability through adaptive forwarding-structure selection, link-quality-based transmission-budget control, and joint utilization of linearly independent coded packets rather than simply increasing redundant transmissions. Conclusions The reliability and redundant transmission overhead challenges in single-source dual-destination underwater acoustic multi-hop transmission are addressed by designing a cooperative transmission structure that enables RLNC to exploit distributed reception and complementary coded information among relay nodes. The proposed NCCRTP protocol adaptively selects the SINGLE, COOP, and BRANCH transmission modes according to residual-hop constraints, link-quality-based transmission-budget control, and two-hop structural utility evaluation. A lightweight coding-vector representation based on CodeID is also adopted to reduce the header overhead associated with carrying complete coding vectors. The protocol is evaluated under both regular and random topologies. The results show that: (1) NCCRTP achieves the highest JPDR among all compared schemes, demonstrating stronger joint recovery capability at the two destination nodes; (2) under medium-to-high packet error rates or multi-hop transmission conditions, NCCRTP improves the JPDR by up to approximately 50%; and (3) the equivalent transmission overhead per successful joint delivery is reduced by up to approximately 40%, indicating that the reliability gain mainly comes from adaptive structure selection, transmission-budget control, and joint utilization of linearly independent coded packets rather than excessive redundant transmissions. Future work will extend NCCRTP to more complex multi-source, multi-destination, multi-hop scenarios and further investigate its implementation and performance under node mobility and realistic underwater acoustic channel dynamics.
Available online , doi: 10.11999/JEIT260375
Abstract:
Objective In Sixth-Generation (6G) Low Earth Orbit (LEO) satellite communication systems, multicast precoding is adopted to mitigate severe inter-beam interference caused by Full Frequency Reuse (FFR). However, conventional precoding algorithms exhibit cubic computational complexity, limiting their applicability to massive Multiple-Input Multiple-Output (MIMO) systems. Existing user grouping methods also fail to satisfy the fixed group-size requirement specified by the DVB-S2X standard. To address these limitations, a joint optimization framework is proposed that combines a low-complexity unsupervised deep learning-based precoding model with improved user grouping algorithms to improve the system sum rate and fairness. Methods An unsupervised deep learning model based on a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) architecture is proposed for precoding (Fig. 2 ). The Convolutional Neural Network (CNN) extracts spatial features from Channel State Information (CSI), while the Long Short-Term Memory (LSTM) network captures high-level feature correlations. The model is trained by directly maximizing the system sum rate while satisfying the Per-Antenna power Constraint (PAC), without requiring supervised labels. For user grouping, two algorithms compatible with the DVB-S2X standard are developed. First, the CK-means algorithm extends conventional K-means clustering to ensure an equal number of users in each group while preserving high intra-group channel similarity. Second, the Fairness-Aware MAUG (FA-MAUG) algorithm prioritizes users with poor channel conditions during grouping, thereby improving system robustness and fairness. Results and Discussions The intra-group similarity metric is used to evaluate user grouping performance. The results show that the CK-means algorithm achieves an average similarity approximately 0.1 higher than that of the MAUG algorithm and nearly 0.5 higher than that of random grouping across different group sizes (Fig. 3 ), resulting in improved beamforming gain. In terms of system sum rate, the proposed CNN-LSTM precoding combined with CK-means grouping consistently outperforms the conventional Minimum Mean Square Error (MMSE) algorithm under different Signal-to-Noise Ratios (SNRs) and total transmit power levels (Fig. 4 and Fig. 5 ). Under different SNR conditions, the proposed CNN-LSTM precoding scheme improves the average system sum rate by 48.59% compared with the MMSE algorithm, whereas CK-means grouping increases the average system sum rate by 30.12% relative to random grouping. The effects of the number of users per group, the number of groups, and the number of antennas are further evaluated (Fig. 6 -Fig. 8 ), demonstrating that the proposed framework maintains superior performance across systems of different scales. Complexity analysis further shows that the proposed precoding method reduces the online computational complexity from the cubic complexity of the conventional MMSE algorithm to linear complexity with respect to the number of antennas, making it well suited for real-time deployment in large-scale LEO satellite communication systems. Conclusions A joint optimization framework is proposed for LEO satellite multicast communication systems to address the high computational complexity of precoding and the limited fairness of conventional user grouping methods. Simulation results demonstrate that the proposed framework substantially improves the system sum rate, achieving an average gain of 48.59% over the conventional MMSE algorithm under different SNR conditions while simultaneously reducing online computational complexity. The proposed framework provides an effective and scalable solution for multi-beam interference mitigation and resource optimization in future 6G LEO satellite communication systems.
Available online , doi: 10.11999/JEIT260751
Abstract:
Objective Image classification depends on complete and discriminative feature representations. Existing convolutional neural networks usually enhance positive and high-amplitude responses through activation functions and attention mechanisms. However, excessive reliance on dominant positive responses may shift attention from the whole object to local salient regions, while negative and low-amplitude responses containing edge, texture, and foreground-background transition information are often weakened. To address this problem, a Complementary Decay Learning Network (CDLNet) is proposed to suppress high-response dominance and preserve complementary feature information. Methods Inspired by the signal attenuation mechanism of the biological visual system, CDLNet introduces complementary decay learning into a residual network. The Spatial Complementary Decay (SCD) module divides features into positive-response, negative-response, and global-response branches, and applies differentiated decay to preserve salient regions, boundary details, and contextual information (Fig.2 ). The Channel Complementary Decay (CCD) module attenuates high-response channels while retaining middle- and low-response channels, thereby reducing channel dominance and promoting cooperative channel representation (Fig.3 ). The Complementary Decay Attention (CDA) module integrates SCD and CCD in parallel and is embedded into ResNet residual blocks to jointly regulate spatial structures and channel semantics (Fig.4 , Fig.6 ). Results and Discussions Experiments are conducted on CIFAR-10, CIFAR-100, SVHN, Imagenette, Imagewoof, and ImageNet datasets. CDLNet achieves classification accuracies of 96.68%, 81.81%, 97.38%, 92.45%, 85.46%, and 62.88%, respectively. Compared with ResNet-34 and representative classification networks, CDLNet obtains higher accuracy on multiple datasets (Table 5 , Table 6 ). Ablation experiments demonstrate that removing either SCD or CCD reduces classification accuracy, indicating that spatial and channel complementary decay both contribute to feature regulation (Table 4 , Table 5 ). Visualization results show that CDLNet can enhance target regions, preserve structural details, and suppress irrelevant background responses (Fig.1 , Fig.12 ). Although CDA increases parameters and computation, the accuracy improvement shows a reasonable balance between performance and complexity (Table 3 ). Conclusions CDLNet introduces spatial and channel complementary decay mechanisms into residual networks. By suppressing excessive high-response dominance and preserving negative-response and low-amplitude information, the proposed method improves the completeness and balance of feature representations, alleviates attention drift, and enhances image classification performance. Future work will further optimize the decay strategy and model complexity to improve efficiency and generalization.
Available online , doi: 10.11999/JEIT260365
Abstract:
Objective The increasing demand for text classification in specialized domains, such as finance, healthcare, and social media, is often hindered by the scarcity of labeled data. Low-resource or data-scarce scenarios significantly limit the effectiveness of conventional supervised learning and pre-trained language models, such as BERT, due to insufficient task-specific supervision. Existing methods that rely on single-view weak supervision or clustering often generate noisy pseudo-labels and fail to fully exploit the discriminative potential of hard-to-classify samples. Therefore, it is essential to develop a robust training method that leverages unlabeled data, mitigates noise sensitivity, and enhances semantic representation in low-label conditions. The primary objective of this research is to propose a text classification training framework specifically designed for data-scarce scenarios. Methods To address these challenges, a Heterogeneous Multi-View Semantic Fusion (HMVSF) training framework is introduced. HMVSF consists of three main stages: (1) Dual Clustering Consistency Screening (DCCS): Text samples are first represented in the word-frequency view using TF-IDF vectors. Two heterogeneous clustering algorithms, sIB and an improved K-means initialized via a genetic algorithm, are applied to the same feature space. Samples that are consistently assigned to intersecting clusters are selected as high-confidence pseudo-labeled instances. These samples are subsequently used for a lightweight supervised update of the shared encoder, ensuring reliable initialization for subsequent training. (2) Latent Dirichlet Allocation (LDA) Guided Hard-Sample Retraining: The remaining hard examples, which are not selected in the first stage, typically exhibit low discriminability in the word-frequency view but carry global semantic and topic-level information. LDA is employed to model the topic distribution of these samples. The optimal number of topics is adaptively determined by a combination of semantic coherence and statistical fitting criteria. Hard and soft pseudo-labels are generated based on the document-topic distribution, and the model undergoes targeted retraining using a combined cross-entropy and symmetric KL divergence loss to integrate topic-level semantic cues. (3) Downstream Fine-Tuning: Finally, the encoder is fine-tuned on a small labeled set by replacing temporary heads with a task-specific classification head. Standard cross-entropy is employed as the objective. The two-stage intermediate training is independent of the pre-training initialization, allowing HMVSF to be applied to both general-purpose BERT and masked language model (MLM)-based initializations, making it widely applicable. Results and Discussions The effectiveness of HMVSF is evaluated on six publicly available datasets, including DBpedia, AGNews, ISEAR, SMS Spam, Subjectivity, and Polarity, which cover both topic-based and non-topic-based text classification tasks. Two initialization routes are tested: Route A (BERT/BERTIT:CLUST) and Route B (BERTIT:MLM/BERTIT:MLM+CLUST). Under extremely low label budgets (e.g., 64 labeled samples), CCS-BERTIT:CLUST and CCS-BERTIT:MLM+CLUST outperform baseline models and single-stage variants across datasets (Figs.2 –3 , Tables 2 –3 ). For instance, in Route A, CCS-BERTIT:CLUST achieves up to 9% accuracy improvement on DBpedia and ISEAR compared with BERTIT:CLUST, while error rates decrease by 10–54% (Table 2 ). Normalized Mutual Information (NMI) and intra-cluster Euclidean distance analyses confirm that the CCS selection produces stable and compact clusters, ensuring reliable pseudo-labels (Tables 4 –5 ). Ablation experiments show that the two-stage design combining CCS and LDA is superior to single-stage variants, demonstrating complementary advantages: CCS provides high-confidence core samples, and LDA captures latent semantic structures of hard examples (Table 7 ). Parameter sensitivity analysis reveals that the model is robust under a wide range of cluster numbers and LDA weight coefficients, achieving stable performance when the LDA weighting coefficient is set to 0.3 (Figs.4 –5 ). HMVSF also maintains competitive performance as labeled data increase. Computational experiments on an NVIDIA A40 GPU demonstrate that the two-stage intermediate training introduces acceptable additional overhead while enhancing representation quality and overall accuracy. Conclusions This study presents a novel heterogeneous multi-view semantic fusion training method (HMVSF) for data-scarce text classification. By integrating cluster-consistent sample selection in the word-frequency view and LDA-guided hard-sample retraining, HMVSF effectively leverages unlabeled data, mitigates pseudo-label noise, and enriches semantic representations. Extensive experiments across multiple datasets confirm that HMVSF significantly improves classification accuracy under low-label conditions while remaining robust to pre-training initialization. The framework is generalizable to various pre-trained language models and represents an effective and practical solution for text classification in low-resource scenarios.
Available online , doi: 10.11999/JEIT260877
Abstract:
Objective High-altitude Electromagnetic Pulse (HEMP) has large area, strong fields. It is difficult for HEMP's damage of critical infrastructures which are crucial to country's electromagnetic security. Especially under E1 environment, the waveform distribution characteristic will have direct influence on the effectiveness of protection design. But for cases which need a lot of waveforms calculation like system level effects simulation and protection scheme optimization etc., traditional calculation method of waveform parameter in coverage region is time consuming, which can not realize the real-time calculation of massive ground waveforms on the field distribution. The rapidly and accurately calculation of other parameters of the waveform function is still an issue for the electromagnetic environment field, limiting the useful application of computed fast HEMP environment results for a particular type of explosion when using these results in another calculation. This approach solves a problem that conventional computations cannot be linked into software to perform on-line computation, which provides the basis for the simulation calculation as well as HEMP damage evaluation, enables a fast comparison between different large scale scenarios, enabling us to calculate waveform functions for hundreds of different cases within a few minutes and thus constitutes an efficient tool for rapid analyses as part of the HEMP robust design and evaluation. Methods In this paper, an approach based on Bayesian optimization for deciding how many nodes should be included within each hidden layer of a multilayer feed forward NN is proposed. Reducing searching times for hidden layer node and ensure model precision, in fast solution modeling of HEMP waveform function. And a neural network is firstly utilized for abstracting all of the physical calculation process as a function and builds up a multilayer feedforward neural network model between input and output. However, in such networks, the choice of hidden layer node count affects network performance. In order to achieve both high prediction accuracy and low model complexity, Bayesian optimization finds the best solution in few iterations. While building the surrogate model, an initial data set is formed, and a gaussian process is utilized for the surrogate model which represents the distribution of the objective function. Updating the acquisition function, the best fit parameters, and For the HEMP waveform function, models for Emax, \begin{document}$ \alpha $\end{document} \begin{document}$ \beta $\end{document} Results and Discussions The above method uses both numerical calculation and Bayesian optimization based neural network algorithm, in order to build up an artificial intelligence model of predicting the ground HEMP waveform parameters, covering arbitrary height of bursts, yield of gammas, position in a certain range. Bayesian optimization neural network method reduces searching times on the number of nodes in hidden layers and improve the predicting precision. In Bayesian optimization process, the searching space of hidden layer as [8, 50] is defined to avoid model over fitting, and gives a best network structure of quite small number of hidden layer nodes. To control algorithm training time, the maximum number of Bayesian evaluations is set to 20, approximately half the search space. Through Bayesian optimization, the optimal hidden layer node counts for Emax, \begin{document}$ \alpha $\end{document} \begin{document}$ \beta $\end{document} Conclusions This paper proposes a Bayesian optimization based multilayer feedforward neural network method to model the simulation computation process of HEMP ground waveforms, enabling rapid calculation of standard waveform functions for all points within the ground field distribution of HEMP over a certain range. The method uses Bayesian optimization to optimize the number of hidden layer nodes in the multilayer feedforward neural network, reducing search iterations and improving model prediction accuracy. Compared with five other artificial intelligence methods, the FBEMP method performs best in terms of all metrics. By combining the neural network with numerical derivation, all waveform functions within the field distribution coverage area can be calculated for different burst heights, gamma yields, longitudes and latitudes. This approach lowers the order of operation count from O(n5) in conventional numerical computation to O(n2), and reduces the calculation time of waveform functions for a given field distribution from hour scale to second scale, with errors on each parameter less than 3.12%. It realizes real time calculation of waveform function in HEMP field distribution, and had become applied to large-scale real-time HEMP waveform function calculations, solving the longstanding technical challenge of time-consuming HEMP environment computations that previously prevented real-time calculation, thereby providing an online computational environmental foundation for digital simulation and assessment in HEMP experiments.
Available online , doi: 10.11999/JEIT260436
Abstract:
Objective Water is essential for all organisms and ecological systems, and the composition and content of dissolved metal elements, especially copper (Cu), sodium (Na), and potassium (K), are crucial for maintaining ecological balance and biological health. Cu is an essential human trace element that forms enzymes with functional proteins, participating in antioxidant, energy supply, and immune processes. However, excessive Cu in water—from industrial wastewater, feed additives, pipeline corrosion, and electronic waste leakage—harms human health, causing vomiting, hypotension, jaundice, and hemolytic anemia. Aquatic organisms and plants, lacking effective detoxification systems, are more vulnerable to Cu pollution, which damages roots, induces oxidative stress, and inhibits growth. Na and K are vital for nerve conduction, muscle movement, and fluid balance, but excessive intake endangers those with hypertension or renal/cardiac insufficiency, and their imbalance in irrigation water causes soil salinization. Conventional detection methods (ICP-MS/AES, AAS, AFS) have excellent sensitivity but are limited by large size, complex pretreatment, high cost, and professional operation. Atmospheric pressure glow discharge (APGD) shows potential for miniaturization, but existing APGD-based technologies (PN-APGD, SCGD) require expensive equipment or complex pretreatment. Thus, a highly sensitive, rapid detector for on-site real-time detection of Cu, Na, K without additional pretreatment or driving equipment is urgently needed. Methods A Ultrasonic Nebulization Glow Discharge (UNGD) detector was designed, consisting of an ultrasonic nebulization unit, a plasma excitation unit, and a spectral signal collection unit. The ultrasonic nebulization unit adopted a self-designed centrifuge tube-based diversion chamber with a detachable microporous atomizing sheet, argon inlet/outlet, and a space-constrained transfer tube to form a short gas path, reducing aerosol loss. The plasma excitation unit used 180° coaxial tungsten needle electrodes (cathode/anode) with a double-layer fixing sleeve, clamped on a 3D platform for precise spacing adjustment, powered by a high-voltage DC power supply with a 20 kΩ ballast resistor. The spectral unit included an optical fiber probe and a three-channel AvaSpec spectrometer for high-resolution signal collection. Performance was evaluated using Cu/Na/K mixed solutions: single-factor experiments optimized parameters; characteristic spectra for qualitative analysis; recovery rates for anti-interference assessment; standard curves, LOD, RSD, and CRM detection for quantitative verification; and comparison with similar technologies. Results and Discussions Qualitative analysis of a water sample (Cu: 5.28 mg/L, Na: 5.23 mg/L, K: 3.89 mg/L) showed OH (281.1～309.0 nm) and N2 (315.0～406.0 nm) molecular bands, with obvious Cu (324.7 nm), Na (589.0 nm), and K (766.5 nm) characteristic peaks (Fig. 2 ); 324.7 nm was selected as Cu’s analytical line. Parameter optimization determined optimal conditions: discharge current 38 mA (Fig. 3 ), argon flow rate 0.5 L/min (Fig. 4 ), electrode spacing 1 mm (Fig. 5 ), sampling distance 22 mm (Fig. 6 ). Under these conditions, Cu, Na, K showed good linearity (R2: 0.9942 , 0.9968 , 0.9973 ), with LODs of 141.29 μg/L, 9.54 μg/L, 12.05 μg/L, and RSDs of 6.8%, 6.1%, 5.5% (n=11) (Table 1 ). Anti-interference tests showed 90%～110% recovery rates with 500 mg/L interfering cations (Fig. 7 ). CRM detection showed 89%～110% recovery rates, consistent with standard values (Table 2 ). Conclusions The UNGD detector achieves accurate quantitative analysis of Cu, Na, K in liquids without additional pretreatment or driving equipment. Its detachable atomizing sheet and short gas path improve sampling efficiency, while coaxial electrodes concentrate excitation energy. With excellent sensitivity, precision, and anti-interference ability, it provides reliable technical support for on-site real-time monitoring of water metal elements and has potential for extending to other metal detections, contributing to water ecological protection and biological health.
Available online , doi: 10.11999/JEIT260651
Abstract:
Objective Spoof surface plasmon polaritons (SSPPs) leaky-wave antennas feature low-profile configuration and inherent frequency-scanning capability, making them promising for modern radar, communication, and intelligent sensing systems. However, strong nonlinear coupling among geometric parameters makes traditional optimization computationally costly, as full-wave simulations require thousands of evaluations and often converge to suboptimal local solutions due to landscape complexity. The leakage dynamics in SSPPs—slow-wave propagation, spatial harmonic coupling, and leakage rate distribution—adds complexity beyond conventional designs. To address this, we propose a patch-sinusoidally modulated SSPPs leaky-wave antenna and a machine learning framework integrating a random forest surrogate with particle swarm optimization (PSO) for efficient high-dimensional global optimization with reduced cost. Methods Unlike conventional groove-depth modulation, our antenna maintains uniform groove depth and a complete metal ground. Patch arrays with sinusoidally varying widths are loaded on both sides of the transmission line, with envelope functions \begin{document}$ Y=A\sin (Tx) $\end{document} \begin{document}$ Y=A\sin (Tx+\pi ) $\end{document} 1200 , depth: 12–24, min samples per split: 2–5). The trained RF surrogate is then embedded into PSO (40 particles, 120 iterations, inertia 0.72, c1=c2=1.5) with a weighted fitness function (G:1.5, η:0.7, SLL:0.55, S11:0.25, S21:0.7, θscan:0.18). Results and Discussions RF achieves the highest average R2 of 0.9554 across six outputs, outperforming GPR (0.9489 ), SVR (0.9212 ), and MLP (0.8952 ). For key radiation indicators, RF attains gain MAE of 0.032 dBi, SLL MAE of 0.168 dB, and efficiency MAE of 0.015. Scatter plots of predicted versus simulated values cluster tightly around the diagonal, and residual histograms show means near zero with no systematic bias, confirming excellent prediction accuracy and generalization. After RF-PSO optimization, full-wave simulation confirms substantial improvements: gain rises from 13.84 dBi to 14.52 dBi, SLL drops from –17 dB to –19 dB,peak total efficiency increases from 84.9% to 92.9%, S11 improves from –23.26 dB to –27.92 dB (4.66 dB), and scanning range expands from 57.2° to 61.5°. The scanning angle versus frequency curve exhibits good linearity across the operating band, and the two-dimensional far-field patterns show improved symmetry. The decrease in S21 (from –3.76 dB to –5.50 dB) together with the gain/efficiency increase indicates that more energy is effectively converted into radiation rather than being dissipated or reflected. Sensitivity analysis with ±2% perturbations (50 samples) shows all coefficients of variation (CV) below 1.3%: gain CV 0.21% (<±0.1 dBi), SLL CV 1.26% (±0.4 dB), efficiency CV 0.88% (±0.01), S11 CV 0.97%, S21 CV 0.91%, scan CV 0.37%. These fluctuations are far smaller than optimization gains, confirming excellent robustness under typical fabrication tolerances. Comparison with recent leaky-wave antennas (both SSPP-based and SIW) demonstrates superior SLL (−19 dB), competitive efficiency (89% vs. 94.95% and90% in prior SSPP work), and scanning range (61.5°) outperforming most single-port SSPP antennas (e.g., 20°, 16°, 33°, 13°). The number of full-wave simulations is reduced by approximately 90% (450 training + 1 validation vs. 4,800 simulations for conventional PSO). Conclusions This paper proposes a patch-sinusoidally modulated SSPPs leaky-wave antenna and an RF-assisted PSO framework for synergistic optimization in nine dimensions. The RF surrogate achieves an average R2 of 0.9554. The optimized antenna shows significantly improved performance across all metrics: gain by 0.68 dBi, SLL by 2 dB, efficiency by 8%, S11 by 4.66 dB, and scanning range by 4.29°, while maintaining compact dimensions. Sensitivity analysis confirms robustness under typical fabrication tolerances. The proposed methodology reduces the number of full-wave simulations by approximately 90%. This work marks a methodological advancement by introducing machine learning surrogate modeling into SSPPs leaky-wave antenna design for efficient high-dimensional optimization. Future work includes fabrication, experimental validation, extension to millimeter-wave bands, and reconfigurable antenna designs.
Available online , doi: 10.11999/JEIT260867
Abstract:
Objective Massive Machine-Type Communications (mMTC) constitutes a fundamental pillar of 5G and emerging 6G wireless networks, dedicated to supporting massive connectivity for the Internet of Things (IoT). In grant-free random access scenarios, sporadic and uncoordinated transmissions by massive terminals inevitably induce severe packet collisions under heavy traffic loads. Traditional scheduled access protocols become inefficient due to prohibitive signaling overhead. Consequently, grant-free slotted ALOHA protocols based on Successive Interference Cancellation (SIC)—such as Contention Resolution Diversity Slotted ALOHA (CRDSA), Irregular Repetition Slotted ALOHA (IRSA), and Coded Slotted ALOHA (CSA)—have garnered widespread attention. However, these classical schemes fundamentally rely on the presence of collision-free (degree-1) slots to trigger and sustain the iterative graph-peeling decoding process. Under practical constraints of finite frame lengths and heavy traffic, collision-free slots are drastically depleted, causing severe decoding stalling and throughput degradation. Furthermore, pure SIC mechanisms are intrinsically vulnerable to error propagation. Although Polar Slotted ALOHA (PSA) introduces polarization transforms across time slots to enhance packet recovery, its collision resolution remains constrained by the initial SIC condition. To resolve these challenges, a joint decoding algorithm combining Polar Slotted ALOHA with Physical-Layer Network Coding (PSA-PNC) is proposed over the Slot Erasure Channel (SEC). The objective is to transform destructive multi-user packet collisions into algebraically solvable linear network coding equations, thereby eliminating the strict reliance on collision-free slots and significantly elevating concurrent multi-user detection capability and throughput performance under heavy traffic loads. Methods A joint physical-layer and MAC-layer random access framework is established over the Slot Erasure Channel (Fig. 1 ). At the transmitter side, active users independently select transmission time slots based on an irregular degree distribution polynomial without inter-user coordination or channel collision feedback. The sender remains blind to multi-user collision patterns in the channel. At the base station receiver, the superimposed signals across slots are equivalently modeled as a sparse global input matrix over a binary finite field, followed by packet-level polar encoding. To resolve dense collisions without relying on clean slots, a closed-loop iterative receiver architecture is developed (Fig. 2 ). In each iteration, channel observation sequences are initially processed by a packet-level Successive Cancellation (pSC) or packet-level Successive Cancellation List (pSCL) decoder to extract reliable equivalent combined packets from information slots. Instead of being discarded, the collided slots corresponding to these reliable packets are utilized to construct a local sparse linear Network Coding (NC) equation system. A Generalized Matrix Inversion (GMI) criterion is subsequently executed to analyze the column-rank characteristics of the access pattern matrix and achieve global algebraic multi-user decoupling. Solvable user packets are directly recovered without requiring full-rank matrix conditions or degree-1 slots. The algebraically decoupled user packets are then utilized as prior information to reconstruct physical-layer codewords and subtracted from the observation buffer via iterative SIC, continuously reducing the dimensionality of the unresolved collision space. High-reliability packets output by the pSCL decoding paths are leveraged in the iterative loop to effectively suppress error propagation and guarantee the linear independence of residual equations. Furthermore, the polarization evolution process of equivalent multi-user packets over the SEC is theoretically proved to be equivalent to that over a scalar Binary Erasure Channel (BEC), enabling rigorous calculation of frame error rate bounds via Bhattacharyya parameters (Fig. 3 ). Results and Discussions Extensive theoretical analyses and Monte Carlo simulations are conducted to evaluate the performance of the proposed PSA-PNC scheme over the Slot Erasure Channel. The theoretical polarization bounds of packet-level polar decoding are verified under various slot erasure probabilities (\begin{document}$ \epsilon \in \left\{0.1,0.2,0.3\right\} $\end{document} Fig. 3 ). In terms of system throughput, simulation results demonstrate that for a frame length of \begin{document}$ N=1024 $\end{document} \begin{document}$ L=8 $\end{document} \begin{document}$ G\approx 0.90 $\end{document} Fig. 4(a) ). In the low-load region, all evaluated schemes exhibit near-identical linear throughput growth due to the abundance of collision-free slots (Fig. 4(a) ). When the slot erasure rate increases to \begin{document}$ \epsilon =0.35 $\end{document} Fig. 4(a) ). For a short frame length of \begin{document}$ N=64 $\end{document} Fig. 4(b) ). In terms of transmission reliability, under a target Packet Loss Rate (PLR) of \begin{document}$ {10}^{-2} $\end{document} \begin{document}$ N=1024 $\end{document} \begin{document}$ G\approx 0.75 $\end{document} \begin{document}$ G\approx 0.84 $\end{document} Fig. 5(a) ). Furthermore, steeper waterfall regions and significantly lower error floors are consistently maintained across various frame lengths from \begin{document}$ N=64 $\end{document} \begin{document}$ N=1024 $\end{document} Fig. 5(b) ). Conclusions A joint decoding scheme combining Polar Slotted ALOHA with Physical-Layer Network Coding (PSA-PNC) is established to resolve the severe decoding stalling problem in grant-free random access. By constructing a closed-loop iterative receiver integrating packet-level polar decoding, GMI-based algebraic equation solving, and iterative SIC cancellation, destructive multi-user collisions are converted into solvable linear equations. The dependence on collision-free slots is effectively eliminated, and the supportable load threshold, peak normalized throughput, and packet recovery reliability are substantially enhanced under heavy traffic loads. Future research will be directed toward non-ideal multipath fading channels, such as Rayleigh fading, and the design of low-complexity sparse receiver architectures for practical massive access implementations.
Available online , doi: 10.11999/JEIT260416
Abstract:
Objective Environment-driven channel prediction requires a structured representation that connects physical objects in a propagation environment with the resulting multipath topology. However, conventional data-driven methods generally treat environmental information as unstructured global features and have difficulty representing the interactions among the transmitter (Tx), receiver (Rx), and surrounding scatterers. This study investigates whether an environmental graph can provide an effective intermediate representation for identifying effective scatterers and predicting candidate propagation paths. Methods An environmental graph is constructed by representing the Tx, Rx, and scatterers as graph nodes. Spatial distance and visibility between nodes are encoded as edge features to describe their geometric relationships. A ScatterGNN based on the Edge-aware Graph Isomorphism Network (EGIN) is developed to extract structural features and identify effective scatterers involved in signal propagation. Candidate single- and two-bounce paths are subsequently generated from the detected scatterers. A path-ranking network, PathRankingNet, is then designed to estimate the validity scores of candidate paths and rank them using a Listwise Ranking Loss. The proposed framework is evaluated in a controlled indoor Industrial Internet of Things scenario generated using Wireless InSite. The scenario covers an area of 150 m × 63 m × 22 m and contains 3,381 Rx sampling locations. Results and Discussions For effective scatterer detection, the proposed method achieves an average precision of 0.957 9, while 77.1% of the test samples achieve complete detection of all effective scatterers. For candidate propagation-path prediction, the model converges stably after approximately 30 training epochs. Precision@3 ranges from 0.60 to 0.64, Recall@3 ranges from 0.42 to 0.45, and Hit@3 ranges from 0.90 to 0.92. These results indicate that the environmental graph preserves useful structural information related to propagation-path topology. In particular, the model retains at least one ground-truth propagation path among the three highest-ranked candidates for more than 90% of the test samples. Conclusions The proposed framework provides a graph-based approach for transforming environmental geometry into structured representations of effective scatterers and candidate propagation paths. Rather than replacing ray tracing or channel measurements, the method is intended to reduce the candidate search space before detailed path-parameter calculation or channel reconstruction. The current results demonstrate its feasibility within a single simulated environment and primarily reflect its ability to approximate the propagation-path topology labels generated by Wireless InSite. Further validation using independent environments, measured channel data, and path-level parameters such as power, delay, and complex gain is required before its cross-scenario generalization and practical applicability can be established.
Available online , doi: 10.11999/JEIT260595
Abstract:
Objective Building change detection from bi-temporal high-resolution remote sensing images is important for urban planning, land resource management, illegal construction monitoring, and disaster damage assessment. Existing fully supervised change detection methods generally achieve high detection accuracy but require pixel-level annotations. However, obtaining pixel-level labels for large-scale remote sensing images is labor-intensive and time-consuming, which limits their application to large-scale monitoring scenarios. Image-level weakly supervised change detection reduces annotation costs by using only image-level labels indicating whether an image pair contains changes. However, the lack of spatial supervision makes accurate localization of changed regions difficult. Existing weakly supervised methods generally rely on Class Activation Maps (CAMs) to generate pseudo labels. CAMs tend to highlight only the most discriminative regions, resulting in incomplete coverage of changed areas or background noise. Vision-language models provide semantic priors for weakly supervised learning. However, directly applying Contrastive Language-Image Pre-training (CLIP) to change detection remains challenging. Fixed text prompts are difficult to adapt to the difference semantics of bi-temporal images, and the original CLIP objective mainly focuses on global image-text alignment rather than local pixel-level localization. To address these problems, a Difference-aware Adaptive Prompt Learning and Dense Alignment method for weakly supervised building change detection, termed DAPL-CD, is proposed. Methods The proposed framework introduces CLIP-based cross-modal semantic knowledge into image-level weakly supervised building change detection. For a pair of bi-temporal remote sensing images, a shared CLIP visual encoder is first used to extract visual representations from the two temporal images. The local visual features are fused along the channel dimension to obtain bi-temporal difference features containing semantic information related to changed and unchanged regions. Based on the difference characteristics of building change detection, a difference-aware adaptive prompt learning strategy is designed. Instead of using manually designed fixed text templates, learnable context vectors are inserted into the text prompts while preserving category-related semantic words. The resulting foreground and background text embeddings are used as foreground and background text prototypes to provide adaptive semantic guidance for change localization. Furthermore, a pixel-text dense alignment mechanism is introduced to extend CLIP’s global image-text alignment capability to local feature matching. The initial CAM generated by the classification branch is used to obtain preliminary foreground and background regions. Visual-text positive and negative sample pairs are then constructed between local difference features and the foreground and background text prototypes. An InfoNCE-based dense alignment loss is used to pull matched visual and textual features closer and push mismatched features apart. Finally, the classification and segmentation branches are jointly optimized using the classification loss, global alignment loss, dense alignment loss, and segmentation loss, with the segmentation loss introduced only after the quality of the generated CAMs has stabilized. Results and Discussions Experiments are conducted on WHU-CD and LEVIR-CD, two public benchmark datasets for building change detection. Only image-level labels are used during training, whereas pixel-level annotations are used only for evaluation. Overall Accuracy (OA), F1-score, and Intersection over Union (IoU) are adopted as the main evaluation metrics. Because changed buildings usually occupy a small proportion of remote sensing images, OA can be strongly affected by the dominant unchanged background pixels. Therefore, F1-score and IoU are emphasized for evaluating the detection quality of changed regions. Quantitative comparisons show that DAPL-CD achieves an OA of 94.7%, an F1-score of 82.8%, and an IoU of 70.6% on WHU-CD, and an OA of 92.3%, an F1-score of 68.0%, and an IoU of 51.5% on LEVIR-CD. The method achieves the best F1-score and IoU among the compared weakly supervised change detection methods (Table 1 ). Visual comparisons further show that the proposed method produces more complete responses for large-scale building changes and more continuous predictions for small and scattered changed buildings (Figs. 3 and 4 ). Ablation experiments verify the effectiveness of difference-aware adaptive prompt learning and pixel-text dense alignment. The baseline model using fixed text prompts without foreground or background alignment achieves an F1-score of 63.2% and an IoU of 46.2%. Introducing both foreground and background alignment increases these metrics to 66.3% and 49.6%, respectively, indicating that dense semantic matching between local visual features and text prototypes improves the discrimination of changed regions. After difference-aware adaptive prompt learning is incorporated, the F1-score and IoU increase to 68.0% and 51.5%, respectively, indicating that learnable context vectors reduce the semantic mismatch between fixed text descriptions and bi-temporal difference features. Under the adaptive-prompt setting, foreground alignment alone achieves an F1-score of 64.4% and an IoU of 47.5%, whereas background alignment alone achieves 65.3% and 48.5%, respectively. Combining the two alignment branches yields the best F1-score and IoU, indicating that foreground and background semantic constraints provide complementary guidance for change localization (Table 2 ). The CAM results further show that pixel-text dense alignment produces stronger and more complete responses over actual changed regions while suppressing irrelevant background activations (Fig. 5 ). Conclusions A weakly supervised building change detection framework based on difference-aware adaptive prompt learning and pixel-text dense alignment is proposed. By introducing CLIP-based cross-modal semantic priors, the proposed method converts text-level semantic knowledge into local change localization capability. The difference-aware adaptive prompt learning strategy improves the representation of change-related semantic descriptions, whereas the pixel-text dense alignment mechanism establishes direct correspondence between local difference features and foreground and background text prototypes. Experimental results on WHU-CD and LEVIR-CD demonstrate that DAPL-CD achieves high performance under image-level supervision and improves the completeness and accuracy of changed building localization. The proposed framework provides an effective approach for reducing annotation requirements in large-scale remote sensing change detection. Future research will focus on improving pseudo-label reliability, reducing dependence on large-scale pre-trained models, and extending the method to multi-temporal and multi-spectral remote sensing data.
Available online , doi: 10.11999/JEIT260609
Abstract:
Objective Architecture parameters of commercial off-the-shelf Field-Programmable Gate Arrays (FPGAs) are fixed by vendors and reused across products. Embedded FPGAs (eFPGAs), in contrast, allow architects to select architecture parameters according to specific application requirements. The LUT input count K, the number of LUTs per cluster N, the interconnect topology, and the types of heterogeneous tiles can therefore be configured for the target application. Architecture design space exploration thus becomes an engineering task in which tens to hundreds of architectures may need to be generated and evaluated before a suitable configuration is selected. Existing architecture description practices rely largely on manually written architecture files and batch scripts. After each parameter change, shared fields across multiple backend toolchains must be updated and aligned manually, making large-scale architecture exploration difficult to support. To address this problem, a parametric architecture description framework is proposed for unified description across multiple backend toolchains. Architecture parameters are organized into three layers according to their independence: process invariants, coupled parameters, and independent parameters. Architecture descriptions for different backends are automatically derived from the same source object through independent derivation functions. The framework currently supports VPR, OpenFPGA, and Yosys and has been extended to COFFE. Its operation is validated across the complete toolchain. Based on the framework, a parameter-sweep design space exploration method is developed, and an open Quality of Results (QoR) dataset covering five application domains and 64 benchmark circuits, including homogeneous and heterogeneous architectures, is released as a public benchmark. Methods HorizonArch, the proposed parametric architecture description framework, organizes architecture parameters into three layers according to parameter independence (Table 1 ). L0 contains process invariants that are fixed once the technology node is determined. L1 contains coupled parameters, including K, N, tier, segment length, switch block type, and channel connectivity, for which a single parameter change can trigger updates across multiple fields and backend architecture descriptions. L2 contains independent parameters that can be specified separately. Architecture construction is formalized by an operator B that maps a parameter vector p to a complete architecture object (Fig. 4 ). Five formal rules are imposed: parameter completeness (R1), fragment independence (R2), type compatibility (R3), explicit coupling (R4), and static checkability (R5). Each backend architecture description is then derived by an independent derivation function from the same source object. Thus, adding a new backend requires only an additional view rather than modifications throughout the existing description structure. Field-level validation and cross-field validation are performed when the architecture object is loaded, before any backend tool is invoked. The class structure (Fig. 3 ) divides the architecture description into synthesis, circuit, and layout views, with each semantic element declared only once. Three extension levels are defined: G1 adds a black-box model, G2 extends the value set of an existing coupled parameter, and G3 adds a new coupled parameter together with its constrained value set. Based on this framework, a parameter-sweep design space exploration method is developed to scan the (K, N) parameter grid and heterogeneous tile configurations. Each configuration is evaluated using three QoR metrics: area, Critical-Path Delay (CPD), and Area-Delay Product (ADP). Results and Discussions End-to-end validation shows that a single source description consistently generates architecture descriptions for VPR, OpenFPGA, and Yosys. COFFE is connected and verified at the interface layer, including SPICE simulation startup (Table 4 , Table 5 ). The G1, G2, and G3 extension experiments pass all cross-field checks. The design space exploration results show different preferences among area, CPD, and ADP across the (K, N) parameter space (Figs. 5 and 6 ). Area favors smaller K values, with K=4 and N=4 providing favorable area and ADP performance for a large proportion of circuits. CPD, in contrast, favors larger K and N values, with optimal configurations concentrated near (K, N)=(8, 10) and (7, 10). Across the twenty (K, N) configurations, the relative-range distribution shows that parameter selection has a much greater effect on area and ADP than on CPD (Table 6 ). The mean relative ranges are 50.9% for CPD, 675.1% for area, and 516.3% for ADP. A comparison of default configurations (Table 7 ) shows that K=4 and N=4 achieves the minimum ADP for 67.9% of the circuits and has an average ADP deviation of 5.8%, although its average CPD deviation reaches 47.9%. In contrast, K=8 and N=10 reduces the average CPD deviation to 5.9%, with 30.8% of the circuits achieving the CPD optimum, but increases the average area and ADP deviations to 673.2% and 487.0%, respectively. A random-forest cross-domain surrogate achieves a top-5 accuracy of approximately 65%. Therefore, parameter sweeping remains necessary when strict design targets are imposed. Conclusions HorizonArch, a parametric architecture description framework for eFPGA exploration, is developed and validated. The framework generates VPR, OpenFPGA, and Yosys backend architecture descriptions from a single source object and provides an extensible interface for COFFE. The parameter-sweep exploration shows that area and CPD favor opposite regions of the (K, N) parameter space. Therefore, eFPGA architecture parameters should be selected according to explicit design targets rather than fixed default values. An open QoR dataset covering five application domains and 64 benchmark circuits is also released as a reusable benchmark for eFPGA architecture design space exploration. Future work will complete the COFFE SPICE topology-rewriting component, refit the routing-area coefficient using measured data, and explore more efficient design space exploration strategies.
Available online , doi: 10.11999/JEIT260442
Abstract:
Objective Accurate grasp detection from language instructions is essential for service robots to achieve natural human-robot interaction. Existing methods primarily rely on large-scale data-driven training or hierarchical feature fusion to align visual perception with textual instructions. However, they generally overlook the strong coupling between target objects and background clutter in low-level visual features, leading to degraded compositional generalization under cross-view and unseen-scene conditions. To address this limitation, a dual-view cross-scene grasp detection and object localization dataset is constructed to systematically evaluate and improve the compositional generalization of existing models. Based on this benchmark, a Simultaneous Grasp detection and object Localization Network (SGL-Net) is proposed to jointly predict object locations and optimal grasp poses. The proposed framework enables service robots to manipulate objects according to natural language instructions in real-world dynamic environments, providing technical support for embodied intelligence. Methods The proposed SGL-Net is illustrated in Fig. 1. First, a Cross-modal Global Context Modulation Module (CGCMM) is proposed to exploit semantic priors from language instructions for adaptive viewpoint correction and background suppression during the early stage of visual feature extraction. Second, a Word-Pixel Cross-modal Alignment Module (WPCAM) is designed to achieve fine-grained semantic decoupling through a flattening-based cross-modal attention mechanism, thereby improving semantic understanding in complex dynamic scenes. Finally, a unified decoder jointly predicts object locations and optimal grasp poses from the fused multimodal features. Results and Discussions Extensive quantitative and qualitative experiments are conducted on the reconstructed dual-view cross-scene dataset containing bottom-view and top-view scenes and on a real-world robotic grasping platform. Comparative results demonstrate that SGL-Net consistently outperforms mainstream CNN-based and CLIP-based methods in both grasp detection and object localization (Tables 2 and 3 ). Ablation studies further verify the effectiveness of CGCMM and WPCAM in improving fine-grained semantic alignment and semantic decoupling (Tables 4 and 5 ). Furthermore, qualitative results (Figs. 4 ～6 ) and real-world robotic experiments (Fig. 7 ) demonstrate that SGL-Net can be reliably deployed in complex physical environments. Overall, the proposed network exhibits strong generalization capability and excellent potential for practical robotic applications. Conclusions To improve cross-view and cross-scene generalization in language-conditioned robotic grasp detection, this paper constructs a dedicated validation dataset and proposes SGL-Net, which jointly performs grasp detection and object localization. By integrating CGCMM and WPCAM, the proposed network accurately localizes instruction-specified objects and predicts optimal grasp poses. Experimental results obtained on multiple benchmark scenarios and a real-world robotic platform demonstrate the superior performance and practical applicability of the proposed method. Future work will focus on integrating Large Multimodal Models (LMMs) and adapting the proposed framework through fine-tuning to further improve zero-shot robotic grasp detection.
Available online , doi: 10.11999/JEIT260420
Abstract:
Objective Terahertz (THz) Ultra-Massive Multiple-Input Multiple-Output (UM-MIMO) systems are expected to support future high-capacity wireless communications. However, accurate channel estimation remains challenging under hybrid near-/far-field propagation and Array-of-SubArrays (AoSA) architectures, where limited Radio-Frequency (RF) chains, low Signal-to-Noise Ratio (SNR), noise uncertainty, and structural perturbations degrade compressed observations. Existing compressed sensing, Bayesian inference, and deep unfolding methods generally rely on fixed statistical assumptions, which limit their cross-SNR generalization under varying noise conditions and statistical mismatches. To address these limitations, this paper proposes a Noise-Conditioned Fixed-Point Network (FPN-NCAS) for robust THz UM-MIMO channel estimation. The proposed method aims to improve estimation accuracy, cross-SNR generalization, robustness, and iterative stability by incorporating noise-aware nonlinear recovery. Methods FPN-NCAS is developed within an Orthogonal Approximate Message Passing (OAMP)-based fixed-point unfolding framework. A coarse noise power estimate is obtained from repeated pilot differences and injected into the nonlinear recovery module as an explicit conditioning variable. After each linear update, the vector-domain estimate is reshaped into an AoSA-aligned tensor to exploit subarray-level structural priors. The nonlinear recovery chain consists of three components. Token-Gate performs lightweight subarray-level reliability pre-calibration to suppress unreliable responses under low-SNR and structurally inconsistent conditions. Block Shrink serves as the core noise-conditioned block-sparse proximal operator, in which a smooth dual-threshold mechanism balances strong denoising at low SNR with structural preservation at medium-to-high SNR. G-HMTD(Gated Hybrid Multi-scale Transformer Denoiser) further refines residual errors by combining Local Multi-scale Enhancement and Global Context Modeling. Bridge relaxation and nonlinear residual scaling are also incorporated to improve inter-stage stability. Results and Discussions Simulation results demonstrate that FPN-NCAS consistently outperforms LS, OAMP, ISTA-Net+, FPN-OAMP, and FPN-OTFN over the 0～20 dB SNR range (Fig. 6 ). At SNR = 0 dB, FPN-NCAS achieves Normalized Mean Square Error (NMSE) gains of approximately 3.0 dB and 1.9 dB over FPN-OAMP and FPN-OTFN, respectively. At SNR = 20 dB, the gains increase to approximately 4.5 dB and 2.6 dB (Fig. 6(a) ). The convergence curves show that FPN-NCAS reaches a stable plateau after approximately three layers at SNR = 5 dB and five layers at SNR = 15 dB, demonstrating stable fixed-point iterative behavior (Fig. 6(b) and Fig. 6(c) ). Analysis of repeated pilots shows that four repeated pilot pairs introduce only 3.13% additional pilot overhead while reducing the relative standard deviation of the coarse noise power estimate to 25.00%. Under moderate noise power mismatch, NMSE degradation remains within 0.3 dB. The proposed method also maintains strong robustness under colored Gaussian noise, impulsive noise, near-/far-field distribution shifts, variation in the number of propagation paths, AoSA subarray shuffling, and RF amplitude/phase mismatch (Fig. 7 and Fig. 8 ). Ablation studies indicate that Block Shrink provides the largest performance gain, whereas Token-Gate and G-HMTD further improve performance through structural calibration and residual refinement (Fig. 9 ). Conclusions This paper proposes FPN-NCAS for noise-conditioned fixed-point channel estimation in THz UM-MIMO systems. By integrating repeated-pilot-based noise conditioning, AoSA-aware feature reshaping, Token-Gate calibration, Block Shrink recovery, and G-HMTD refinement, the proposed method improves NMSE performance, robustness, and iterative stability under different SNR conditions and non-ideal scenarios. The improved performance is achieved at the cost of higher inference complexity. Future work will focus on lightweight implementations and extensions to wideband, multi-user, and hardware-impaired THz communication systems.
Available online , doi: 10.11999/JEIT260512
Abstract:
Objective In port operation videos, the grab is a continuously moving target, and accurate trajectory extraction is essential for operation monitoring, equipment coordination, and collision warning. However, complex backgrounds, scale variations, partial occlusion, and boundary degradation often reduce the stability of target region segmentation, leading to centroid deviation, trajectory jitter, missed detections, and trajectory discontinuity. To address these challenges, a Spatial-Temporal Collaborative Optimization Method is proposed for stable and continuous grab trajectory extraction. While maintaining high inference speed, the proposed method improves both trajectory extraction accuracy and trajectory stability, providing a practical solution for stable perception of continuously moving targets in port industrial video scenarios. Methods Built on YOLOv8-seg, the proposed framework integrates Spatial Representation Enhancement (SRE) and Temporal CONSistency constraint (TCONS). First, CBAM, BiFPN-lite, and shallow feature aggregation are incorporated to improve target-background separability, enhance multi-scale feature representation, and preserve boundary details. TCONS is then imposed on prototype features through global average pooling, a cache-based pairing mechanism, and a weighted Charbonnier loss to suppress the temporal accumulation of local errors. In addition, a stage-wise training strategy with warm-up epochs and a joint optimization objective is adopted to ensure stable convergence. Results and Discussions Experiments are conducted on DAVIS2016, SegTrackV2, and a real portal crane grab dataset to evaluate the proposed method in terms of segmentation performance, trajectory stability, and occlusion robustness. The proposed method achieves the best segmentation performance on the real portal crane grab dataset, with J and F scores of 90.05% and 98.56%, respectively. It also improves performance on DAVIS2016 while maintaining comparable performance with slight gains on SegTrackV2 (Tables 1 and 2 , Fig. 2 ). In terms of trajectory stability, compared with YOLOv8-seg, the proposed method reduces MAE and RMSE by approximately 55.3% and 52.6%, respectively, and decreases the miss rate to 0.56% (Table 5 ). It also produces a more concentrated trajectory error distribution and a smaller fluctuation range (Fig. 3 ). Occlusion robustness experiments further demonstrate that, under different occlusion ratios, the proposed method maintains good region integrity and continuous target extraction capability, reducing the maximum number of consecutive missed frames from 52 to 47 (Table 6 , Figs. 4 and 5 ). Ablation studies verify the complementary effects of SRE and TCONS, whereas parameter analysis shows that a TCONS weight of 0.3 provides the best balance between segmentation quality and trajectory stability (Tables 7 and 8 ). Conclusions A Spatial-Temporal Collaborative Optimization Method is proposed to address the challenge of stable grab trajectory extraction in port operation videos. Experimental results demonstrate that the proposed method achieves high segmentation accuracy and stable trajectory extraction on DAVIS2016 and the real portal crane grab dataset, while maintaining comparable segmentation performance on SegTrackV2. It also exhibits strong continuous target extraction capability under occlusion without significantly sacrificing inference speed. Since the current study is limited to fixed crane viewpoints, future work will focus on cross-scene generalization and long-term continuous perception under more complex operating conditions to further improve the robustness and applicability of the proposed method in real-world environments.
Available online , doi: 10.11999/JEIT260426
Abstract:
Objective Deepfake detectors often exhibit performance degradation when applied to unseen manipulation methods, cross-dataset distribution shifts, diffusion-generated faces, and common image degradations. In forensic applications, the generation process, data source, and post-processing conditions of a questioned sample are usually unknown. Existing methods often represent the fake class with a single feature center, although different generation methods, data sources, and processing conditions produce heterogeneous patterns. Transferable forensic cues may therefore be mixed with mode-specific artifacts, limiting generalization to unknown domains. To address this problem, a hierarchical prototype learning framework with shared subspace factorization (HPL-SF) is proposed. Within-class diversity is first used to observe latent fake modes, followed by estimation of a low-rank structure shared across these modes and sample-level exploitation of the shared structure during inference. Methods A pretrained DINOv2 Vision Transformer (ViT-L/14) is used as the backbone. Its original parameters are frozen, and Low-Rank Adaptation (LoRA) modules are inserted into the query and value mappings of the self-attention layers for parameter-efficient training. All features and prototypes are L2-normalized, and cosine similarity is used for prototype assignment, binary classification, and test-time representation adaptation. HPL-SF consists of three successive stages (Fig. 1). First, one real prototype and multiple mode-specific fake prototypes are maintained in the normalized feature space. Each training sample is softly assigned to the fake prototypes according to its cosine similarity to each prototype, and the weighted prototypes are aggregated to obtain a sample-adaptive fake representation. The response distribution thus provides an observable representation of latent within-class modes without using forgery-source labels. A binary classification loss, a sample–prototype contrastive loss, and a prototype diversity loss are jointly optimized to separate real and fake samples, improve sample–prototype alignment, and prevent prototype collapse. Second, the normalized mode-specific fake prototypes are arranged into a prototype matrix. Singular Value Decomposition (SVD) is applied to this matrix, and the largest gap between consecutive singular values is used to determine the dimension of the shared fake subspace. Each fake prototype is then decomposed into a projection onto the shared fake subspace and a mode-specific residual. Only the shared projections are aggregated to update the shared fake prototype, whereas the residuals retain mode-specific information and are excluded from this update. The real prototype and shared fake prototype are updated using Exponential Moving Average (EMA). Gradients are not propagated through subspace construction, dimension selection, or semantic prototype updates. Third, the backbone, LoRA modules, prototypes, and subspace basis are fixed during inference. A test feature is compared with the real and shared fake prototypes, and their relative responses determine a sample-specific adaptation weight. The original feature is blended with its projection onto the shared fake subspace and then L2-normalized. The resulting feature is classified according to its cosine similarities to the two semantic prototypes. Thus, the estimated shared structure is exploited on a sample-by-sample basis without updating model parameters during testing. Results and Discussions The method is evaluated under cross-dataset, cross-forgery-type, diffusion-forgery, repeated-run, image-degradation, ablation, and mechanism-analysis protocols. Across seven unseen datasets, HPL-SF achieves the highest area under the receiver operating characteristic curve (AUC) on all seven datasets, with an average AUC of 91.67%, which is 2.10 percentage points higher than the second-highest average (Table 1). When trained on FaceForensics++ and evaluated on the Diffusion Facial Forgery dataset, HPL-SF achieves the highest AUC on the text-to-image, image-to-image, face-swapping, and face-editing subsets, with an average AUC of 84.39% (Table 2). Across four cross-forgery-type settings on FaceForensics++, the average accuracy and AUC reach 85.29% and 92.41%, respectively, both ranking first among the compared methods. When DeepFakes under high-quality compression is held out for testing, HPL-SF trails the best-performing method by only 0.91 percentage points in accuracy and 0.39 percentage points in AUC (Table 3). Repeated experiments with multiple random seeds yield the highest mean values for all four aggregate metrics, with standard deviations ranging from 0.38 to 0.62 percentage points (Table 4). Under five severity levels of compression, blur, and noise on the Deepfake Detection Challenge Preview dataset, HPL-SF achieves the best or joint-best AUC at most severity levels and remains relatively stable under moderate and severe degradations (Fig. 2). All six ablated variants perform worse than the complete HPL-SF model on the four unseen test sets, indicating complementary contributions from mode-specific prototype learning, sample–prototype contrastive loss, prototype diversity loss, shared subspace factorization, and test-time representation adaptation (Fig. 3). Performance generally increases as the number of mode-specific fake prototypes increases and plateaus when the number reaches 10. The largest spectral gap occurs between the fifth and sixth singular values, yielding a shared subspace dimension of 5. With this setting, HPL-SF achieves a diffusion-forgery AUC of 84.39% ± 0.62%, compared with 79.68% for direct mean aggregation and 81.56% ± 0.80% without test-time representation adaptation (Figs. 4(a)–4(c)). Linear-probe analysis further shows that the shared component provides stronger discrimination in unknown domains and weaker domain-identifying capability than the mode-specific residual, supporting the separation of transferable and domain-related cues (Fig. 4(d)). The t-distributed stochastic neighbor embedding (t-SNE) visualization shows clearer real–fake separation and closer distributions of same-class samples from different data sources (Fig. 5). Prototype allocation statistics show differentiated responses across fake types, with prototype usage remaining close to the uniform baseline. The adaptation weights are higher for fake samples, particularly for correctly classified fake samples, whereas misclassified samples show responses closer to the balanced-response line (Fig. 6). False positives are mainly associated with low resolution, compression, filters, or occlusion, whereas false negatives usually contain high-quality forgeries or weak forgery traces (Fig. 7). Conclusions HPL-SF organizes generalizable deepfake detection into three successive stages: observation of within-class diversity, estimation of shared structure, and sample-level exploitation of that structure. Experimental results show that separating shared projections from mode-specific residuals provides transferable discriminative cues under dataset shifts, unseen forgery types, diffusion-generated manipulations, and common image degradations. The framework requires no parameter updates during testing and adaptively exploits the shared structure according to each sample’s relative responses to the real and shared fake prototypes. Errors remain for degraded real images whose artifacts resemble forgery traces and for high-quality forgeries with weak traces. Future work will extend the framework to temporal prototypes for video, multimodal forensic evidence, and adaptive discrimination in broader open-world settings.
Available online , doi: 10.11999/JEIT260462
Abstract:
Objective Active Reconfigurable Intelligent Surface (RIS) enhances wireless communication performance by integrating radio frequency amplifiers to mitigate the multiplicative fading inherent to passive RIS. However, amplification noise and additional power consumption are introduced. Furthermore, high-precision digital phase control at the base station incurs considerable communication overhead. Employing low-precision phase shifters is therefore an effective approach for practical RIS deployment. Therefore, characterizing the Energy Efficiency (EE) performance of active RIS-assisted communication systems and quantifying the effect of finite-bit phase quantization errors on EE are essential for system design and practical implementation. To this end, a discrete phase-shifted active RIS-assisted communication system over Rayleigh fading channels is investigated. The EE loss caused by phase quantization errors is analyzed, approximate optimal solutions for the power allocation factor and the number of RIS elements that maximize EE are derived, and the relationship between RIS EE and user EE is established, providing theoretical guidance for the practical deployment of active RIS. Methods Based on the law of large numbers and Taylor series expansion, closed-form expressions for the user EE loss and its approximation are derived. The effects of system parameters on EE are investigated by expressing EE as explicit univariate functions. Ferrari’s method and the Lambert W function are then employed to derive approximate optimal solutions for the power allocation factor and the number of RIS elements that maximize EE. Finally, the relationship between RIS EE and user EE is established using the law of large numbers and the Lambert W function. Results and Discussions User EE is expressed as a function of six parameters: the number of quantization bit (\begin{document}$ k $\end{document} \begin{document}$ \beta $\end{document} \begin{document}$ N $\end{document} \begin{document}$ {P}_{\text{t}} $\end{document} \begin{document}$ \sigma _{\text{r}}^{2} $\end{document} \begin{document}$ \sigma _{\text{u}}^{2} $\end{document} \begin{document}$ k $\end{document} \begin{document}$ k $\end{document} Fig. 3 ). Therefore, 3-bit to 4-bit discrete phase shifters achieve performance close to that of continuous phase shifters. Second, user EE exhibits a unimodal dependence on both \begin{document}$ \beta $\end{document} \begin{document}$ N $\end{document} \begin{document}$ \beta $\end{document} Fig. 4 ), whereas the approximate and exact optimal solutions for \begin{document}$ N $\end{document} Fig. 5 ), demonstrating the high accuracy of the proposed approximations. Third, user EE exhibits a unimodal trend as \begin{document}$ {P}_{\text{t}} $\end{document} \begin{document}$ {P}_{\text{t}} $\end{document} Fig. 6 ). In addition, user EE decreases as both \begin{document}$ \sigma _{\text{r}}^{2} $\end{document} \begin{document}$ \sigma _{\text{u}}^{2} $\end{document} Fig. 7 ). Finally, user EE first increases and then decreases sharply to zero as RIS EE increases. The signal-to-noise ratio at the RIS is identified as the key factor governing the relationship between RIS EE and user EE (Fig. 8 ). Conclusions The EE performance of active RIS-assisted wireless networks employing discrete phase shifters over Rayleigh fading channels is investigated. First, closed-form expressions are derived for the user EE in the lossless case, the lossy case, and the approximate-loss case. Simulation results demonstrate that 3-bit to 4-bit discrete phase shifters closely approach the performance of continuous phase shifters. Next, explicit functions describing the effects of key system parameters on user EE are established. Ferrari’s method and the Lambert W function are employed to derive approximate optimal solutions for the power allocation factor and the number of RIS elements that maximize EE, and both exhibit negligible errors relative to the exact solutions. Finally, the relationship between RIS EE and user EE is established, demonstrating that user EE initially increases and subsequently decreases to zero as RIS EE increases.
Available online , doi: 10.11999/JEIT260423
Abstract:
Objective Signals acquired from industrial systems often exhibit irregular sampling due to sensor instability, intermittent operation, communication dropout, and multi-source asynchronous acquisition. This irregularity violates the uniform sampling assumption of most signal analysis methods and challenges anomaly detection. Interpolation and resampling may distort the underlying temporal dynamics, while continuous-time models based on neural ordinary differential equations as well as time-aware Transformers suffer from high training costs and strong dependence on large training sets, making them impractical under limited training resources. Model space learning offers an alternative by fitting each signal with a dynamic model and analyzing the fitted models instead of the raw signals. However, existing model space methods for irregular sampling rely on fixed reservoir configurations and lack adaptive optimization of the model space. This paper aims to develop an anomaly detection framework for irregularly sampled signals that is simultaneously robust to non-uniform intervals and efficient to train. Methods An adaptive model space learning framework based on the Adaptive Decay Reservoir Network (ADRN) is proposed (Fig. 1 ). ADRN extends the echo state network by introducing an exponential decay mechanism derived from a linear ordinary differential equation (Fig. 2 ). Between consecutive observations, the hidden state decays according to a learnable decay rate over the actual elapsed time, so that the state update naturally adapts to non-uniform sampling intervals without numerical ordinary differential equation solvers. Each new observation then updates the decayed state through a nonlinear activation. Ridge regression with a closed-form solution then fits a readout model mapping hidden states to the original signal, and the fitted readout weights serve as a compact fixed-dimensional representation regardless of signal length. The model space is further optimized by two complementary losses. A time-interval-weighted reconstruction loss assigns higher weights to larger intervals, preventing densely sampled segments from dominating the optimization and improving fitting quality under non-uniform sampling. A separability loss inspired by Fisher discriminant analysis acts through a learnable projection matrix to minimize intra-class scatter and maximize inter-class separation in the projected model space. The two losses are combined into a joint objective that simultaneously updates the reservoir parameters, the decay rate, and the projection matrix, and gradients propagate through the differentiable closed-form ridge regression to enable end-to-end optimization. A downstream classifier, namely a support vector machine on CWRU and SU and a random forest on the higher-dimensional TEP model space, performs the final detection. The echo state property of ADRN is formally established, and the resulting spectral-norm condition is more relaxed than the classical one, allowing richer reservoir dynamics. Results and Discussions Experiments cover the CWRU bearing dataset (five subsets, 50% missing rate), the SU gearbox dataset (30%, 50%, and 70% missing rates), and the Tennessee Eastman Process (TEP) chemical dataset (19 classes, three missing rates), with only 200 labeled signals per subset for training on CWRU and SU. The proposed method achieves the highest accuracy in 7 of the 8 CWRU and SU settings (Table 1 ), with accuracies ranging from 87.3% to 93.8% on CWRU. On SU, it maintains 94.4% accuracy even at a 70% missing rate, and the fluctuation across missing rates is only 2.8%, in contrast to 18.7% for ODE-RNN, while Neural CDE drops from 86.8% to 62.6%. Ablation studies confirm the contribution of each component (Table 1 , Table 2 ). Removing the exponential decay reduces accuracy by up to 28.4 percentage points, and the interval weighting and the separability loss contribute complementary gains of 2.0 and 3.1 percentage points on SU at the 70% missing rate. t-SNE visualization shows that the optimized model space exhibits compact and clearly separated classes (Fig. 3 ). Training on a CWRU dataset completes in about 50 seconds, over two orders of magnitude faster than neural ordinary differential equation methods, which require 3000 to 7000 seconds (Table 3 ). Hyperparameter analysis indicates that a loss balance coefficient between 0.1 and 0.3 performs well and that a small reservoir suffices (Table 4 ). On TEP, a non-rotating-machinery industrial object whose faults manifest as changes in process dynamics, the proposed method attains the highest accuracy among all compared methods at every missing rate, with the largest margin of 7.0 percentage points at the highest missing rate (Table 5 ). Conclusions The ADRN based model space learning framework provides an effective solution for anomaly detection on irregularly sampled signals. The exponential decay mechanism enables interval-aware state updates without numerical ordinary differential equation solvers, ridge regression yields efficient closed-form readout fitting, and the joint optimization of the reconstruction and separability losses produces a model space with both high fitting quality and strong discriminative structure. The framework requires only a small reservoir of 10 to 50 dimensions and completes training within one minute, making it well suited to scenarios with limited training resources and irregular sampling. Future work includes adaptive reservoir sizing, extension to multivariate joint modeling, and validation in further domains such as structural health monitoring and biomedical or meteorological time series.
Available online , doi: 10.11999/JEIT260600
Abstract:
Objective Max transition violations in digital integrated circuit physical design may increase gate delay, reduce timing margin, and introduce additional power-consumption and signal-integrity risks. With technology scaling and the increasing complexity of high-performance SoC and CPU designs, long interconnects and heavy effective loads make such violations more prominent in post-routing optimization and ECO stages. Conventional repair methods usually rely on manual analysis or global heuristics, such as total-wire-length-ratio-based buffer insertion, which may lead to low repair efficiency, inaccurate repair targeting, and redundant buffer insertion in multi-load nets. Although these approaches can alleviate some violations in simple cases, they often fail to accurately identify the real violating branches in multi-load nets. As a result, repair targeting becomes insufficient and redundant buffer insertion is likely to occur. To address these limitations, a path-level buffer insertion method is proposed for max transition violation repair in multi-load nets. Methods The proposed method first reconstructs the physical topology of the target net from routing information extracted from the Design Exchange Format (DEF) file. Routing endpoints, turning points, and via connection points are abstracted as physical nodes with coordinate and metal-layer attributes, and a weighted undirected graph is established to preserve branch structures and cross-layer connectivity (Fig. 3 ). To reduce the influence of small coordinate deviations, a spatial-tolerance-based node merging strategy is introduced during graph construction. Since the violation coordinates reported by timing analysis tools may not lie exactly on valid routed segments, a vector-projection-based coordinate snapping strategy is then adopted to align logical violation coordinates with the actual physical topology (Fig. 4 ). After endpoint binding, the actual physical propagation path from the driver to each violating load is recovered by Dijkstra shortest-path search. Based on the Elmore-model intuition that inserting a buffer near the midpoint of a long interconnect can effectively segment the distributed RC load, the midpoint of each recovered path is selected as the initial candidate insertion point. The exact insertion coordinate is obtained by accumulating segment lengths along the path and interpolating on the segment where half of the total path length is reached. To improve robustness, the initial candidate point is further expanded into an effective candidate interval with a spatial tolerance factor. For multi-load nets, different violating branches may share long common physical segments. Therefore, an interval-intersection-based shared-buffer optimization strategy is introduced to merge overlapping candidate intervals into shared insertion regions, thereby reducing redundant buffer insertion (Fig. 5 -Fig. 7 ). Results and Discussions Experiments are conducted on five designs, including CPU, SAS, RAID, PCIe, and HBA, implemented in the UMC 28 nm process. Synopsys IC Compiler II is used for physical implementation, and PrimeTime is used for timing analysis. A violation-margin threshold of -8 ps is adopted, and only paths below this threshold are included in the repair and evaluation. The complete ECO flow retains the existing repair method for single-load nets and applies the proposed path-level method to multi-load nets. To illustrate the repair mechanism, a representative multi-load violating net is selected for detailed analysis. The net contains one driver and fourteen loads, among which thirteen violating load paths share a long common routed trunk and exhibit max transition violation margins ranging from -75.2 ps to -38.3 ps. If these paths are repaired independently, thirteen nearby buffer insertion demands are generated on the shared trunk. After interval-intersection-based merging, one shared buffer is sufficient to repair all thirteen violating paths jointly (Fig. 8 ). This case study indicates that path recovery improves repair targeting, while shared-buffer optimization improves resource efficiency. Across the five designs, the complete ECO flow reduces the total number of max transition violations from 10,616 to 108, corresponding to an overall repair rate of 98.98% (Table 1 ). The CPU, SAS, and RAID designs are completely repaired, while only 17 and 91 violations remain in PCIe and HBA, respectively. Among the 8,216 multi-load violating paths, 8,166 are successfully repaired, corresponding to a multi-load violation repair rate of 99.39%. If these successfully repaired paths were handled independently, 8,166 buffer insertions would be required. After shared-buffer optimization, only 1,548 buffers are inserted, giving an overall compression ratio of 81.04%. The setup worst negative slack, total negative slack, and number of violating paths remain generally stable before and after repair, with only minor fluctuations observed in individual designs (Table 2 ). These results show that the repair flow does not cause obvious degradation in setup timing quality. Further analysis indicates that the residual violations in PCIe and HBA mainly occur when candidate buffer locations fall inside standard-cell placement blockages around hard macros or IP cores. Metal routing is allowed through these regions, but buffers cannot be legally placed, revealing a current limitation in the physical-feasibility handling of candidate insertion locations. Conclusions A path-level buffer insertion method is proposed for max transition violation repair in multi-load nets. By combining physical topology reconstruction, violation-coordinate snapping, shortest-path-based path recovery, midpoint-guided candidate generation, and interval-intersection-based shared-buffer optimization, the proposed method improves repair targeting and reduces redundant buffer insertion. Experimental results on five designs show that the complete ECO flow reduces the number of max transition violations from 10,616 to 108, achieving an overall repair rate of 98.98%. Among the 8,216 multi-load violating paths, 8,166 are successfully repaired, corresponding to a repair rate of 99.39%. For these successfully repaired paths, shared-buffer optimization reduces the number of buffer insertion demands from 8,166 to 1,548, corresponding to a compression ratio of 81.04%, without causing obvious degradation in setup timing quality. The remaining violations are mainly associated with standard-cell placement blockages around hard macros or IP cores, indicating that the physical-feasibility handling of candidate insertion locations should be further improved.
Available online , doi: 10.11999/JEIT260631
Abstract:
Objective Compared with post-silicon side-channel security analysis, pre-silicon side-channel security verification during the design phase enables the earlier identification of potential side-channel security vulnerabilities in cryptographic core designs, thereby effectively reducing the cost and time of post-silicon remediation. However, most existing pre-silicon side-channel security assessment approaches rely on data-driven statistical analysis or artificial intelligence techniques and require complex calculations on large amounts of data to mitigate the impact of insufficient coverage on the assessment results. In addition, existing methods typically adopt independent modeling strategies for different types of side channels, lacking a unified side-channel security modeling approach. A cryptographic side-channel security modeling and formal verification method is proposed, supporting unified and automated modeling of different types of side channels by constructing a side-channel security model. The method can identify potential timing side-channel, power side-channel and fault injection vulnerabilities in cryptographic core designs, and analyze the effectiveness of side-channel countermeasures based on side-channel security property checking. Methods The proposed cryptographic side-channel security modeling and formal verification method includes side-channel security model construction, side-channel security property extraction, and side-channel security verification. The side-channel security model uses information flow analysis to characterize timing side-channel leakage, power side-channel leakage, and fault propagation behavior in cryptographic core designs, providing an effective mathematical model for side-channel security verification. Specifically, the side-channel security model utilizes changes in signal labels to analyze information flows during the encryption process by assigning a label to a signal bit and defining label propagation rules. Side-channel security properties formally describe the behavioral characteristics of side-channel leakage, including timing properties, power properties, and fault properties, providing theoretical support for side-channel security verification. Side-channel security verification uses the extracted security properties as verification constraints and employs formal verification tools to identify potential timing side-channel, power side-channel, and fault injection vulnerabilities in cryptographic core designs. Furthermore, the method can analyze the effectiveness of masking and fault injection countermeasures against side-channel vulnerabilities. Results and Discussions The proposed side-channel security verification method utilizes formal verification techniques to accurately identify potential side-channel security vulnerabilities in various block cipher core designs, and evaluate the effectiveness of side-channel countermeasures based on side-channel security property constraints. The timing side-channel verification results demonstrate that the proposed method can accurately identify timing side-channel security vulnerabilities in AES, SM4, LED, PRESENT, and IDEA core designs within 20s (Table 2 , Fig. 7 ). No timing side-channel vulnerabilities are identified in the other cryptographic core designs, except for IDEA, which exhibits timing side-channel vulnerabilities caused by modular multiplication operations. The power side-channel security verification results show that formal checks based on controllability property, key–power distinguishability coupling property and key–power nonlinear coupling property can accurately identify target modules with potential power side-channel security vulnerabilities in AES, SM4, LED and PRESENT within 1 minute (Table 3 ). The key expansion module in cryptographic core designs does not cause key leakage through key-dependent power consumption, as it fails to satisfy the controllability property. In addition, the experimental results indicate that the masking protection in the RSM core design can prevent the correct key from being distinguished through random masking (Fig. 8 ). The fault injection security verification results for three AES core designs with infective countermeasures demonstrate that the proposed method can analyze the effectiveness of fault infection countermeasures. The results show that the infection countermeasure requires not only altering the fault propagation path but also disrupting the algebraic relationships among faults (Table 4 , Fig. 10 ). Conclusions This paper proposes a cryptographic side-channel security modeling and formal verification method to address the lack of formal mathematical models and the limited completeness of existing data-driven side-channel security assessment methods. The proposed method first achieves unified modeling of timing leakage, power leakage, and fault propagation behaviors from the perspective of information flow analysis. Based on the constructed side-channel security model, side-channel security properties are extracted to formally characterize the behavioral features of side-channel information leakage and propagation during the encryption process. Potential side-channel security vulnerabilities in cryptographic core designs are then identified through formal checking using the extracted security properties as constraints. The proposed method provides an effective solution for the unified modeling and formal security verification of different types of side channels. Experimental results obtained from the side-channel security verification of various block cryptographic core designs demonstrate that: (1) the proposed method can uniformly model timing side-channel leakage, power side-channel leakage, and fault propagation behaviors in cryptographic core designs; (2) the proposed method can accurately identify timing side-channel vulnerabilities, power side-channel vulnerabilities, and fault injection vulnerabilities in cryptographic core designs, including AES, SM4, IDEA, LED and PRESENT; (3) the proposed method can analyze the effectiveness of masking and fault infection countermeasures. However, this study only qualitatively identifies side-channel security vulnerabilities in cryptographic core designs; pre-silicon quantitative assessment of side-channel leakage should be investigated in future work.
Available online , doi: 10.11999/JEIT260464
Abstract:
Objective With the rapid proliferation of Unmanned Aerial Vehicles (UAVs) and the escalating demand for airspace security, radio frequency (RF) fingerprint recognition has emerged as a pivotal technology for identifying non-cooperative UAVs. However, existing methods grapple with significant challenges, including poor robustness in low signal-to-noise ratio (SNR) environments and prohibitive computational complexity, which severely hinder their deployment on resource-constrained tactical edge devices. To address these critical limitations, this paper proposes a novel lightweight dual-stream convolutional network tailored for UAV RF recognition. This network is designed to extract static spectral texture features and dynamic temporal gradient features in parallel, complemented by a meticulously crafted lightweight feature fusion strategy. Methods The proposed network architecture is ingeniously designed to process RF signals. The input signal undergoes a Short-Time Fourier Transform (STFT) to generate a two-dimensional spectrogram, which serves as the primary input. The network is bifurcated into two parallel streams: a static stream and a dynamic stream. The static stream is engineered to capture the inherent static spectral patterns and energy distributions within the STFT spectrogram. It comprises a series of stacked convolutional blocks, each integrating convolutional layers, batch normalization, and ReLU activation functions, followed by max-pooling layers to progressively downsample the feature maps and increase the channel depth. Conversely, the dynamic stream is dedicated to enhancing feature discriminability, particularly in low-SNR scenarios. It begins by computing the temporal gradient of the input spectrogram, effectively suppressing static background noise and accentuating dynamic signal variations. This gradient map is then processed by a symmetric set of convolutional blocks, mirroring the structure of the static stream. To maintain model efficiency, an element-wise addition fusion strategy is employed to integrate the features from both streams, ensuring a balance between feature complementarity and computational overhead. The fused features are subsequently fed into a classification head, consisting of an adaptive average pooling layer, dropout layers for regularization, and fully connected layers to produce the final classification output. Extensive experiments are conducted on the publicly available DroneRF dataset, encompassing ablation studies to dissect the contribution of each component, comparative analyses of various fusion strategies, and rigorous evaluations of the model’s lightweight characteristics. Results and Discussions The experimental results unequivocally demonstrate the efficacy of the proposed method. The dual-stream network achieves a remarkable 95.65% accuracy on the test set, representing a substantial 5 percentage point improvement over the best-performing single-stream network. A critical analysis reveals that the temporal gradient operation contributes significantly to this enhancement by improving the average SNR by 1 dB, thereby bolstering feature discriminability in challenging low-SNR environments. Furthermore, the model’s lightweight design is a standout feature, with a mere 0.58 million parameters, making it eminently suitable for deployment on tactical edge devices. Ablation studies and feature visualization analyses provide compelling evidence for the complementary nature of static and dynamic features. The static stream adeptly captures broad spectral contours, while the dynamic stream focuses on fine-grained temporal variations. The element-wise addition fusion strategy proves superior, outperforming other approaches like feature concatenation and attention-based fusion in terms of both performance and computational efficiency, thereby validating the rationale behind the lightweight design. Conclusions This paper presents a comprehensive solution to the challenges of UAV RF recognition in complex environments by proposing a lightweight dual-stream convolutional network. The method effectively enhances recognition accuracy and robustness through the synergistic combination of dual-stream feature extraction and the SNR-enhancing properties of temporal gradient features, all while maintaining a lightweight architecture suitable for edge deployment. The proposed approach offers a significant advancement in the field, providing a robust and efficient solution for UAV identification. Future research endeavors will focus on further enhancing the model’s adaptability to complex electromagnetic environments, incorporating the effects of sensor noise, and extending the framework to multi-UAV cooperative scenarios.
Available online , doi: 10.11999/JEIT260499
Abstract:
Objective The rapid development of the low-altitude economy is increasing the use of unmanned aerial vehicle (UAV) swarms in emergency communication, urban logistics, reconnaissance, and low-altitude network coverage. These applications require reliable spectrum awareness to support cooperative communication, dynamic spectrum access, and interference avoidance. However, low-altitude electromagnetic environments often contain low-SNR signals, multipath propagation, non-cooperative interference, and multiple coexisting transmissions. Conventional energy, cyclostationary-feature, and matched-filter detectors are sensitive to noise uncertainty, computational cost, or prior waveform knowledge. Learning-based methods can improve robustness, but many rely on power spectral density (PSD) or short-time Fourier transform (STFT) representations. These representations may weaken phase information or require costly two-dimensional time-frequency preprocessing. Existing methods also focus mainly on spectrum occupancy and provide limited information about overlapping transmissions. This study therefore develops a low-latency joint sensing method that preserves magnitude and phase information while estimating spectrum occupancy and interference overlap for each frequency bin. The method is intended for local spectrum sensing at UAV nodes under resource and latency constraints. Methods The proposed RadioSEUnet pipeline contains two stages: magnitude-phase feature construction and joint spectrum-state estimation. First, each complex baseband in-phase/quadrature (I/Q) sequence is multiplied by a Hann window and transformed using a one-dimensional fast Fourier transform (FFT). The resulting complex spectrum is decomposed into logarithmic magnitude and phase components. The two components are normalized separately and stacked as a two-channel feature tensor. Binary labels indicate spectrum occupancy and multi-signal overlap at each frequency bin. RadioSEUnet adopts a U-shaped encoder-bottleneck-decoder architecture with four encoder stages containing 64, 128, 256, and 512 channels. Each RadioSEBlock combines a complex-parameterized convolution, squeeze-and-excitation channel attention, and a residual connection. The convolution couples the magnitude and phase feature streams through constrained cross-channel operations. Two prediction heads convert the shared representation into a spectrum-occupancy probability mask and an interference-overlap probability mask. The model is optimized using an equally weighted sum of two binary cross-entropy losses. Training uses AdamW, cosine-annealing learning-rate scheduling, early stopping, a batch size of 64, and at most 200 epochs. The complete data collection contains 72,000 complex I/Q records, including 48,000 simulated records and 24,000 measured records. The simulated subset covers Wi-Fi, BLE, ZigBee, LoRa, QPSK/16QAM, FM, and AM signals. Signal-to-noise ratios range from –15 dB to 10 dB under additive white Gaussian noise and Rayleigh fading. The measured subset was collected using a USRP N310 in the 2.4–2.5 GHz ISM band at 100 MS/s over a 1 ms observation interval. The controlled quantitative evaluation uses an 8:1:1 split of the simulated subset. A separate simulated-to-measured protocol is defined in the main text to examine cross-domain generalization. RadioSEUnet is compared with six PSD- or STFT-based baselines under matched data splits and hardware conditions. Performance is measured using intersection over union (IoU), precision, recall, preprocessing time, inference time, and total sensing latency. Results and Discussions The SNR-dependent quantitative results reported here are obtained using the controlled simulated-data protocol. At -15 dB, RadioSEUnet achieves an IoU of 0.768 and a recall of 0.846 for spectrum occupancy detection. Compared with the second-best STFT-RADN baseline, these values correspond to absolute improvements of 0.186 and 0.166, respectively. For interference-overlap detection, RadioSEUnet achieves an IoU of 0.456 and a precision of 0.768 at –15 dB. The corresponding improvements over STFT-RADN are 0.246 and 0.275. The lower IoU for interference-overlap detection indicates that weak overlap boundaries remain difficult to separate from strong-signal sidelobes and background noise. The latency evaluation is conducted on the workstation specified in the main text. Magnitude-phase preprocessing requires 21.04 ms, and network inference requires 4.77 ms, producing a total sensing latency of approximately 25.81 ms. STFT-YOLOv3 requires 98.9 ms under the same hardware setting, so the proposed pipeline is approximately 3.8 times faster in this comparison. Ablation experiments show that magnitude-phase preprocessing, complex-parameterized feature coupling, and channel attention each improve low-SNR sensing performance. Removing the magnitude-phase preprocessing produces the largest degradation. These results indicate that preserving complementary magnitude and phase information is useful for weak-signal and interference-overlap detection. They do not, however, establish performance on airborne hardware or across unreported radio environments. Conclusions RadioSEUnet combines a magnitude-phase representation, constrained cross-channel feature coupling, channel attention, multiscale feature fusion, and dual-head prediction. It jointly estimates spectrum occupancy and interference-overlap states while avoiding two-dimensional STFT preprocessing. Under the controlled simulated-data protocol, the method provides higher point estimates than the six evaluated baselines at low SNR and reduces total sensing latency on the evaluated workstation. The present evidence is limited to the reported signal types, channel models, hardware configuration, and the 2.4–2.5 GHz measurement band. Quantitative simulated-to-measured results, tests on wider bands, additional interference types, repeated trials, and deployment on airborne edge hardware are still required. Future work will therefore focus on cross-domain validation, lightweight deployment, boundary-aware interference modeling, and integration with spectrum resource management for UAV networks.
Available online , doi: 10.11999/JEIT260396
Abstract:
Objective Gallium Oxide Metal-Oxide-Semiconductor Field-Effect Transistor (Ga2O3 MOSFET) is regarded as a promising power device for high-voltage applications, particularly in aerospace and satellite power systems, because of its ultra-wide bandgap and high critical breakdown field. However, the Conventional MOSFET (C-MOSFET) exhibits limited reliability in space radiation environments. Under off-state conditions, the electric field is highly concentrated near the gate edge. Heavy-ion irradiation generates dense electron-hole pairs along the ion track. Driven by the intense electric field, these carriers undergo avalanche multiplication through impact ionization, causing the drain current to increase sharply without recovery and ultimately leading to irreversible Single-Event Burnout (SEB) at relatively low drain bias. This failure mechanism severely limits the application of Ga2O3 MOSFETs in harsh radiation environments. Furthermore, the lack of reliable and efficient p-type doping restricts the implementation of conventional radiation-hardening techniques, including junction termination extension and junction isolation. Therefore, ionization-induced carriers readily accumulate in sensitive regions, increasing susceptibility to Single-Event Effect (SEE). The extremely low thermal conductivity of Ga2O3 further promotes local heat accumulation following heavy-ion irradiation, producing localized hot spots that increase the likelihood of thermal burnout. Existing hardening approaches, including field-plate optimization and dielectric engineering, provide only limited improvement. Moreover, the application of heterojunction structures for radiation hardening has rarely been investigated, and systematic hardening strategies have not yet been established. To address these limitations, this paper proposes a Comb-Shaped Gate Metal-Oxide-Semiconductor Field-Effect Transistor (CSG-MOSFET) incorporating a NiO heterojunction. The proposed structure redistributes the channel electric field, suppresses electric-field crowding at the conventional gate edge, and significantly improves SEB tolerance, providing an effective solution for Ga2O3 power devices operating in harsh radiation environments. Methods Technology Computer-Aided Design (TCAD) simulations are performed to evaluate the electrical characteristics and SEB performance of the proposed CSG-MOSFET in comparison with the C-MOSFET. The simulations incorporate high-field mobility, Shockley-Read-Hall recombination, Auger recombination, impact ionization, and heavy-ion models. Based on the charge-compensation effect of the p-NiO/n-Ga2O3 heterojunction, the proposed structure utilizes the extended depletion region formed at the heterointerface to redistribute the channel electric field. This heterojunction-induced depletion region improves electric-field uniformity and enhances SEB tolerance. Furthermore, the comb-shaped gate columns, operating together with the extended gate field plate, relocate the peak electric field away from the conventional gate edge, suppress local electric-field crowding, and improve device reliability under high-voltage and radiation conditions. Results and Discussions Simulation results demonstrate that the optimized Double Comb-Shaped Gate MOSFET (DCSG-MOSFET) significantly improves radiation hardness compared with the C-MOSFET. The SEB Threshold Voltage (VSEB) increases from 240 V to 2 280 V, while the Breakdown Voltage (BV) increases from 2 000 V to 3 500 V. Meanwhile, the specific on-resistance decreases. Therefore, the Baliga Figure of Merit (BFOM) and the SEB-based figure of merit are substantially improved. The NiO heterojunction and comb-shaped gate columns effectively redistribute the electric field, shifting the peak electric field from the conventional gate edge to the outer gate-column edge and suppressing local electric-field crowding. These improvements substantially enhance the radiation hardness of the device. Conclusions A radiation-hardened DCSG-MOSFET incorporating a NiO heterojunction is proposed and evaluated using TCAD simulations. The optimized structure significantly improves SEB tolerance while maintaining excellent electrical performance. Compared with the C-MOSFET, both VSEB and BV are substantially increased, demonstrating enhanced blocking capability. Charge compensation at the p-NiO/n-Ga2O3 heterojunction forms an extended depletion region that effectively redistributes the channel electric field and suppresses electric-field crowding near the conventional gate edge. Furthermore, the comb-shaped gate columns, operating together with the extended gate field plate, relocate the peak electric field to the outermost gate-column edge, thereby suppressing impact ionization induced by heavy-ion irradiation and effectively mitigating SEB. The reduced specific on-resistance further improves the BFOM and the SEB-based figure of merit. These results demonstrate that the proposed DCSG-MOSFET is a promising candidate for power electronic applications in harsh radiation environments, including aerospace and satellite systems.
Available online , doi: 10.11999/JEIT260401
Abstract:
Objective Modern electronic reconnaissance systems must deinterleave dense and overlapping radar pulse streams in non-cooperative environments. As radar emitters increasingly employ agile waveforms, similar pulse descriptor words, and low-intercept-probability strategies, conventional single-station methods based on carrier frequency, pulse width, and Pulse Repetition Interval (PRI) become less reliable. Multi-station deinterleaving based on Time Difference of Arrival (TDOA) provides a more stable geometric observable, but severe pulse loss still causes sparse cross-station pairing, weak true TDOA peaks, ambiguity-induced spurious peaks, isolated pulses, and fragmented trajectories across time slices. These effects increase false alarms and weaken track continuity. To address these issues, a closed-loop multi-station emitter TDOA deinterleaving method is proposed for severe pulse-loss environments, with Time of Arrival (TOA) sequences used as the core observables. Methods A slice-based framework is developed for continuous reconnaissance. Residual unmatched pulses are carried forward by a sliding window to alleviate cross-slice misalignment. First, candidate pulse pairs satisfying geometric TDOA constraints are generated, and pulse descriptor word constraints on carrier frequency and pulse width are used to remove inconsistent pairs. To reduce the sparsity and binning sensitivity of conventional histograms, multiscale Kernel Density Estimation (KDE) is introduced to reconstruct the TDOA density from sparse candidate differences. Gaussian kernels with different bandwidths are fused, and candidate peaks are adaptively extracted using local statistics and peak widths. Second, a dynamic memory matrix is designed to suppress ambiguity-induced spurious peaks in high pulse repetition frequency scenarios. Since dependent spurious peaks collapse after the dominant peak is extracted and removed, a collapse-rate criterion is defined, and the spurious regions are recorded in a memory mask for subsequent iterations. Third, Dynamic Time Warping (DTW) is used to compare incomplete TOA sequences of isolated pulses with extracted pulse sequences, enabling reassignment of unequal-length and incomplete sequences. Finally, a Kalman-filter-based state-space model tracks multi-baseline TDOA trajectories across successive slices. Predicted and observed TDOA residuals are jointly used for association, so intermittent observations can still be linked to the correct track. In this way, the proposed method forms a closed-loop processing chain that links weak-peak reconstruction, spurious-peak suppression, isolated-pulse reassignment, and trajectory association (Fig. 3 ). Results and Discussions Four simulation scenarios are designed: a high pulse repetition frequency scenario dominated by ambiguity-induced spurious peaks, a parameter-overlapping scenario dominated by isolated pulse reassignment, an ablation scenario for evaluating the memory matrix and DTW modules, and a 10-emitter mixed-regime scenario including fixed PRI, staggered, jittered, frequency-agile, pulse-group frequency-agile, frequency-agile jittered-PRI, linear-sliding, and sinusoidal-sliding PRI signals. In the mixed-regime scenario, the total reconnaissance duration is 1 s and the slice duration is 0.1 s. The environmental pulse loss rate is fixed at 10%, and the receiver-specific loss rate increases from 0% to 40%. Both loss rates are calculated with respect to the initial theoretical number of transmitted pulses; therefore, the total loss rate is their sum, ranging from 10% to 50%. The proposed method is compared with an extended TDOA histogram method under constrained criteria, a cloud-model-based multi-station sorting method, and a Dirichlet Process Mixture Model (DPMM)-based method (Table 5 ). In the high pulse repetition frequency scenario, the proposed method maintains near-zero false alarms by identifying the collapse of dependent spurious peaks and suppressing them through the memory matrix, whereas the comparison methods show severe false alarms (Figs. 4 and 5 ). In the parameter-overlapping scenario, DTW-based reassignment improves isolated-pulse recovery, while the memory matrix suppresses spurious TDOA peaks. Their combination improves extraction reliability and reduces false alarms (Figs. 6 and 7 ). The ablation results verify their complementary roles: at a 50% loss rate, the memory matrix reduces the TDOA false alarm rate from 29.92% to 3.32%, DTW increases pulse extraction accuracy from 66.71% to 91.99%, and the complete method achieves a TDOA detection rate of 99.25% with a false alarm rate of 0.88% (Fig. 8 ). In the 10-emitter mixed-regime scenario, the proposed method achieves a favorable overall trade-off. At an overall pulse loss rate of 50%, its pulse extraction accuracy remains 92.72%, and the TDOA false alarm rate is limited to 7.75%, lower than 35.39%, 35.05%, and 37.58% for the DPMM, cloud-model, and constrained recursive histogram methods, respectively. After cross-slice trajectory association, the mean number of identity switches decreases to 3.83, compared with 10.64, 9.85, and 12.64 for the three comparison methods (Fig. 9 and Table 5 ). Conclusions A closed-loop multi-station emitter TDOA deinterleaving method is proposed for severe pulse-loss environments. By integrating multiscale KDE-based weak peak reconstruction, dynamic memory-matrix-based spurious peak suppression, DTW-based isolated pulse reassignment, and Kalman-filter-based trajectory association, the method addresses the coupled failure mechanisms caused by severe pulse loss. Simulation results demonstrate high extraction accuracy, low TDOA false alarm rates, and strong trajectory continuity in high-loss and mixed-regime scenarios. These results demonstrate the effectiveness of the method under the simulated conditions and indicate its application potential for persistent multi-station passive reconnaissance.
Available online , doi: 10.11999/JEIT260400
Abstract:
Objective Industrial Internet of Things (IIoT) routing forwards and schedules control commands, equipment status information and sensing data to support critical tasks such as collaborative equipment control, safe system operation and environmental monitoring, but the routing process is prone to congestion and resource exhaustion under flooding attacks. Existing intelligent secure routing methods apply reinforcement learning (RL) to optimize next-hop selection based on network topology, but the heterogeneity in queue capacity and link bandwidth of IIoT terminals is often overlooked, leading to load imbalance and local congestion, and limiting performance under high load or malicious traffic attacks. Therefore, we propose a large language model (LLM)-based global situation-aware assisted secure routing method in IIoT against flooding attacks, which applies RL to optimize multi-path selection and achieve load balancing across the network. Methods Based on global security awareness, queue congestion of neighboring nodes, queue capacity, link bandwidth, and service types, the proposed secure routing method applies RL to optimize multi-path selection against flooding attacks. The cloud–edge large model infers global security situational awareness including global load distribution and anomalous traffic distribution based on network topology, node resource occupancy and link state information, and feeds the inference result back to IIoT terminals to construct RL states and evaluate routing policies risks. In addition, a risk-aware function is formulated to quantify the routing disruption potential by integrating end-to-end latency, packet delivery ratio and node vulnerability to attacks. An experience replay buffer that incorporates both reward and risk is constructed, where both factors are considered during routing parameter updates to guide routing policy selection, thereby balancing safe path exploration and optimization efficiency. Results and Discussions Simulations are conducted using 30 industrial nodes under varying configurations, including bandwidths of 5 MHz, 10 MHz, 20 MHz, and queue capacities ranging from 100 to 500 packets. The global security situational awareness is inferred by the Qwen3.5-27B-AWQ-4bit, which is deployed on a cloud–edge server equipped with dual 24 GB RTX 4090 GPUs. In each time slot, each terminal sends 5 packets of 2 KB each to the industrial gateway. A flooding attacker injects \begin{document}$ y\in \{10,20,30\} $\end{document} Conclusions This paper proposes an LLM-based global situation-aware assisted secure routing method for IIoT against flooding attacks, which applies RL to optimize multi-path selection based on global security situational awareness including load distribution and anomalous traffic distribution. A risk assessment network is constructed based on attack behavior characteristics and service requirements to evaluate the risk level of routing performance degradation, thereby enabling risk-aware rerouting. Simulation results show that the proposed method increases the packet delivery ratio by 39.4%, reduces the end-to-end latency by 48.2% and the routing energy consumption by 41.1%.
Available online , doi: 10.11999/JEIT260798
Abstract:
Objective Twin-fisheye cameras provide near-360° coverage with low hardware complexity and are widely used in immersive imaging, surveillance, and mobile robotics. Their panoramic output depends on registration over a narrow overlapping band, so geometric accuracy and temporal consistency directly affect seam quality and video smoothness. After the two fisheye views are unfolded into the Equirectangular Projection (ERP), three coupled problems arise. First, the near-co-centric lens pair is ideally related by a pure rotation R ∈SO(3), whereas a conventional 8-Degree-of-Freedom (DoF) homography introduces five redundant parameters that may couple with matching noise. Second, ERP sampling is nonuniform with latitude, so identical pixel residuals do not represent identical spherical angular errors. Third, the cyclic ±π longitude boundary splits structures that are continuous on the sphere and weakens correspondences around the seam. Existing planar pipelines and generic learned matchers rarely combine these constraints under a unified rotation-referenced evaluation. This study therefore develops a lightweight registration framework that explicitly exploits spherical rotation geometry while addressing ERP boundary discontinuity, temporal fluctuation, and long-tail residuals. Methods The proposed framework contains three modules (Fig. 1 ). First, an overlapping-band Region-of-Interest (ROI) is cropped around the ERP seam and rearranged with modulo-W wrap-around (Fig. 2 ). A default longitude half-width of ±15° and an approximately 3° margin on each side preserve cross-boundary feature continuity while restricting the search region. XFeat detects, describes, and matches features under a fixed Top-K budget. Second, the two-dimensional matches are restored to global ERP coordinates, mapped to unit-sphere direction vectors, and processed by rotation-only SO(3)-RANSAC. Spherical angular residuals are used as the inlier criterion with a 0.8° threshold, a maximum of 2000 iterations, confidence 0.999, and at least 12 inliers; the iteration bound is updated adaptively. All inliers are then used for Kabsch/SVD closed-form rotation re-estimation, which reduces the randomness of a minimal sample while preserving the SO(3) constraint. Third, a local increment on the Lie algebra so(3) is optimized under a Huber loss by the Levenberg–Marquardt algorithm. The refinement is triggered only when the inlier-residual P95 exceeds 0.90° and the inlier ratio is below 0.58, thereby concentrating nonlinear optimization on difficult image pairs. Results and Discussions Experiments are conducted on PanoraMIS Sequences 3 and 4 under a unified relative inter-frame rotation protocol (Table 1 ). The proposed method achieves a 97.10% success rate, a 0.549° P95 angular residual, and a 0.313° temporal-stability error. Compared with SuperPoint+LightGlue, the P95 and temporal-stability errors are reduced by 22.8% and 77.3%, respectively. Compared with Efficient LoFTR, peak GPU memory and runtime are reduced by 52.7% and 60.9%, although Efficient LoFTR retains the lowest overall P95. Under a unified SO(3)-RANSAC back-end (Table 2 ), XFeat provides the largest average inlier count of 625.4 and the lowest temporal-stability error of 0.313° at 38.04 ms. The ablation and sensitivity results (Tables 3 –4 ) show that the ±15° ROI reduces the P95 from 0.720° for the full ERP to 0.600°. Replacing H-RANSAC with SO(3)-RANSAC reduces temporal instability from 0.931° to 0.313°, a 66.3% reduction, while increasing runtime from 24.91 ms to 35.24 ms. Adaptive refinement operates on approximately one third of the image pairs and improves both P95 and temporal stability with lower overhead than always-on refinement; its five-seed mean and median trigger rate are both 36.23%. A Top-K budget of 2048 reaches the saturated accuracy level, because increasing the budget to 4096 yields no further P95 or stability improvement. Five fixed-seed repetitions produce standard deviations no greater than 0.011° for P95 and stability, indicating that the main conclusions are insensitive to RANSAC randomness. In a 5×5 threshold sweep, the maximum changes in P95 and inter-frame rotation jitter within the central neighborhood are 4.40% and 0.037%, respectively, and changing the robust-error truncation from 3° to 2° or 5° does not alter the relative ranking. On the more difficult Sequence 4, characterized by weak texture and unstable overlap, the proposed method obtains 919.7 average inliers and a P95 of 0.383°, the lowest among the evaluated learning-based matchers, although robust-estimation time increases. With an identical standardized stitching back-end, it produces a lower seam-band gradient than H-RANSAC in the small-rotation example (26.90 versus 28.71) and a lower truth-referenced temporal-stability error (0.313° versus 0.931°; Fig. 6 ). On outdoor Sequence 7-L2, 324 of 346 correspondences are retained, yielding a 93.6% inlier ratio and a 0.524° P95 residual (Fig. 5 ). Because sequence-specific calibration is unavailable and the fixed inter-lens baseline may cause depth-dependent parallax, this result serves only as a diagnostic consistency check, not as evidence of absolute pose accuracy. Conclusions By restoring feature continuity across the ERP boundary, replacing the redundant planar homography with an explicit SO(3) rotation model, and selectively refining difficult image pairs on so(3), the proposed method balances registration accuracy, temporal consistency, and resource cost. It provides a lightweight front-end for twin-fisheye panorama stitching. The current evaluation is limited to pairwise registration on a small number of sequences; future work will address translation compensation, multi-frame global optimization, end-to-end integration with seam finding, exposure compensation, and blending, as well as generalization across additional platforms, dynamic scenes, and illumination conditions.
Available online , doi: 10.11999/JEIT260567
Abstract:
Objective Existing Alamouti-coding-based Non-Orthogonal Multiple Access (NOMA) schemes adopt an equal-number symbol transmission mode for both cell-center and cell-edge users, which overlooks the significant channel disparity between the two types of users. This leads to two drawbacks: on one hand, the superior channel condition of cell-center users is not fully exploited, resulting in wasted transmission resources; on the other hand, the equal-number transmission inevitably weakens the performance of Space-Time Block Coding (STBC) in suppressing intra-group multi-user interference. Therefore, a novel design that adapts to user channel differences is urgently needed to improve spectral efficiency and interference mitigation capability. Methods In light of the significant channel disparity between cell-center and cell-edge users, this paper proposes an unequal-number symbol transmission scheme for Alamouti coding in NOMA. Specifically, when constructing Alamouti group transmission codes, the number of symbols required for the cell-edge user is made smaller than that for the cell-center user. By reducing the number of transmitted symbols for the cell-edge user, two benefits are achieved: first, the multi-user interference imposed on the cell-center user is directly reduced; second, under a total power constraint, reducing the number of symbols for the cell-edge user equivalently increases the per-symbol transmission power, thereby significantly improving the received signal-to-interference-plus-noise ratio (SINR). Furthermore, the paper derives perfect closed-form solutions for the achievable sum rate and outage probability of the proposed scheme, and validates its effectiveness via numerical simulations. Results and Discussions The theoretical derivations yield closed-form analytical expressions for the achievable sum rate and outage probability, providing an accurate basis for system performance evaluation. Numerical simulation results demonstrate that, compared with existing equal-symbol Alamouti coding schemes, the proposed unequal-symbol Alamouti coding scheme effectively reduces the intra-group interference from the cell-edge user to the cell-center user, while significantly enhancing the SINR of the cell-edge user through power reallocation. Under typical channel parameters, the system achieves a notable sum-rate gain and a substantial reduction in outage probability, confirming the high efficiency of the proposed scheme. Conclusions The proposed unequal-number symbol transmission scheme based on Alamouti coding for cell-edge and cell-center users fully exploits the potential benefits arising from channel disparity. By reducing the number of symbols transmitted by the cell-edge user, the scheme achieves both suppression of intra-group interference and enhancement of the edge user’s transmission power. The theoretical closed-form solutions and simulation results consistently show that the proposed scheme outperforms conventional equal-number transmission schemes, providing an effective new approach for mitigating multi-user interference and optimizing resource allocation in NOMA systems.
Available online , doi: 10.11999/JEIT260645
Abstract:
Objective One of the core challenges in applying machine learning techniques to electronic design automation (EDA) lies in learning high-quality circuit representations from large-scale gate-level netlists. As modern digital integrated circuits scale to tens of millions of gates, existing methods based on graph neural networks (GNNs) and graph transformers (GTs) suffer from excessive computational and memory overhead, rendering them impractical for industrial-scale designs. The fundamental issue is the lack of a proper tokenization mechanism for netlists—unlike natural language, where subword tokenization effectively compresses long sequences, the circuit domain lacks an analogous decomposition strategy that preserves functional semantics while reducing the effective graph size. This work aims to bridge this gap by introducing a hypergraph-partitioning-based circuit tokenizer that decomposes massive netlists into functionally cohesive sub-circuits, termed circuit elements, thereby enabling scalable and fine-grained representation learning. Methods This paper proposes a function-aware hypergraph partitioning driven framework for large-scale circuit representation learning. Inspired by the tokenization paradigm of large language models, the framework first models a gate-level netlist as a directed hypergraph, where gates are nodes and signals are hyperedges that can connect multiple gates. A novel optimization objective, the Functional Independence Ratio (FIR), is introduced to guide the partitioning process. FIR incorporates circuit structural priors and a bus recognition correction mechanism that identifies bus-structured signals based on structural similarity (gate type purity across predecessor/successor levels) and spatial similarity (topological distance variance within candidate groups). The bus recognition module corrects the effective interface count, ensuring that functionally cohesive modules are not penalized for using wide buses. An iterative greedy refinement procedure accepts only moves that strictly decrease FIR, converging to a locally optimal partition. On top of the partitioned circuit elements, a two-stage self-supervised pretraining framework is designed. In the first stage, a masked autoencoder with edge prediction tasks is applied to the coarse-grained circuit-element graph, learning global inter-element dependencies. The graph transformer encoder is then frozen and circuit element embeddings are saved. In the second stage, within each circuit element, a contrastive learning scheme is employed at the gate level. Positive pairs are constructed via Boolean equivalence transformations (e.g., associativity, De Morgan's laws), which preserve the Boolean function while altering the gate-level structure. Negative pairs are drawn from functionally different circuit elements within the same batch. The training jointly optimizes node-level and local-global alignment losses. A feature-wise modulation mechanism injects circuit-element-level context into gate-level representations, enabling the same gate type to acquire different embeddings depending on its functional context. Results and Discussions Extensive experiments are conducted on circuits collected from multiple sources, including ITC99, EPFL, OpenCores, and three RISC-V SoC designs (Rocket, BOOM, and OpenC910), with the largest design containing 22.3 million gates. For the tokenizer evaluation, FIR-based partitioning is compared against Mt-kahypar using the Adjusted Mutual Information (AMI) and Adjusted Rand Index (ARI) metrics, with the original module hierarchy serving as ground truth. Across all three RISC-V designs, FIR achieves AMI improvements of 24.6% to 27.1% and ARI improvements of 25.5% to 35.8% over Mt-kahypar, demonstrating consistent and substantial gains in functional coherence. On downstream tasks, the proposed method consistently outperforms state-of-the-art baselines including DeepGate4 and NetTAG across all comparable datasets. For full-circuit function recognition, the proposed method achieves F1 scores of 0.896, 0.861, 0.817, and 0.773 on ITC99, EPFL, OpenCores, and Rocket, respectively, representing improvements of 10.0% to 15.4% over the strongest baseline. Crucially, on BOOM (8.12M gates) and OpenC910 (22.3M gates), the proposed method is the only approach capable of completing training without running out of memory, achieving F1 scores of 0.771 and 0.748 for full-circuit tasks and 0.723 and 0.679 for gate-level tasks, respectively. Ablation studies with six model variants reveal a clear division of labor: the module-level pretraining stage dominates full-circuit performance (20.9% F1 drop when removed), while the gate-level pretraining stage dominates gate-level performance (26.1% F1 drop when removed). The FIR partitioning objective contributes 8.6% and 11.7% F1 improvements at the full-circuit and gate levels, respectively. The bus recognition module and the two-stage architecture also show consistent positive contributions across all metrics. Conclusions This paper presents a novel framework that addresses the scalability challenge in circuit representation learning by introducing a function-aware hypergraph partitioning tokenizer and a two-stage self-supervised pretraining architecture. The key insight is that by abstracting the intermediate circuit element level between individual gates and the full circuit, one can achieve both scalability to multi-million-gate designs and fine-grained gate-level discriminative capability. The proposed FIR objective effectively captures functional cohesion during partitioning, and the two-stage pretraining framework decouples global context learning from local representation refinement. Experimental results demonstrate state-of-the-art performance on function recognition tasks and, more importantly, the unique ability to scale to industrial-sized designs where existing methods fail. Future work includes extending the framework to other EDA tasks such as logic synthesis and physical design, exploring more aggressive hierarchical strategies for billion-gate designs, and adapting the bus recognition mechanism to non-standard cell libraries.
Available online , doi: 10.11999/JEIT260695
Abstract:
Objective Traditional optimization-based power allocation algorithms for collocated MIMO radar have two fundamental limitations. First, they optimize tracking performance only for the next time step and therefore lack a full-time-horizon view of the power allocation process. This myopic strategy cannot achieve optimal multi-target tracking accuracy over extended periods, particularly when target trajectories vary substantially. Second, these algorithms rely on iterative nonlinear constrained optimization, resulting in high computational complexity. Therefore, they cannot satisfy the real-time requirements of dynamic battlefield environments where target states change rapidly. To address these limitations, this paper proposes a Reinforcement Learning (RL)-driven power allocation algorithm. Unlike conventional methods, the proposed approach formulates the power allocation problem as a Markov Decision Process (MDP) that maximizes long-term cumulative tracking accuracy. The algorithm adaptively allocates limited transmit power among multiple beams according to the current system state, balancing immediate tracking performance with long-term cumulative tracking accuracy. Methods The Posterior Cramér-Rao Lower Bound (PCRLB) is employed to quantify the theoretical lower bound of the tracking error for each target. The state space is constructed by combining the motion states (position and velocity) of all targets with the normalized PCRLB from the previous allocation step. The action space consists of discrete transmit power levels for each beam, subject to the total power budget and individual beam power constraints. All feasible power allocation vectors are enumerated and encoded to reduce the action-space dimensionality. The reward function is defined as the negative weighted sum of the normalized PCRLB, encouraging the agent to minimize tracking errors. The power allocation process is formulated as an MDP and solved using the Dueling Double Deep Q-Network (D3QN) algorithm. The D3QN framework incorporates three major enhancements: (1) a double-network architecture comprising an online Q-network and a target Q-network to improve training stability; (2) a dueling architecture that decomposes the Q-value into a state-value function and an action-advantage function to improve action discrimination; and (3) off-policy learning with experience replay to improve the use of historical trajectories. An ε-greedy strategy is adopted for exploration, with ε gradually decreasing during training. After offline training, the learned network directly generates real-time transmit power allocation decisions from the current system state without iterative optimization. Results and Discussions Simulations are conducted using three targets following the Constant Velocity (CV) model. Fixed power allocation yields the lowest tracking accuracy because of inefficient resource utilization. The traditional optimization method, which minimizes the instantaneous tracking error, achieves moderate tracking performance but remains myopic. When the discount factor \begin{document}$ \gamma =0 $\end{document} \begin{document}$ \gamma =0.99 $\end{document} \begin{document}$ \gamma =0.99 $\end{document} Conclusions This paper proposes an RL-driven power allocation algorithm for collocated MIMO radar multi-target tracking that overcomes the myopic behavior and high computational complexity of conventional optimization methods. The proposed algorithm constructs the state space and reward function using the PCRLB, models the power allocation process as an MDP, and solves it using the D3QN algorithm. Simulation results demonstrate that, with an appropriate discount factor (\begin{document}$ \gamma =0.99 $\end{document}
Available online , doi: 10.11999/JEIT260296
Abstract:
Objective Unmanned systems play an increasingly important role in emergency response, public safety, intelligent transportation, and other mission-critical applications. Reliable communications in complex electromagnetic environments are essential for autonomous operation. However, communication links are directly exposed to open, non-cooperative electromagnetic environments and are therefore vulnerable to intentional jamming and unintentional interference. In practical scenarios, prior information regarding the desired signal, jamming sources, and multipath propagation is often unavailable, substantially degrading the performance of conventional spatial-domain anti-jamming methods. To address this challenge, this paper proposes a spatial-domain anti-jamming framework for unmanned systems operating under limited prior information. Methods The proposed method first applies a spatial smoothing algorithm to the received signals to decorrelate coherent multipath components. Capon spatial spectrum estimation is then performed to detect the Direction Of Arrival (DOA) of potential incident signals. Spectrum peaks corresponding to individual incident signals are subsequently identified. A Covariance Matrix Reconstruction (CMR)-based beamforming algorithm is then applied by traversing all detected spectrum peaks to sequentially extract the signal associated with each peak, thereby separating the mixed signals. After signal separation, a signal classification method based on spectral similarity and time delay is employed. Kullback-Leibler (KL) divergence between the spectrum of each separated signal and the reference spectrum is calculated to identify jamming signals. The remaining communication signals are further classified into direct-path and multipath signals according to their relative time delays. Finally, different processing strategies are applied according to the identified signal type. Specifically, multipath signals are either suppressed as interference or coherently combined with the direct-path signal after time-delay and phase alignment. Results and Discussions Two simulation scenarios, including jamming only and combined jamming and multipath, are designed to evaluate the proposed method in terms of the output Signal-to-Interference-plus-Noise Ratio (SINR), beam pattern, Bit Error Rate (BER), and Error Vector Magnitude (EVM). Simulation results demonstrate that, under the jamming-only scenario, the proposed method achieves performance close to the theoretical optimum. The output SINR increases with the input Signal-to-Noise Ratio (SNR) at a fixed Jamming-to-Signal Ratio (JSR) (Fig. 3(a) ) and remains nearly unchanged as JSR increases at a fixed SNR (Fig. 3(b) ), indicating stable jamming suppression capability. The recovered time-domain waveform and spectrum remain highly consistent with the transmitted signal (Fig. 4 ). The BER curve nearly overlaps that of the optimal beamformer (Fig. 5 ). At \begin{document}$ {E}_{\rm b}/{N}_{0}=10\;{\mathrm{dB}} $\end{document} Fig. 6 ). Under simultaneous jamming and multipath conditions, the proposed framework flexibly suppresses or exploits multipath signals. Compared with multipath suppression, multipath utilization further improves both the output SINR and BER (Fig. 7(a) and Fig. 7(b) ). The corresponding beam pattern forms a beam toward the multipath direction rather than a null, demonstrating effective multipath exploitation (Fig. 7(c) ). Conclusions This paper proposes a spatial-domain anti-jamming framework for unmanned systems operating under limited prior information. Using only the received mixed signals, the proposed framework estimates the directions of arrival, separates incident signals, and classifies them as direct-path, multipath, or jamming signals. Appropriate suppression or preservation strategies are then applied according to the identified signal type. Therefore, the framework flexibly suppresses or exploits multipath signals while preserving the direct-path signal and mitigating jamming. Simulation results demonstrate the effectiveness of the proposed method in terms of output SINR and demodulation accuracy, confirming reliable jamming suppression and communication performance even when prior information regarding the desired signal, jamming sources, and multipath propagation is unavailable. Future work will investigate the effects of array perturbations, intelligent jamming, and heterogeneous communication modes on the proposed framework and extend it to more complex unmanned-system communication environments.
Available online , doi: 10.11999/JEIT260516
Abstract:
Objective In recent years, memristors have been employed to emulate neuronal synapses with dynamically adjustable synaptic weights, enabling the construction of Memristive Hopfield Neural Networks (HNNs). Compared with conventional HNNs, Memristive HNNs more accurately reproduce the nonlinear dynamical behavior of biological neural systems. Multi-scroll attractors have attracted considerable attention in secure communication because of their complex topological structures and strong state-space ergodicity. However, previous studies have primarily focused on conventional multi-scroll attractors with single structural patterns, whereas multi-scroll attractors with special structures remain largely unexplored. Therefore, this paper proposes a nested multi-scroll Memristive HNN system that generates nested multi-scroll attractors, thereby overcoming the limitations of conventional single-structure multi-scroll attractors. Methods A Four-Dimensional (4D) Memristive HNN system is constructed from a three-neuron HNN by incorporating a multi-segment nonlinear magnetically controlled memristor into the Memristive self-connected synapse of neuron 2. Equilibrium-point and stability analyses are performed to investigate the regulatory effects of the Memristive self-connected synapse coupling strength and system initial conditions on the system dynamics. The number of multi-scroll attractors is regulated by adjusting the memristor control parameters. Building on this framework, a Multi-level Logic Pulse current (IMLP) is introduced to construct a nested multi-scroll Memristive HNN system. The proposed system generates nested multi-scroll attractors with enhanced dynamical complexity. Finally, the MATLAB numerical simulation results are validated through Multisim circuit simulations and Field-Programmable Gate Array (FPGA)-based hardware experiments. Results and Discussions The results demonstrate that regulating the Memristive self-connected synapse coupling strength enables the proposed 4D Memristive HNN system to exhibit period-doubling bifurcations and chaotic behavior, as illustrated by the bifurcation diagrams and Lyapunov exponent spectra (Fig. 3 ). Various types of coexisting attractors are generated under different coupling strengths (Fig. 4 ). By adjusting the memristor control parameters, multi-scroll attractors with different numbers of scrolls are generated through one-directional extension (Figs. 5 ～8 ). After the introduction of the IMLP, the proposed nested multi-scroll Memristive HNN system generates nested multi-scroll attractors while preserving the controllable scroll-number extension property (Figs. 8 ～10 ). Spectral Entropy (SE) analysis demonstrates that the IMLP increases the dynamical complexity of the proposed system compared with the original 4D Memristive HNN system (Figs. 9 and 10 ). The strong agreement among MATLAB numerical simulations, Multisim circuit simulations, and FPGA-based hardware experiments confirms the physical realizability of the proposed nested multi-scroll Memristive HNN system (Figs. 12 ～14 ). Conclusions A 4D Memristive HNN system is constructed by incorporating a multi-segment nonlinear magnetically controlled memristor into the Memristive self-connected synapse of a three-neuron HNN. Equilibrium-point and stability analyses reveal the regulatory effects of the Memristive self-connected synapse coupling strength and the evolution of coexisting attractors associated with different initial conditions. The results show that the system enters chaos through the period-doubling route to chaos and generates single-scroll and double-scroll chaotic attractors. The number of multi-scroll attractors is continuously increased by adjusting the memristor control parameters. Furthermore, introducing the IMLP produces a nested multi-scroll Memristive HNN system capable of generating nested multi-scroll attractors with increased dynamical complexity. The strong agreement among MATLAB numerical simulations, Multisim circuit simulations, and FPGA-based hardware experiments validates the physical realizability of the proposed nested multi-scroll Memristive HNN system.
Available online , doi: 10.11999/JEIT260514
Abstract:
Objective The Ka-band has become a key frequency band for satellite communications, placing stringent requirements on the millimeter-wave power amplifier, a core component of the transmitter, to provide high efficiency, compact size, and broadband operation. To maximize spectral efficiency, millimeter-wave satellite communication signals typically exhibit a high Peak-to-Average Power Ratio (PAPR), making high back-off efficiency particularly important. Although the Doherty Power Amplifier (DPA) is widely adopted because of its high efficiency under power back-off conditions, its operating bandwidth is inherently limited. In addition, both saturated efficiency and back-off efficiency degrade substantially at millimeter-wave frequencies. Therefore, extending the operating bandwidth while maintaining high efficiency remains a major challenge for millimeter-wave DPAs used in satellite communication transmitters. Methods An ultra-wideband enhanced active load modulation technique is proposed to overcome the trade-off between bandwidth and back-off efficiency in DPAs. The proposed method achieves optimal load impedance modulation for both the carrier and peaking amplifiers over an ultra-wide frequency range by introducing an Impedance Tunable Bias Network (ITBN) and a dual-drive impedance control mechanism. These techniques improve load modulation while extending the load modulation bandwidth, thereby enhancing both efficiency and bandwidth in the millimeter-wave DPA architecture. Furthermore, a broadband phase-compensation technique is integrated into an unequal power division network to achieve sufficient load modulation across the entire operating band while accurately compensating for the phase difference between the carrier and peaking paths. The proposed ultra-wideband phase-compensated power division network further extends the high-efficiency operating bandwidth while reducing the chip area. Results and Discussions To validate the proposed method, a millimeter-wave ultra-wideband high-efficiency DPA was designed and fabricated using a 0.15-μm GaN process. Across the 24～33 GHz frequency band, corresponding to a relative bandwidth of 31.6%, the fabricated chip achieves a small-signal gain of 21.8～23.8 dB with a gain flatness of ±1 dB. The measured saturated output power is 28.9～31.0 dBm, with a Power-Added Efficiency (PAE) of 25.2%～33.5% at saturation and a 6 dB back-off PAE of 15.5%～19.8%. Compared with previously reported GaN DPAs operating over similar frequency bands, the proposed design achieves the highest reported small-signal gain and saturated PAE while maintaining high saturated output power and high 6 dB back-off PAE. Furthermore, it occupies the smallest chip area among reported three-stage DPA MMICs. Conclusions An ultra-wideband enhanced active load modulation method is proposed to achieve sufficient active load modulation through multi-frequency impedance tuning across the entire operating band. A novel ultra-wideband phase-compensated unequal power division network is also proposed to reduce the chip area while maintaining accurate phase compensation. To validate the proposed method, a millimeter-wave ultra-wideband high-efficiency DPA was fabricated using a 0.15-μm GaN process. Measurement results demonstrate that, over the 24～33 GHz frequency band, the fabricated chip achieves a small-signal gain of 21.8～23.8 dB, a saturated output power of 28.9～31.0 dBm, a saturated PAE of 25.2%～33.5%, and a 6 dB back-off PAE of 15.5%～19.8%. Compared with previously reported GaN DPAs operating in similar frequency bands, the proposed design achieves a maximum relative bandwidth of 31.6%, the highest reported small-signal gain and saturated PAE, high saturated output power, and high 6 dB back-off PAE, while occupying the smallest chip area among reported two- and three-stage DPA MMICs. These measurement results validate the proposed method and demonstrate its strong potential for millimeter-wave satellite communication transmitters.
Available online , doi: 10.11999/JEIT260418
Abstract:
Objective In Ambient Backscatter Communication (AmBC) systems, a low-complexity Energy Detector (ED) is commonly employed at the reader to recover symbols transmitted by the tag. The detection performance of ED depends strongly on the accurate setting of the detection threshold, which is determined by the average received signal power corresponding to tag symbols “1” and “0”. Existing parameter estimation methods assume that the two symbols are transmitted with equal probability and therefore divide the sorted received signal power samples into two equal groups. However, because the number of transmitted symbols is finite, the actual numbers of symbols “1” and “0” are generally unequal. Therefore, equal partitioning introduces sample misclassification, causing the estimated threshold to deviate from its optimal value and reducing detection performance. To address this limitation, an iterative threshold parameter estimation method is proposed to reduce the parameter estimation bias caused by sample misclassification and improve the accuracy of detection threshold estimation. Methods An iterative threshold parameter estimation method is proposed to overcome the sample misclassification introduced by conventional sorting-based grouping. Because the initial detection threshold obtained by the sorting-based grouping method provides reliable decisions for most received samples, these initial decisions are used as the basis for sample reclassification. The received signal samples are then reclassified to iteratively update the threshold parameters, progressively refining the detection threshold. The proposed method is evaluated through simulations under three representative ambient radio-frequency source conditions: complex Gaussian, Phase-Shift Keying (PSK), and Quadrature Amplitude Modulation (QAM) sources. Results and Discussions Simulation results show that, over a wide range of Signal-to-Noise Ratio (SNR) values, the proposed iterative method substantially reduces the Bit Error Rate (BER) compared with the conventional sorting-based grouping method and approaches the theoretical lower bound obtained with perfect parameter estimation. At a given SNR, the proposed method improves BER by approximately 0.1, 1.5, and 1.3 orders of magnitude under complex Gaussian, PSK, and QAM sources, respectively (Fig. 2 ). These results demonstrate that the proposed iterative method effectively corrects sample misclassification and reduces the performance loss caused by parameter estimation bias. Moreover, most of the performance gain is achieved after only one iteration, indicating rapid convergence with minimal additional computational overhead. Under different numbers of sampling points, BER improvements of approximately 0.5, 1.6, and 1.1 orders of magnitude are achieved under complex Gaussian, PSK, and QAM sources, respectively (Fig. 3 ). These results indicate that using the initial decisions for sample reclassification effectively reduces the estimation bias introduced by fixed equal partitioning, thereby improving detection performance under limited-sample conditions. Under different Relative Channel Difference (RCD) values, BER improvements of approximately 0.56 and 0.7 orders of magnitude are achieved under complex Gaussian and QAM sources, respectively (Fig. 4 ). As the RCD increases, the separation between the received signal power distributions becomes more pronounced, improving the accuracy of the initial decisions and enabling more reliable sample reclassification. This positive feedback process further refines the parameter estimates and improves detection performance. Conclusions An iterative threshold parameter estimation method is proposed to address the sample misclassification introduced by conventional sorting-based grouping in ED. The proposed method uses the initial decisions to reclassify the received signal samples and iteratively update the threshold parameters. In addition, closed-form expressions for the detection threshold and BER under QAM sources are derived. Simulation results demonstrate that the proposed method effectively reduces parameter estimation bias with only one iteration while maintaining robust performance under limited-sample and varying channel conditions. Significant BER improvements are achieved with minimal additional computational overhead, making the proposed method well suited for practical, high-reliability AmBC systems.
Available online , doi: 10.11999/JEIT260447
Abstract:
Objective Direction Of Arrival (DOA) estimation is a key technique for sonar target localization. As the demand for high-precision DOA estimation in complex environments continues to increase, the number of array elements used for estimation is steadily growing, leading to massive arrays. Although larger arrays improve DOA estimation accuracy and resolution, they also impose a substantial computational burden on conventional DOA estimation algorithms. To address this issue, a low-complexity composite hierarchical receiving array structure is constructed, and two fast phase ambiguity resolution algorithms are proposed: Composite HierArchical Global Nearest-Neighbor Matching (CHA-GNNM) and Composite HierArchical Cross-Correlation Covariance Merging (CHA-CCM). Methods The CHA-GNNM algorithm constructs multiple candidate solution sets by exploiting the auto-covariance and cross-covariance relationships among the subarrays within each group. The true solution in each candidate solution set is identified through nearest-neighbor matching based on source consistency, and the final DOA estimate is obtained through multilevel coherent combining. This approach achieves phase ambiguity resolution and angle matching with relatively low computational cost. However, because the correlation information among all array elements is not fully exploited, some estimation performance is sacrificed. To improve DOA estimation performance, the CHA-CCM algorithm reorganizes the composite hierarchical structure into evenly partitioned groups, which are regarded as several large subarrays. Multiple large candidate solution sets are first constructed from the cross-correlation relationships among these groups. Each group is then divided into multiple small subarrays, from which additional candidate solution sets are generated using the corresponding auto-covariance and cross-covariance relationships. A coarse DOA estimate is obtained through coprime clustering, followed by a more accurate initial DOA estimate derived from the small candidate solution sets. This initial DOA estimate is subsequently used to eliminate spurious solutions from the large candidate solution sets, yielding the final DOA estimate. Combined with a low-complexity covariance block-processing strategy, this approach avoids computationally expensive operations while improving DOA estimation accuracy. Results and Discussions Simulation results demonstrate that both proposed algorithms substantially reduce the computational burden as the number of array elements increases, while effectively achieving phase ambiguity resolution through the proposed composite hierarchical receiving array structure (Fig. 5 ). Compared with the conventional Root-MUSIC algorithm, CHA-GNNM achieves coarse DOA estimation with nearly four orders of magnitude lower computational complexity (Fig. 7 ), making it suitable for applications with stringent real-time requirements. In contrast, CHA-CCM requires only a modest increase in computational cost (Fig. 7 ) while achieving DOA estimation performance close to the Cramér-Rao Lower Bound (CRLB) above a certain signal-to-noise ratio threshold (Fig. 6 ). Therefore, a favorable balance is achieved between DOA estimation accuracy and computational complexity. Conclusions To address the rapid increase in computational complexity associated with massive arrays, a composite hierarchical receiving array structure is constructed for efficient DOA estimation. By hierarchically grouping the array elements, the proposed structure provides a new framework for low-complexity DOA estimation. Based on this structure, two fast DOA estimation algorithms are developed. Both algorithms achieve effective phase ambiguity resolution with low computational complexity by exploiting the structural differences among array groups and the consistency of observations from the same source across different groups, thereby enabling rapid DOA estimation. CHA-GNNM primarily exploits the phase relationships among subarrays to perform phase ambiguity resolution and angle matching through a simple computational procedure, making it suitable for applications requiring high computational efficiency and real-time processing. Because the cross-correlation information among all array elements is not fully exploited, some estimation performance is reduced under challenging signal conditions. To overcome this limitation, CHA-CCM reorganizes the composite hierarchical receiving array into evenly partitioned groups while preserving the low-complexity advantage of the hierarchical structure. Group-level cross-correlation information is further exploited so that the intrinsic relationships among different groups are more fully utilized. In addition, the signal processing procedure is simplified by eliminating unnecessary computational steps, thereby improving the robustness and accuracy of DOA estimation while maintaining manageable computational complexity. Compared with CHA-GNNM, CHA-CCM incurs only a small increase in computational cost and achieves a better balance between computational complexity and DOA estimation performance. Overall, the proposed composite hierarchical receiving array structure and the two fast DOA estimation algorithms provide an effective solution for efficient DOA estimation in massive arrays. CHA-GNNM is more suitable for applications with stringent real-time requirements, whereas CHA-CCM is better suited for applications requiring higher DOA estimation accuracy and robustness. The proposed structure achieves efficient phase ambiguity resolution and accurate DOA estimation and provides both theoretical significance and practical value for engineering applications of massive array signal processing.
Available online , doi: 10.11999/JEIT260404
Abstract:
Objective With the widespread deployment of extra-large-scale antenna arrays in 6G networks, user terminals are increasingly located in the near-field region. Existing Near-Field Integrated Sensing And Communication (NF-ISAC) algorithms face key challenges, including off-grid power leakage, severe model mismatch, and strong pilot dependence. These limitations make them unsuitable for low-overhead, high-performance 6G transmission. This paper aims to design an off-grid blind NF-ISAC algorithm and derive the theoretical performance bound for near-field sensing. Methods To overcome the limitations of analytical geometric steering vectors and adapt to more accurate electromagnetic propagation characteristics without closed-form expressions, an amplitude-phase separation method is first proposed. This method decomposes the nonlinear near-field steering vector into amplitude and phase terms, enabling high-precision characterization of the steering vector using a single-hidden-layer neural network. Second, the NF-ISAC problem is formulated as a constrained matrix factorization problem, and a corresponding factor graph model is constructed. The trained neural network is embedded into the factor graph as a function node. Message passing through the embedded neural network is then achieved, enabling joint blind coordinate sensing, channel estimation, and signal detection in a pilot-free manner. Finally, the Cramér-Rao Lower Bound (CRLB) for multi-user near-field joint distance and angle sensing in polar coordinates is derived based on the neural-network-fitted steering vector. Results and Discussions Extensive Monte Carlo simulations are conducted to evaluate the performance of the proposed algorithm. The simulation results show that the proposed algorithm achieves millimeter-level position sensing. Compared with existing mainstream algorithms, it improves both communication Bit Error Rate (BER) and sensing accuracy. The proposed algorithm achieves a 2～3 dB gain in sensing accuracy over the state-of-the-art near-field off-grid algorithm, and its performance is closest to the derived theoretical CRLB. These results indicate that the proposed algorithm effectively mitigates off-grid power leakage and model mismatch. Conclusions The proposed off-grid blind NF-ISAC algorithm overcomes the pilot dependence and model mismatch of existing NF-ISAC schemes. It achieves integrated high-precision sensing and reliable communication for near-field users in a pilot-free manner. The derived CRLB provides a theoretical benchmark for evaluating the sensing performance of NF-ISAC systems. This work provides technical support for the design of 6G NF-ISAC systems.
Available online , doi: 10.11999/JEIT260458
Abstract:
Objective With the growing demand for high-capacity and high-spectral-efficiency transmission in LEO satellite communications, massive MIMO has become a promising enabling technology. However, conventional fully digital beamforming is difficult to implement in practice due to the strict constraints on power consumption, hardware complexity, and payload cost of satellite platforms. Although partially connected hybrid beamforming can reduce hardware complexity, the conventional fixed subarray structure lacks flexibility and suffers from performance degradation, especially under low-resolution PSs constraints. To address this issue, this paper investigates adaptive antenna-RF chain mapping for hybrid beamforming design in LEO satellite massive MIMO systems. Methods This paper first establishes a system model for LEO satellite multi-user downlink massive MIMO hybrid beamforming and formulates a joint optimization problem with the objective of maximizing spectral efficiency. Considering the constant-modulus discrete phase constraints of low-resolution PSs, antenna-RF chain connection constraints, and transmit power constraint, the resulting problem is highly non-convex. To solve it, the original problem is transformed into an equivalent WMMSE formulation, and auxiliary variables are introduced to decouple the coupled variables. Based on this reformulation, a double-layer iterative optimization framework is developed by combining the BCD method and the PDD method. For adaptive antenna-RF chain mapping, the mapping problem is reformulated as a capacity-constrained linear assignment problem, and an optimal adaptive mapping method based on the Hungarian algorithm is proposed. Furthermore, to reduce the computational burden in large-scale antenna array scenarios, a low-complexity adaptive mapping method based on antenna priority sorting is developed. Results and Discussions Simulation results show that the proposed methods exhibit good convergence behavior. Specifically, the spectral efficiency increases rapidly in the initial iterations and then gradually converges, while the constraint violation decreases continuously, confirming the effectiveness of the proposed iterative optimization framework (Fig. 3 ). In terms of spectral efficiency, the proposed adaptive mapping methods consistently outperform the conventional fixed subarray and the existing greedy dynamic subarray scheme over different transmit powers and antenna scales. Among them, the Hungarian-based method achieving better spectral efficiency, whereas the antenna priority sorting-based method attains near-optimal performance with significantly reduced computational complexity (Figs. 4 and 5 ). As the number of PSs quantization bits increases, the system performance gradually approaches that of the continuous-PSs case, demonstrating the effectiveness of the proposed low-resolution PSs-based design (Fig. 6 ). In terms of energy efficiency, the proposed methods also outperform the conventional fully digital beamforming, fully connected, fixed subarray, and greedy dynamic subarray hybrid beamforming under different transmit powers and antenna scales (Figs. 7 and 8 ). Conclusions This paper proposes an adaptive hybrid beamforming design for LEO satellite massive MIMO systems under low-resolution PS constraints. By combining WMMSE reformulation with a PDD-BCD based optimization framework, joint design of digital precoding, analog precoding, and adaptive antenna-RF chain mapping is achieved. Simulation results demonstrate that the proposed methods provide superior performance in both spectral efficiency and energy efficiency. In particular, the Hungarian-based method provides better system performance, while the antenna priority sorting-based method achieves a favorable trade-off between performance and computational complexity. The proposed design provides an effective solution for high-performance hybrid beamforming in hardware-constrained LEO satellite massive MIMO systems.
Available online , doi: 10.11999/JEIT260328
Abstract:
Objective Refractory metal tungsten, particularly single crystal tungsten, serves as a critical material for high-performance field emission cathodes, which are core components in advanced electron microscopes and electron beam lithography systems. The fabrication of single crystal tungsten microtips with precisely controlled geometry and surface cleanliness remains a significant challenge, especially for applications requiring tip radii ranging from sub-100nm for cold field emission to 0.3–1.0 μm for Schottky-type thermal field emission. Currently, the domestic development of high-end electron microscopes in China faces a major bottleneck due to heavy reliance on imported field emission cathodes. Electrochemical corrosion, the mainstream method for tip preparation, suffers from limitations such as difficulty in cutoff timing control, susceptibility to tip bending or passivation, residual surface impurities, and low yield rates. Moreover, the anisotropic nature of single crystal tungsten introduces additional complexity in morphology control. This study aims to establish a controllable fabrication method for single crystal tungsten tips, enabling tailored geometry and surface quality to meet demanding requirements of different field emission applications while significantly improving fabrication yield. Methods Single crystal tungsten wires (diameter 0.12 mm, purity 99.95%, (100) orientation) were used as the starting material. The tips were first pre-shaped by electrochemical corrosion in 1 mol/L NaOH solution under a 10 V DC voltage with pulsed control (6 kHz, 100 μs pulse width) for 10 min. Following corrosion, the tips underwent surface cleaning by sequential immersion in ultrapure water and anhydrous ethanol, followed by drying with high-purity nitrogen. Subsequently, the samples were placed in an ultra-high vacuum system (base pressure < 1 × 10–6 Pa) and subjected to high-temperature oxygen treatment. Process parameters were systematically varied, including temperature (1500 –1600 °C), oxygen flow rate (0.015–1.0 sccm), and treatment duration (5–1440 min). During treatment, high-purity oxygen was introduced while maintaining vacuum levels better than 10–4 Pa. Tip morphologies were characterized by scanning electron microscopy, and surface compositions were analyzed by energy dispersive X-ray spectroscopy. Geometric parameters, including half-cone angle and tip radius, were measured following standardized protocols. Results and Discussions The combination of electrochemical corrosion and high-temperature oxygen treatment enabled both effective surface purification and controlled morphological reconstruction. Scanning electron microscopy characterization revealed distinct evolutionary pathways depending on oxygen partial pressure (Fig. 3, Fig. 4). Under oxygen-rich conditions (1.0 sccm, 1500 –1600 °C), the tips underwent “blunting reconstruction,” evolving from an initial inverted cone into a characteristic “cylindrical segment + hemispherical cap” structure (Samples 2–5, Table 2 ). For Sample 3 treated at 1500 °C for 10 min, the tip radius increased to 327 nm with a half-cone angle reduction of 21%; for Sample 4 treated for 15 min, further evolution occurred with tip radius decreasing to 275 nm and cylindrical segment height increasing substantially. With increasing temperature from 1500 °C (Sample 2) to 1600 °C (Sample 5), the half-cone angle change rate increased from 11% to 25%, while tip radius decreased from 380 nm to 172 nm, indicating accelerated kinetics. This anisotropic behavior is attributed to orientation-dependent surface energies of tungsten and preferential oxidation along specific crystallographic planes. Under oxygen-lean conditions (0.015 sccm, 1500 °C, 1440 min), “sharpening reconstruction” occurred, with tip radius decreasing dramatically from 143 nm to 28 nm and half-cone angle reducing from 8° to 1° (Sample 6). In contrast, treatment in vacuum at 1500 °C primarily removed surface contaminants without altering tip geometry (Sample 1). EDS analysis confirmed that treated surfaces were free from impurities other than trace carbon contamination (Table 3 ), demonstrating the dual functionality of the process in achieving both purification and reconstruction. The findings reveal that oxygen partial pressure serves as the key determinant of reconstruction direction, with oxygen-rich conditions favoring blunting and oxygen-lean conditions promoting sharpening. Conclusions A combined process of electrochemical corrosion followed by high-temperature oxygen treatment was successfully developed for the controllable fabrication of single crystal tungsten field emitter tips. The process achieves dual functionality: effective surface purification through high-temperature vacuum treatment, which removes residual surface contaminants from electrochemical corrosion, and controlled morphological reconstruction enabled by the introduction of oxygen under precisely regulated conditions. By systematically adjusting process parameters including treatment temperature, oxygen flow rate, and treatment duration, tip morphology and dimensions can be tailored to meet specific application requirements. The oxygen partial pressure plays a decisive role in determining the reconstruction pathway. Under oxygen-rich conditions (e.g., 1500 °C, 1.0 sccm), blunting reconstruction yields a “cylindrical segment + hemispherical cap” structure ideally suited for Schottky-type thermal field emission cathodes requiring tip radii between 0.3 μm and 1.0 μm. Under oxygen-lean conditions (e.g., 1500 °C, 0.015 sccm), sharpening reconstruction produces nanoscale sharp tips with tip radius as low as 28 nm, suitable for cold field emission cathodes and scanning tunneling microscope probes. This anisotropic reconstruction mechanism is explained by the selective reaction of oxygen atoms with different tungsten crystal planes under high-temperature conditions, where the formation of volatile oxides modifies local surface energy distribution and promotes the exposure of specific crystallographic planes. Control experiments confirmed that such morphological reconstruction does not occur under oxygen-lean or oxygen-free conditions at the same temperatures, further demonstrating the critical role of oxygen in triggering this process. Importantly, this approach provides an effective method to correct imperfections from the initial electrochemical corrosion step, significantly improving fabrication yield and process robustness, thereby offering a viable technical pathway for the domestic fabrication of high-performance field emission cathodes.
Available online , doi: 10.11999/JEIT260370
Abstract:
Objective With the rapid development of the low-altitude economy and 6G intelligent networks, Unmanned Aerial Vehicle (UAV) image communication shows strong potential in target reconnaissance, emergency communication, and intelligent inspection. However, conventional pixel-level transmission cannot meet the requirements of efficient, low-latency, and intelligent communication because UAVs are constrained by limited bandwidth, payload capacity, and onboard computational resources. Semantic communication, which transmits only task-relevant information, provides an effective solution for improving communication efficiency in resource-constrained scenarios. However, current studies on UAV image transmission face several challenges. First, fixed network architectures use unified semantic encoding and transmission strategies for all users and cannot adapt to different personalized requirements. Second, new user access usually requires interest pre-training or model fine-tuning, which increases deployment overhead. Third, most models have high computational complexity. To address these issues, this paper proposes the Lightweight Personalized UAV Semantic Communication (LPUSC) system to balance computational cost, transmission bandwidth, and personalized requirements. The system enables personalized transmission through low-overhead semantic index interaction and a lightweight semantic extraction module, without pre-training for new users. A dual-branch end-to-end network is also designed. In this network, the semantic index transmission network works with the semantic image transmission network trained by a weighted hybrid loss function, thereby supporting high-precision and high-quality transmission of personalized semantic images. Methods The proposed LPUSC system adopts a dual-branch architecture for accurate task-driven semantic content transmission. In the semantic index interaction branch, the lightweight object detection model YOLO11s is used to perform semantic perception on UAV-captured visual scenes. Complex image information is compressed into low-dimensional semantic index vectors, which reduces transmission redundancy and communication overhead. On this basis, an end-to-end semantic index transmission network is designed to improve the robustness of semantic index transmission under complex wireless channel conditions. Through the semantic index interaction mechanism, the system accurately identifies targets of user interest and provides prior guidance for subsequent semantic content extraction. In the semantic image transmission branch, the lightweight and high-precision MobileSAM model is adopted for semantic region extraction. This branch uses the target bounding boxes returned by the semantic index interaction branch as box-prompt inputs, enabling pixel-accurate segmentation and extraction of specific semantic targets. To further improve semantic image reconstruction quality, a weighted hybrid loss function is designed. This function integrates Mean Squared Error (MSE), L1-norm loss, Structural Similarity Index Measure (SSIM) loss, gradient loss, perceptual loss, and background suppression loss. These losses jointly optimize pixel accuracy, structural preservation, and fine-detail restoration. Through the joint constraints of multiple loss terms, the proposed system improves semantic region reconstruction and achieves high-quality semantic image transmission. Results and Discussions Simulation results validate the proposed LPUSC system in semantic extraction and end-to-end transmission. For semantic extraction, three schemes are compared: YOLO11s-seg, YOLO11s + Segment Anything Model (SAM), and YOLO11s + MobileSAM (Fig. 4 ). The results show that the detection-segmentation decoupled architecture achieves better semantic boundary localization accuracy. Combined with the quantitative analysis in Table 1 , the YOLO11s + MobileSAM scheme reduces resource use while maintaining high extraction accuracy. This confirms its suitability for resource-constrained UAV platforms. For end-to-end transmission, the semantic index vector transmission results (Fig. 5 ) show that the Bit Error Rate (BER) decreases monotonically as the Signal-to-Noise Ratio (SNR) increases in all three channel environments. The rural environment achieves the best performance, followed by the suburban and urban environments. These differences are mainly caused by variations in scatterer density and link blockage across environments. The proposed transmission network maintains stable BER under different Doppler frequencies, demonstrating its robustness under dynamic channel conditions. For semantic image transmission, the proposed weighted hybrid loss function shows good training stability (Fig. 6 ), and LPUSC consistently outperforms the Deep Joint Source-Channel Coding (DeepJSCC) and JPEG + Low-Density Parity-Check (LDPC) baselines across the full SNR range (Fig. 7 ). Specifically, LPUSC achieves SSIM and Peak Signal-to-Noise Ratio (PSNR) gains of 1.3% and 4.8% over DeepJSCC, respectively, and gains of 43% and 79.5% over JPEG + LDPC, respectively. These results indicate that the proposed personalized semantic image transmission network achieves high-quality reconstruction and remains robust to channel variations. Conclusions To improve the efficiency and flexibility of UAV image communication, this paper proposes LPUSC, a lightweight personalized semantic communication system. The system uses a dual-branch transmission architecture that integrates lightweight, high-precision object detection and semantic segmentation models. It enables personalized content transmission without interest pre-training. This design satisfies personalized user requirements while maintaining low computational and communication overhead. Simulation results show that the LPUSC system achieves stable and reliable semantic index interaction and outperforms the DeepJSCC and JPEG + LDPC baselines in semantic region reconstruction. The proposed system provides a useful reference for efficient UAV image semantic communication in 6G low-altitude intelligent networks.
Available online , doi: 10.11999/JEIT260430
Abstract:
Objective Aerial-Ground Person Re-IDentification (AG-PReID) aims to match the same person across Unmanned Aerial Vehicle (UAV) and ground-camera views. Compared with conventional same-platform person re-identification, this task faces larger cross-view appearance variation and more severe cross-domain distribution shifts. Under these conditions, identity-consistent cues are often weakened by strong viewpoint asymmetry and cross-domain appearance distortion. Existing methods mainly focus on feature extraction and cross-view representation alignment. However, the classification supervision branch still relies heavily on linear feature transformation, which limits its ability to model complex nonlinear discriminative relationships in high-dimensional feature spaces. A stronger nonlinear supervision mapping is therefore needed to better exploit high-order feature interactions and local discriminative variations. To address this issue, this paper proposes a Kolmogorov-Arnold Nonlinear Enhancement Module (KANEM). KANEM replaces the conventional fully connected feature transformation between backbone features and the linear classifier. It uses learnable nonlinear mappings to adaptively enhance features for more discriminative cross-view representation learning. Methods The backbone follows the View-Decoupled Transformer (VDT), which introduces an additional view token and performs layer-wise view decoupling. This design separates view-related factors from identity features and reduces representation bias between aerial and ground domains. Based on this framework, KANEM replaces the conventional fully connected feature transformation between backbone features and the linear classifier, thereby providing adaptive nonlinear mappings for feature enhancement. Specifically, KANEM consists of a base activation branch and a spline branch, which are stacked into cascaded function-mapping layers. This design enables more flexible nonlinear modeling than conventional linear or MultiLayer Perceptron (MLP)-based transformations. It allows the model to capture local nonlinear variations and complex correlations among feature dimensions. To improve discriminability and further separate identity and view information, the network is jointly optimized using identity classification loss, view classification loss, triplet loss, and orthogonality loss. KANEM is used only during training and is removed during inference, so no extra inference cost is introduced. Results and Discussions Comprehensive evaluations are conducted on the CARGO and AG-ReID datasets. The results show that the proposed method consistently performs better than the baseline model and existing state-of-the-art methods. On CARGO, the proposed method achieves 70.19%/63.16%/51.34% in Rank-1 accuracy, Mean Average Precision (mAP), and Mean Inverse Negative Penalty (mINP), respectively, under the overall ALL retrieval protocol. It also achieves 58.75%/53.27%/41.11% under the most challenging aerial-ground (A↔G) cross-view retrieval protocol (Table 1). On AG-ReID, the proposed method achieves the best performance under both retrieval protocols. It reaches 84.41%/76.21%/53.05% in Rank-1/mAP/mINP for aerial-to-ground (A→G) retrieval and 86.69%/77.99%/52.28% for ground-to-aerial (G→A) retrieval (Table 2 ). Ablation studies on CARGO further verify the effectiveness of KANEM. They show that KANEM achieves better overall performance than conventional linear transformation and MLP-based alternatives. This result indicates that the proposed nonlinear enhancement strategy is more suitable for supervision mapping in CARGO (Tables 3 and 4 ). In addition, integrating KANEM into other person re-identification tasks further demonstrates its potential generalization ability across different scenarios (Table 5 ). Parameter analysis shows that setting λ to 0.001 enables the model to better balance the complexity difference between view classification and identity classification (Fig. 2(a) ). When G and P are set to 5 and 3, respectively, the model effectively fits nonlinear variations in the feature space while preserving the smoothness and continuity of spline functions. This setting achieves effective nonlinear feature enhancement (Fig. 2(b) ～(d) ). The two-dimensional t-distributed Stochastic Neighbor Embedding (t-SNE) visualization shows that the enhanced features have higher intra-class compactness and better inter-class separability (Fig. 3 ). The top-5 retrieval comparisons further provide qualitative evidence that the proposed method improves ranking quality and retrieval robustness under all four retrieval protocols on CARGO. It promotes correct matches to higher positions and returns more relevant samples among the top-ranked results (Fig. 4 ). Conclusions This paper presents KANEM for AG-PReID. The proposed module is motivated by the large discrepancy between UAV and ground-camera views and by the limited capacity of linear feature transformation in the classification branch to capture complex nonlinear discriminative relationships. By replacing the conventional fully connected feature transformation between backbone output features and the linear classifier, KANEM provides a more flexible nonlinear supervision mechanism for cross-view representation learning. Through adaptive nonlinear enhancement, it better models complex feature interactions in high-dimensional spaces and strengthens the representation of cross-view consistency and fine-grained discriminative cues. Experimental results on CARGO and AG-ReID demonstrate the effectiveness of the proposed method, particularly in challenging scenarios with large view discrepancies. Future work will further refine the nonlinear mapping mechanism of KANEM and explore its use in more complex cross-view settings to improve model discriminability and generalization performance.
Available online , doi: 10.11999/JEIT260505
Abstract:
Objective This study addresses key challenges in Fifth-Generation (5G) millimeter-wave terminal antennas by designing a compact, high-performance dual-polarized array. Existing designs often face trade-offs among bandwidth, beam-scanning range, and integration complexity. To address these limitations, this paper proposes a differentially fed magnetoelectric dipole array. A stacked stripline-slot-stripline balun is used to enable efficient single-ended-to-differential conversion, and the array design is optimized. The objective is to realize an integrated solution with wideband operation, low cross-polarization, wide-angle beam scanning, and high integration density for practical 5G millimeter-wave applications. Methods A structured design method is adopted. First, a stacked differential balun based on a stripline-slot-stripline configuration is developed to achieve efficient single-ended-to-differential conversion. A single-polarized magnetoelectric dipole antenna element is then designed and integrated with the balun, and its performance is characterized. The design is further extended by orthogonally integrating two elements to form a dual-polarized unit, which is used to construct a 1×4 linear array. Iterative full-wave electromagnetic simulation and optimization are conducted to balance wideband impedance matching, high port isolation, stable wide-angle beam scanning, grating-lobe suppression, and mutual-coupling reduction. Results and Discussions The optimized 1×4 dual-polarized differentially fed magnetoelectric dipole antenna array uses an element spacing of 4.6 mm, corresponding to 0.4 free-space wavelength at 26 GHz. This spacing achieves a favorable balance between grating-lobe suppression and inter-element mutual-coupling reduction. The measured –10 dB reflection coefficient bandwidths are 25～29.4 GHz for the +45° polarization port and 25～27.7 GHz for the –45° polarization port (Fig. 20 ). The slight matching difference is attributed to the incomplete structural symmetry of the baluns under the two polarization modes (Fig. 13 ). At 26 GHz, both polarization modes provide a peak gain of 10.7～11 dBi and support ±60° wide-angle beam scanning, with main-lobe gain attenuation no greater than 3 dB (Fig. 21 ). The measured radiation performance agrees well with the simulated results. Minor deviations are mainly caused by the high dimensional sensitivity of millimeter-wave structures and small errors in fabrication and test assembly. The array also maintains stable low cross-polarization and high port isolation across the operating band. These results are achieved through equal-length feed lines, symmetric layout, ground-pad shielding, and metallized-via electromagnetic isolation (Fig. 16 ), which suppress mutual coupling and parasitic radiation and ensure consistent dual-polarized radiation performance. Conclusions This paper presents a dual-polarized magnetoelectric dipole antenna array with differential feeding for 5G millimeter-wave applications. By using a stacked stripline-slot-stripline balun and optimizing the radiating structure and array layout, the design achieves wide bandwidth, high gain, low cross-polarization, and wide-angle beam scanning. The differential balun enables efficient single-ended-to-differential conversion with good amplitude and phase balance across the target band. The implemented 1×4 array, with an optimized element spacing of 4.6 mm, achieves a simulated peak gain of 11 dBi at 26 GHz and supports ±60° beam scanning, with gain variation below 3 dB. The overall design verifies the feasibility of a differentially fed magnetoelectric dipole architecture for compact, high-performance 5G millimeter-wave terminal antenna modules. Future work may focus on larger array configurations and further integration with BeamForming Integrated Circuits (BFICs).
Available online , doi: 10.11999/JEIT260399
Abstract:
Significance Constrained by size, power consumption, and cost, emerging intelligent terminals often face excessive energy consumption and limited battery life. These limitations have become major bottlenecks to large-scale deployment. Compared with Fifth-Generation (5G) wireless networks, Sixth-Generation (6G) wireless networks are expected to enhance the Radio Access Network (RAN) architecture and move computing capability toward the RAN side. High-energy and compute-intensive Artificial Intelligence (AI) tasks that are originally executed by end devices can therefore be processed by the network. Through End-Edge Collaboration, emerging intelligent terminals can be upgraded toward lightweight design, low cost, and long battery life, thereby supporting the large-scale deployment of ubiquitous intelligence in 6G networks. Progress Current progress in Terminal Energy Consumption Optimization through 6G End-Edge Collaboration is reviewed, with emphasis on local execution, full offloading, and partial offloading. In local execution, User Equipment (UE) processes all tasks locally, which leads to high computing energy consumption. In full offloading, all tasks are transferred to the RAN. This reduces terminal-side computing energy consumption but can increase transmission energy consumption, especially under poor channel conditions. Partial offloading combines the benefits of both modes and optimizes energy consumption according to real-time network conditions. For partial offloading, four representative optimization techniques are summarized. (1) Feature Extraction and Filtering. Semantic encoding and information extraction are performed at the UE, and only task-relevant data are transmitted to the RAN. This reduces redundant data transmission and lowers transmission energy consumption. (2) Split Offloading. A large Deep Neural Network (DNN) is divided into layers according to its structure. Simpler shallow layers are processed at the UE, whereas more complex deep layers are offloaded to the RAN. This method balances terminal-side and RAN-side computational loads through End-Edge Collaborative Inference. (3) Model Lightweighting. Model complexity is reduced through pruning, quantization, and knowledge distillation, which lowers computational overhead while maintaining task performance. (4) Incremental Inference. Only changed data or updated features are processed, while historical computations are reused. This reduces redundant computation. Together, these techniques improve terminal performance and energy efficiency within the 6G End-Edge Collaboration framework. Conclusions This paper systematically reviews Terminal Energy Consumption Optimization for 6G End-Edge Collaboration. It summarizes the functional evolution of enhanced RAN, constructs an End-Edge Collaborative service framework for Communication-Computing Integration, and establishes a theoretical model of terminal computing energy consumption and transmission energy consumption. The composition and influencing factors of energy consumption under different offloading modes are clarified. Key energy optimization technologies, including Feature Extraction and Filtering, Split Offloading, Model Lightweighting, and Incremental Inference, are then discussed. To address energy consumption fluctuations caused by dynamic wireless channels, the paper proposes energy optimization mechanisms based on Adaptive Semantic Compression, Dynamic Split Offloading, Adaptive Model Pruning, and Incremental Inference. These mechanisms maintain a dynamic balance between energy optimization and task performance. Using embodied intelligent robot video understanding as a typical application scenario, a test platform is developed to verify the effectiveness of the proposed mechanisms. Current challenges and future research directions are also analyzed. Prospects Although End-Edge Collaborative energy-saving technologies have achieved initial progress, practical deployment still faces challenges in real network environments, dynamic wireless channels, and large-scale user access. Future research should examine the trade-off between optimization overhead and system robustness. It should also study dynamic communication-computing resource substitution modeling in stochastic resource environments, multi-user collaboration strategies, and global energy-efficiency optimization. As the technology matures, standardization and engineering implementation of End-Edge Collaborative energy-saving frameworks will become critical to the large-scale adoption of 6G applications. Future studies should further integrate algorithm design with network architecture, enabling practical deployment of low-power and high-efficiency intelligent communication systems.
Available online , doi: 10.11999/JEIT260218
Abstract:
Objective Reliable inference at the network edge is indispensable for 6G-enabled ubiquitous AI, yet the deployment of large language models (LLMs) in such environments remains a cornerstone challenge. In resource-constrained edge settings, single-LLM inference often proves unreliable due to knowledge limitations and inherent biases, severely hampering real-world deployment. Collaborative inference leveraging multiple heterogeneous LLMs emerges as a promising remedy to boost robustness, but it introduces nontrivial hurdles under stringent latency and energy budgets, especially when wireless channel conditions and query content vary unpredictably. These challenges include the need for dynamic sequential decision-making for LLM selection and resource allocation, the fundamental paradigm mismatch between bit-level reliability protocols and semantic-level error correction, and the lack of adaptive mechanisms to align and fuse disparate LLM outputs effectively. To fill these critical gaps, this paper presents a novel framework that fundamentally reinterprets collaborative inference as a semantic-driven, closed-loop process, thereby transitioning from conventional bit-retransmission to semantic-retransmission and offering a practical path toward reliable 6G edge intelligence. Methods In response to these critical challenges, we propose the Gating Adaptive Repeat Query (G-ARQ) framework. Its core innovation is the Semantic-Space Alignment and Error-Guided Retransmission (SEMAR) mechanism. SEMAR first aligns the token-level probability distributions from heterogeneous LLMs into a unified semantic space using relative representation, enabling comparable outputs. It then models the collaborative process probabilistically, explicitly capturing error dependencies among models, and uses an Expectation-Maximization (EM) algorithm to infer a latent error direction, which guides the selection of the next LLM for query retransmission, steering it towards outputs orthogonal to previous errors. To jointly optimize the LLM gating and uplink power allocation under communication constraints without requiring explicit system dynamics—often unavailable in practice—we design a black-box trajectory optimizer. This optimizer formulates the sequential decision problem as sampling from a target distribution that encodes dynamic feasibility, optimality, and constraints. It employs a diffusion-based sampling process with a model-guided prior and Monte Carlo estimation to generate near-optimal policy trajectories that satisfy hard latency and energy limits. Results and Discussions To evaluate the practical viability of G-ARQ under realistic edge conditions, simulations are conducted in a scenario with five base stations hosting five heterogeneous 7B-parameter LLMs (Mistral-7B, Vicuna-7B, Nous-Capybala-7B, Gemma-7B, and Llama-2-7B). The user equipment (UE) performs a question-answering task evaluated on a mixed SQuAD and TriviaQA dataset. Component-level evaluations, each designed to isolate the contribution of a single innovation, validate the effectiveness of every key element. The error-guided gating of SEMAR, compared to a Top-k gating baseline, improves accuracy by 0.23 % on average, and its dynamic weight ensemble contributes an additional 0.7 % gain (Fig. 3). The black-box trajectory optimizer, which operates without any explicit channel model, achieves accuracy close to that of the unconstrained model-greedy strategy while ensuring strict latency constraints (Fig. 4). The convergence of the optimizer is verified by tracking the evolution of \begin{document}$ J(\boldsymbol{S}) $\end{document} \begin{document}$ {K}_{\max }=1 $\end{document} \begin{document}$ {K}_{\max }=2 $\end{document} Conclusions This paper proposed G-ARQ framework, an innovative closed-loop framework that transforms collaborative edge inference into a semantics-guided retransmission process. By introducing SEMAR for error-based alignment and selection of heterogeneous LLMs, and employing a black-box trajectory optimizer for joint model selection and power allocation, the framework achieves up to a 2.2% accuracy improvement under strict resource constraints. The results validate G-ARQ as an effective and practical approach toward reliable and efficient 6G edge intelligence.
Available online , doi: 10.11999/JEIT260494
Abstract:
Objective Traditional roadheader bearing fault diagnosis methods often struggle with high-dimensional and nonlinear multi-sensor data. They also fail to effectively perceive cross-modal, multi-scale fault information or integrate local and global structural features. To address these limitations, this paper proposes a Cross-modal Kernel Fusion-sphere Space Learning (CKFSL) method. By perceiving cross-modal multi-scale fault information, CKFSL extracts highly discriminative features from roadheader bearing cross-modal fault samples and improves diagnostic accuracy. Methods CKFSL first maps roadheader bearing cross-modal fault samples into a high-dimensional kernel space through implicit transformation. Dual extremal point anchoring and polar neighbor allocation mechanisms are then used to capture fault sample clusters with similar isomorphic information, forming kernel fusion-spheres. An adaptive binary partitioning strategy is designed according to the geometric span of internal fault samples. This strategy tightens isomorphic boundaries, constructs micro-neighbor kernel fusion-spheres, and achieves highly isomorphic manifold aggregation at the microscopic scale. A micro-neighbor kernel fusion-sphere space is further formed to re-evaluate local isomorphism (Fig. 1 ). To characterize wide-area topological correlations, a wide-area topological isomorphism constraint is proposed. This constraint constructs a wide-area dynamic isomorphism graph among micro-neighbor kernel fusion-spheres (Fig. 1 ). Finally, an objective optimization function is formulated within the space learning framework. It integrates local manifold isomorphism and wide-area topological correlations of roadheader bearing cross-modal fault samples, as shown in the CKFSL diagnostic flowchart (Fig. 2 ). The analytical solution for spatial projection is theoretically derived to obtain discriminative cross-modal kernel fusion-sphere space isomorphic features from roadheader bearing cross-modal fault samples. Results and Discussions CKFSL is first validated on the self-built AUST roadheader bearing cross-modal fault dataset, with the experimental platform shown in Fig. 3 . The average recognition rates obtained with increasing numbers of training fault samples are shown in Fig. 4. On the AUST dataset, CKFSL achieves a recognition rate of 99.49% with only 70 training fault samples and reaches 100% as the number of training fault samples increases. Table 1 summarizes the standard deviations under different training fault sample sizes. The results show that CKFSL has the lowest standard deviation and stronger robustness than the other seven comparison algorithms. Three-dimensional fault feature distributions are shown in Fig. 5 . The results confirm that CKFSL effectively separates highly overlapping fault samples into different clusters and reduces the boundary confusion observed in the comparison algorithms. To verify generalization capability, CKFSL is further evaluated on the public Paderborn dataset, with the experimental setup shown in Fig. 6 . As shown in Fig. 7 and Fig. 8 , CKFSL achieves a 100% average recognition rate across four complex fault categories. It also outperforms the comparison algorithms, which have difficulty exceeding an 85% recognition rate for the F4 fault category. Conclusions CKFSL effectively addresses the inability of traditional roadheader bearing fault diagnosis methods to perceive complex multi-scale fault information. By using the wide-area dynamic isomorphism graph learned in the micro-neighbor kernel fusion-sphere space, CKFSL integrates local manifold isomorphism with wide-area topological correlations of roadheader bearing cross-modal fault samples. This process enables CKFSL to extract highly discriminative cross-modal kernel fusion-sphere space isomorphic features. It improves the accuracy of roadheader bearing fault diagnosis and supports the reliability and continuous operation of roadheaders.
Available online , doi: 10.11999/JEIT260194
Abstract:
Objective With the deployment of intelligent connected vehicle-road-cloud cooperative systems, roadside infrastructure is evolving from traffic-state sensing units into intelligent decision-support platforms for multi-vehicle interaction analysis and dynamic risk inference. In highway and urban freeway scenarios, traffic operation is affected not only by the kinematic responses of individual vehicles but also by lane-changing intentions, car-following competition, and conflict propagation under local interactions. From a roadside perspective, a unified framework is therefore needed to continuously represent traffic risk, reveal its spatiotemporal evolution, and couple risk information with behavior decision-making and trajectory planning. Existing car-following models, such as the Optimal Velocity Model (OVM), Full Velocity Difference (FVD) Model, and Intelligent Driver Model (IDM), can describe speed-spacing evolution. However, these models mainly focus on longitudinal interactions and usually embed risk implicitly in safety-distance or acceleration constraints. They cannot explicitly characterize the coupling between longitudinal following and lateral lane changing, nor can they provide a continuous risk representation suitable for regional traffic assessment. Although Artificial Potential Field (APF) methods and Model Predictive Control (MPC) methods can improve trajectory safety, existing studies still lack a unified mechanism that links risk assessment, behavior decision-making, and motion planning. In addition, discrete behavior choices and continuous control actions are difficult to process efficiently within a single optimization framework. Methods An intelligent connected traffic risk-field model oriented toward vehicle-road cooperation is first established. The model integrates vehicle-interaction risk, lane-marking constraint risk, and road-boundary repulsive risk (Fig. 1 ). In the vehicle-interaction layer, motion-state-induced risk is formulated by considering the relative speed and relative orientation between the ego vehicle and surrounding vehicles. Distance-induced risk is modeled to reflect attenuation as separation distance increases (Fig. 2(a) –(b) ). To represent the stronger influence of forward hazards than lateral and rear hazards, a directional non-uniformity coefficient is used. This coefficient adjusts the angular attenuation of field strength and enables anisotropic spatial risk representation around the vehicle. In the road-constraint layer, lane markings and road boundaries are modeled separately. The total driving risk field is obtained by weighting and combining the lane-marking field, road-boundary field, and multi-vehicle interaction field (Fig. 2(e) –(f) ). Based on this representation, a dual-MPC hierarchical decision and motion-planning architecture is designed (Fig. 3 ). In each control cycle, the upper layer evaluates candidate behavior modes according to the vehicle state, surrounding traffic state, and dynamic risk field, and then outputs a unique behavior mode. The lower layer activates the corresponding control branch. When lane keeping is selected, longitudinal speed-planning MPC is used. When lane changing is selected, lane-change trajectory-planning MPC is activated under road-boundary, lane-marking, and safe-gap constraints. Results and Discussions The proposed framework reconstructs microscopic traffic risk evolution under different datasets and time scales. In the HighD highway scenario, when the slicing interval is 0.4 s, the local evolution of following and lane-changing interactions is captured in detail. This includes the process in which the ego vehicle initially follows a preceding vehicle and then starts changing to the adjacent lane (Fig. 4(a) –(d) ). When the interval is increased to 1.0 s, a wider spatiotemporal interaction range becomes visible, and lane-change completion and the subsequent return maneuver are identified more clearly (Fig. 4(e) –(h) ). In the NGSIM scenario, a smaller interval provides a finer description of the lane-change disturbance process. By contrast, a larger interval reveals the wider reconstruction of interaction relationships among the original lane, target lane, and surrounding vehicles (Fig. 4(i) –(p) ). These results indicate that the proposed roadside-oriented risk field can describe both local interaction details and larger-scale evolution trends, depending on the selected reconstruction interval. Sensitivity experiments further confirm the role of the directional non-uniformity coefficient. As this coefficient increases, forward risk concentration becomes stronger, local peak risk increases, and the coverage of high-risk regions decreases (Table 2 ). This finding shows that the coefficient effectively regulates anisotropic field distribution. Comparative experiments with IDM, OVM, FVD, and APF show that the proposed method performs better in most representative scenarios and error metrics (Fig. 5 , Table 3 ). In the lateral cut-in scenario, its advantage lies in the early representation of lateral intrusion risk, which enables the behavior decision layer to anticipate conflict and the motion-planning layer to generate continuous evasive actions. In congested scenarios, the superposition of forward congestion risk, lateral neighboring-vehicle influence, and road-boundary constraints allows the dual-MPC controller to evaluate safety and feasibility simultaneously in local space. Conclusions A unified framework for roadside-oriented traffic-risk modeling and behavior-driven trajectory planning is developed. By integrating multi-vehicle interaction risk, lane-marking constraint risk, and road-boundary repulsive risk into a continuously evolving dynamic risk field, the spatial quantification of multi-vehicle interaction risk is realized. The directional non-uniformity coefficient further enables asymmetric risk perception modeling in forward, lateral, and rear directions. On this basis, a dual-MPC hierarchical architecture is constructed to couple behavior decision-making with motion planning, so that lane-keeping and lane-changing behaviors can be adaptively selected and optimized under a unified risk-driven mechanism. Experiments based on HighD and NGSIM datasets show that the proposed method can effectively characterize the spatiotemporal evolution of traffic-risk fields and outperform representative comparison models in most typical scenarios and error metrics.
Available online , doi: 10.11999/JEIT260090
Abstract:
Objective Low-altitude edge computing networks provide flexible computing services and extended coverage for user equipment. However, quality of service is often degraded by uncertainty in task data size and by Unmanned Aerial Vehicle (UAV) position jitter caused by environmental disturbances. Existing robust methods commonly rely on deterministic uncertainty sets, which tend to be conservative and cannot accurately describe the stochastic distribution of task demands. To address these challenges, a robust energy minimization framework is proposed for multi-UAV-assisted Mobile Edge Computing (MEC) networks. The objective is to minimize the weighted sum of system energy consumption. This is achieved by developing a joint optimization model that coordinates UAV flight trajectories, task splitting decisions, and computation and communication resource allocation. The model explicitly accounts for the dual uncertainties of task data size and UAV trajectory. Methods To handle the nonconvexity and strong coupling among optimization variables, the problem is first modeled as a Markov Decision Process (MDP). A comprehensive state space is defined to characterize real-time system dynamics, and a continuous action space is designed for trajectory control and resource management. A Distributionally Robust Optimization Soft Actor-Critic (DRO-SAC) algorithm is then developed to solve the MDP. In this framework, an ambiguity set based on the L1-norm distance is constructed to characterize the distributional uncertainty of the task demand distribution. A maximum-entropy reinforcement learning mechanism is used to learn an optimal policy under the worst-case distribution within the ambiguity set. In this way, UAV trajectories, task splitting, and computation and communication resource allocation are jointly optimized to improve system robustness under dynamic environmental fluctuations. Results and Discussions The performance of the proposed DRO-SAC algorithm is evaluated through simulations. DRO-SAC achieves faster convergence and higher rewards than Deep Deterministic Policy Gradient (DDPG) and Proximal Policy Optimization (PPO) algorithms (Fig. 3 ). For energy consumption, the proposed method consistently achieves higher efficiency under different user densities (Fig. 4 ). The robustness of the system against position errors is also verified, with energy fluctuations kept at a low level (Fig. 5 ). Dynamic trajectory adjustment further confirms that the proposed method can provide effective user coverage while reducing system energy consumption (Fig. 6 ). Conclusions A DRO-SAC-based joint optimization framework is proposed to address uncertainty in task data size and UAV position jitter in multi-UAV-assisted MEC networks. By constructing an ambiguity set for the task demand distribution and optimizing the worst-case expected objective, the proposed method mitigates the limitations of traditional deterministic models in dynamic environments. Weighted system energy consumption is minimized while latency and safety constraints are satisfied. Simulation results demonstrate that the proposed scheme achieves stable convergence and high energy efficiency, even when communication and computation resources are limited and environmental parameters fluctuate strongly.
Available online , doi: 10.11999/JEIT260377
Abstract:
Objective With the rapid growth of the Internet of Things, 5G/6G, and satellite Internet, resource-constrained devices increasingly require high-quality random numbers for key generation, authentication, masking, and other security functions. Although pseudo-random number generators are efficient, their outputs may be predictable once the seed or internal state is compromised. True random number generators (TRNGs) offer a hardware root of trust by extracting entropy from physical randomness, but many existing designs rely on multiple entropy sources or complex post-processing, leading to increased area and power consumption. To address this issue, this paper proposes a lightweight TRNG based on chain-coupled oscillation rings for high-quality randomness with very low FPGA overhead. Methods Starting from the state evolution of a Galois oscillation ring (GARO), this work demonstrates that ideal matched-delay conditions can result in periodic and predictable oscillation. However, in practical circuits, delay mismatch, jitter, and process variation disturb the ideal evolution and can be exploited as entropy sources. On this basis, a compact delay-feedback XOR ring is proposed to enhance state uncertainty, introduce feedback competition, and improve randomness through inter-stage delay differences. In addition, a second-order oscillation ring is incorporated to eliminate the all-zero stop state and provide continuous excitation. Multiple rings are then chain-coupled, enabling adjacent rings to mutually interfere with one another and thereby generate stronger irregular oscillations. The proposed design is modeled in MATLAB and implemented on a Xilinx Artix-7 FPGA. Finally, we evaluate its performance by NIST SP 800-22, NIST SP 800-90B, bias, autocorrelation, and voltage-temperature robustness tests. Results and Discussions Simulation confirms that the proposed structure avoids stable periodic locking and produces sustained irregular oscillation. Experimental results show that the TRNG passes all NIST SP 800-22 tests and achieves an average minimum entropy of 0.9936 in NIST SP 800-90B test, outperforming conventional RO and GARO-based TRNGs under similar conditions. The measured bias is only 0.0228 %, and the autocorrelation remains well below the threshold, indicating excellent statistical independence. The design also maintains high entropy over temperatures from 0 °C to 80 °C and supply voltages from 0.9 V to 1.1 V. Implemented on Artix-7, our proposed TRNG achieves 200 Mbps throughput using only 11 LUTs and 4 DFFs, with 0.108 W power consumption. Conclusions This paper presents a lightweight chain-coupled oscillation-ring TRNG that exploits delay mismatch, phase disturbance, and feedback competition to generate high-quality physical randomness. The theoretical analysis clarifies how practical nonidealities transform ideal periodic oscillation into irregular oscillation, providing a design basis for compact oscillator-based entropy sources. By combining delay-feedback XOR rings with chain-coupled mutual disturbance and continuous excitation, the proposed design enhances entropy while avoiding excessive hardware overhead and complex post-processing. FPGA implementation and statistical evaluations verify high entropy, low bias, and high randomness under voltage and temperature variations. Therefore, the proposed TRNG achieves high randomness quality and high throughput while effectively reducing hardware overhead, making it suitable for resource-constrained security applications such as IoT terminals, lightweight cryptographic modules, and embedded authentication systems.
Available online , doi: 10.11999/JEIT251300
Abstract:
Objective In the Artificial Intelligence of Things (AIoT), Edge Servers (ESs) provide intelligent content generation services to AIoT devices by utilizing cached Artificial Intelligence Generated Content (AIGC) models. However, the limited computing resources and caching capacity of ESs make it difficult to support the large-scale caching demands of heterogeneous AIGC services. To address this issue, a communication, computation, and caching resource collaboration scheme is proposed based on a combined cloud-edge and edge-edge collaborative framework. The scheme considers three representative AIGC services: lightweight AIGC services, computation-intensive AIGC services, and preprocessing-based AIGC services. The objective is to minimize the total AIGC service latency through joint optimization of transmit power, computing resource allocation, model caching strategies, and offloading decisions. Methods Communication, computation, and caching resource collaboration for heterogeneous AIGC services is investigated. First, an AIGC service-oriented AIoT system model is established to incorporate both cloud-edge and edge-edge collaboration. An optimization problem is then formulated to minimize the total latency of AIGC services through joint optimization of transmit power, computing resource allocation, model caching strategies, and offloading decisions. Because the formulated problem is non-convex, an Alternating Optimization (AO) algorithm is proposed. The original problem is decomposed into three subproblems. These subproblems are solved using the Successive Convex Approximation (SCA) method, Karush-Kuhn-Tucker (KKT) conditions, and an improved Harris Hawks Optimization (HHO) algorithm. Results and Discussions Simulation experiments compare the proposed joint optimization scheme with three baseline methods: Particle Swarm Optimization (PSO), fixed resource allocation, and random offloading and caching. First, the convergence of the proposed AO algorithm is verified (Fig. 2 ). The results show that the algorithm converges rapidly within a limited number of iterations across different subproblems. Second, increasing transmission bandwidth significantly reduces the total AIGC service latency (Fig. 3 ). This occurs because each device obtains more bandwidth resources for task transmission, and the ES can allocate more bandwidth to deliver generated content in the downlink. Furthermore, the total AIGC service latency decreases as the ES storage capacity increases for all schemes (Fig. 4 ). Greater storage capacity enables the ES to store more AIGC models, which reduces the transmission delay between the ES and the cloud server. Moreover, when the required floating-point operations per bit increase, the total AIGC service latency rises significantly across all schemes (Fig. 5 ). Finally, the total AIGC service latency decreases as the maximum transmit power of the Base Station (BS) increases (Fig. 6 ). This occurs because higher BS transmit power improves the downlink signal-to-noise ratio, which increases the downlink transmission rate and reduces overall service latency. The proposed scheme demonstrates better performance than the baseline schemes, particularly under high computational demand. Conclusions Communication, computation, and caching resource collaboration for heterogeneous AIGC services is investigated. The objective is to minimize total AIGC service latency through joint optimization of the transmit power of AIoT devices and BSs, computing resource allocation, AIGC model deployment, and service offloading decisions under computation and caching resource constraints. Because the formulated problem is a mixed-integer nonlinear programming problem, an efficient AO algorithm is developed. The original optimization problem is decomposed into three subproblems, which are solved using the SCA algorithm, KKT conditions, and the HHO algorithm, respectively. Simulation results show that the proposed algorithm reduces the total AIGC service latency compared with the baseline schemes.
Available online , doi: 10.11999/JEIT251244
Abstract:
Objective In complex and dynamically changing deep-sea environments, the performance of underwater acoustic communications shows substantial variability. Feedback-based channel estimation and parameter adaptation are impractical in long-range scenarios because platform constraints prevent reliable feedback channels and the slow propagation of sound introduces significant delay. In typical long-range systems, environmental dynamics are often ignored and communication parameters are selected heuristically, which frequently leads to mismatches with actual channel conditions and causes communication failures or reduced efficiency. Predictive methods able to assess performance in advance and support feed-forward parameter adjustment are therefore required. This study proposes a deep-learning-based framework for performance analysis and rapid prediction of long-range underwater acoustic communications under uncertain environmental conditions to enable efficient and reliable parameter–channel matching without feedback. Methods A feed-forward method for underwater acoustic communication performance analysis and rapid prediction is developed using deep-learning-based sound-field uncertainty estimation. A neural network is first used to estimate probability distributions of Transmission Loss (TL PDFs) at the receiver under dynamic environments. TL PDFs are then mapped to probability distributions of the Signal-to-Noise Ratio (SNR PDFs), enabling communication performance evaluation without real-time feedback. Statistical channel capacity and outage capacity are analyzed to characterize the theoretical upper limits of achievable rates in dynamic conditions. Finally, by integrating the SNR distribution with the bit-error-rate characteristics of a representative deep-sea single-carrier communication system under the corresponding channel, a rate–reliability prediction model is constructed. This model estimates the probability of reliable communication at different data rates and serves as a practical tool for forecasting link performance in highly dynamic and feedback-limited underwater acoustic environments. Results and Discussions The method is validated using simulation data and sea trial data. The TL PDFs predicted by the deep learning model show strong consistency with the traditional Monte Carlo (MC) method across multiple receiver locations (Fig. 6 ). Under identical computational settings, deep-learning-based TL PDF prediction reduces computation time by 2\begin{document}$ \sim $\end{document} Fig. 7 and Fig. 8 ). The rate–reliability curves derived from the deep-learning-based TL PDFs are highly consistent with MC-based results. In the high sound-intensity region, prediction errors for reliable communication probabilities across data rates range from 0.1% to 3%, and in the low sound-intensity region errors are approximately 0.3% to 5% (Fig. 12 ). Sea trial results further indicate that predicted rate–reliability performance agrees well with measured data. In the convergence zone, deviations between predicted and measured reliability probabilities at each rate range from 0.9% to 4%, and in the shadow zone from 1% to 9% (Fig. 18 ). Under a 90% reliability requirement, the maximum achievable rates predicted by the method match the measurements in both the convergence and shadow zones, demonstrating accuracy and practical applicability in complex channel environments. Conclusions A deep-learning-based framework for performance analysis and rapid prediction of long-range underwater acoustic communications in uncertain deep-sea environments is developed and validated. The framework builds a chained mapping from environmental parameters to TL PDFs, SNR PDFs, and communication performance metrics, enabling quantitative capacity assessment under dynamic ocean conditions. Predictive “rate–reliability’’ profiles are obtained by integrating probabilistic propagation characteristics with the performance of a representative deep-sea single-carrier system under the corresponding channel, providing guidance for parameter selection without feedback. Sea trial results confirm strong agreement between predicted and measured performance. The proposed approach offers a technical pathway for feed-forward performance analysis and dynamic adaptation in long-range deep-sea communication systems, and can be extended to other communication scenarios in dynamic ocean environments.
Available online , doi: 10.11999/JEIT251221
Abstract:
Objective Multimodal Federated Learning (MFL) uses complementary information from multiple modalities, yet in wireless edge networks it is restricted by limited energy and frequent missing modalities because many clients store only images or only reports. This study presents Cross-modal Retrieval Enhanced Energy-efficient Multimodal Federated Learning (CREEMFL), which applies selective completion and joint communication–computation optimization to reduce training energy under latency and wireless constraints. Methods CREEMFL completes part of the incomplete samples by querying a public multimodal subset, and processes the remaining samples through zero padding. Each selected user downloads the global model, performs image-to-text or text-to-image retrieval, conducts local multimodal training, and uploads model updates for aggregation. An energy–delay model couples local computation and wireless communication and treats the required number of global rounds as a function of retrieval ratios. Based on this model, an energy minimization problem is formulated and solved using a two-layer algorithm with an outer search over retrieval ratios and an inner optimization of transmission time, Central Processing Unit (CPU) frequency, and transmit power. Results and Discussions Simulations on a single-cell wireless MFL system show that increasing the ratio of completing text from images improves test accuracy and reduces total energy. In contrast, a large ratio of completing images from text provides limited accuracy gain but increases energy consumption (Fig. 3 , Fig. 4 ). Compared with four representative baselines, CREEMFL achieves shorter completion time and lower total energy across a wide range of maximum average transmit powers (Fig. 5 , Fig. 6 ). For CREEMFL, increased system bandwidth further reduces completion time and energy consumption (Fig. 7 , Fig. 8 ). Under different user modality compositions, CREEMFL also attains higher test accuracy than local training, zero padding, and cross-modal retrieval without energy optimization (Fig. 9 ). Conclusions CREEMFL integrates selective cross-modal retrieval and joint communication–computation optimization for energy-efficient MFL. By treating retrieval ratios as variables and modeling their effect on global convergence rounds, it captures the coupling between per-round costs and global training progress. Simulations verify that CREEMFL reduces training completion time and total energy while preserving classification accuracy in resource-constrained wireless edge networks.
Available online , doi: 10.11999/JEIT250352
Abstract:
This study presents a breakthrough in addressing NP-complete problems using a newly developed Electronic Probe Computer (EPC60). The system employs a hybrid serial–parallel computational model and performs large-scale parallel operations through seven probe operators. In benchmark tests on 3-coloring problems in graphs with 2,000 vertices, EPC60 achieves 100% accuracy, outperforming the mainstream solver Gurobi, which succeeds in only 6% of cases. Computation time is reduced from 15 days to 54 seconds. The system demonstrates high scalability and offers a general-purpose solution for complex optimization problems in areas such as supply chain management, finance, and telecommunications. Objective NP-complete problems pose a fundamental challenge in computer science. As problem size increases, the required computational effort grows exponentially, making it infeasible for traditional electronic computers to provide timely solutions. Alternative computational models have been proposed, with biological approaches—particularly DNA computing—demonstrating notable theoretical advances. However, DNA computing systems continue to face major limitations in practical implementation. Methods Computational Model: EPC is based on a non-Turing computational model in which data are multidimensional and processed in parallel. Its database comprises four types of graphs, and the probe library includes seven operators, each designed for specific graph operations. By executing parallel probe operations, EPC efficiently addresses NP-complete problems.Structural Features:EPC consists of four subsystems: a conversion system, input system, computation system, and output system. The conversion system transforms the target problem into a graph coloring problem; the input system allocates tasks to the computation system; the computation system performs parallel operations via probe computation cards; and the output system maps the solution back to the original problem format.EPC60 features a three-tier hierarchical hardware architecture comprising a control layer, optical routing layer, and probe computation layer. The control layer manages data conversion, format transformation, and task scheduling. The optical routing layer supports high-throughput data transmission, while the probe computation layer conducts large-scale parallel operations using probe computation cards. Results and Discussions EPC60 successfully solved 100 instances of the 3-coloring problem for graphs with 2,000 vertices, achieving a 100% success rate. In comparison, the mainstream solver Gurobi succeeded in only 6% of cases. Additionally, EPC60 rapidly solved two 3-coloring problems for graphs with 1,500 and 2,000 vertices, which Gurobi failed to resolve after 15 days of continuous computation on a high-performance workstation.Using an open-source dataset, we identified 1,000 3-colorable graphs with 1,000 vertices and 100 3-colorable graphs with 2,000 vertices. These correspond to theoretical complexities of O(1.3289n) for both cases. The test results are summarized in Table 1.Currently, EPC60 can directly solve 3-coloring problems for graphs with up to n vertices, with theoretical complexity of at least O(1.3289n).On April 15, 2023, a scientific and technological achievement appraisal meeting organized by the Chinese Institute of Electronics was held at Beijing Technology and Business University. A panel of ten senior experts conducted a comprehensive technical evaluation and Q&A session. The committee reached the following unanimous conclusions:1. The probe computer represents an original breakthrough in computational models.2. The system architecture design demonstrates significant innovation.3. The technical complexity reaches internationally leading levels.4. It provides a novel approach to solving NP-complete problems.Experts at the appraisal meeting stated, “This is a major breakthrough in computational science achieved by our country, with not only theoretical value but also broad application prospects.” In cybersecurity, EPC60 has also demonstrated remarkable potential. Supported by the National Key R&D Program of China (2019YFA0706400), Professor Xu Jin’s team developed an automated binary vulnerability mining system based on a function call graph model. Evaluation of the system using the Modbus Slave software showed over 95% vulnerability coverage, far exceeding the 75 vulnerabilities detected by conventional depth-first search algorithms. The system also discovered a previously unknown flaw, the “Unauthorized Access Vulnerability in Changyuan Shenrui PRS-7910 Data Gateway” (CNVD-2020-31406), highlighting EPC60’s efficacy in cybersecurity applications.The high efficiency of EPC60 derives from its unique computational model and hardware architecture. Given that all NP-complete problems can be polynomially reduced to one another, EPC60 provides a general-purpose solution framework. It is therefore expected to be applicable in a wide range of domains, including supply chain management, financial services, telecommunications, energy, and manufacturing. Conclusions The successful development of EPC offers a novel approach to solving NP-complete problems. As technological capabilities continue to evolve, EPC is expected to demonstrate strong computational performance across a broader range of application domains. Its distinctive computational model and hardware architecture also provide important insights for the design of next-generation computing systems.
Available online , doi: 10.11999/JEIT221203
Abstract:
To overcome the limitation of the Federated Learning (FL) when the data and model of each client are all heterogenous and improve the accuracy, a personalized Federated learning algorithm with Collation game and Knowledge distillation (pFedCK) is proposed. Firstly, each client uploads its soft-predict on public dataset and download the most correlative of the k soft-predict. Then, this method apply the shapley value from collation game to measure the multi-wise influences among clients and quantify their marginal contribution to others on personalized learning performance. Lastly, each client identify it’s optimal coalition and then distill the knowledge to local model and train on private dataset. The results show that compared with the state-of-the-art algorithm, this approach can achieve superior personalized accuracy and can improve by about 10%.
To overcome the limitation of the Federated Learning (FL) when the data and model of each client are all heterogenous and improve the accuracy, a personalized Federated learning algorithm with Collation game and Knowledge distillation (pFedCK) is proposed. Firstly, each client uploads its soft-predict on public dataset and download the most correlative of the k soft-predict. Then, this method apply the shapley value from collation game to measure the multi-wise influences among clients and quantify their marginal contribution to others on personalized learning performance. Lastly, each client identify it’s optimal coalition and then distill the knowledge to local model and train on private dataset. The results show that compared with the state-of-the-art algorithm, this approach can achieve superior personalized accuracy and can improve by about 10%.
Available online , doi: 10.11999/JEIT210265
Abstract:
The application of Frequency Diverse Array and Multiple Input Multiple Output (FDA-MIMO) radar to achieve range-angle estimation of target has attracted more and more attention. The FDA can simultaneously obtain the degree of freedom of transmitting beam pattern in angle and range. However, its performance is degraded due to the periodicity and time-varying of the beam pattern. Therefore, an improved Estimating Signal Parameter via Rotational Invariance Techniques (ESPRIT) algorithm to estimate the target’s parameters based on a new waveform synthesis model of the Time Modulation and Range Compensation FDA-MIMO (TMRC-FDA-MIMO) radar is proposed. Finally, the proposed method is compared with identical frequency increment FDA-MIMO radar system, logarithmically increased frequency offset FDA-MIMO radar system and MUltiple SIgnal Classification (MUSIC) algorithm through the Cramer Rao lower bound and root mean square error of range and angle estimation, and the excellent performance of the proposed method is verified.
The application of Frequency Diverse Array and Multiple Input Multiple Output (FDA-MIMO) radar to achieve range-angle estimation of target has attracted more and more attention. The FDA can simultaneously obtain the degree of freedom of transmitting beam pattern in angle and range. However, its performance is degraded due to the periodicity and time-varying of the beam pattern. Therefore, an improved Estimating Signal Parameter via Rotational Invariance Techniques (ESPRIT) algorithm to estimate the target’s parameters based on a new waveform synthesis model of the Time Modulation and Range Compensation FDA-MIMO (TMRC-FDA-MIMO) radar is proposed. Finally, the proposed method is compared with identical frequency increment FDA-MIMO radar system, logarithmically increased frequency offset FDA-MIMO radar system and MUltiple SIgnal Classification (MUSIC) algorithm through the Cramer Rao lower bound and root mean square error of range and angle estimation, and the excellent performance of the proposed method is verified.
Available online , doi: 10.11999/JEIT251357
Abstract:
Objective In recent years, next Point Of Interest (POI) recommendation has played an increasingly important role in Location-Based Social Networks (LBSNs). However, existing Graph Representation Learning (GRL)-based recommendation methods have struggled to balance node distributions across different domains (i.e., node types) effectively and have often overlooked feature differences among heterogeneous relations. Thus, complex semantic dependencies in contextual information cannot be fully captured when users’ temporal preference patterns are modeled. Methods To address these issues, a next POI recommendation method based on Semantic Relation-enhanced adaptive Graph Representation Learning (SR-GRL) is proposed. A heterogeneous transition graph is constructed to integrate three entity types, namely POIs, POI categories, and regions, and their complex interrelationships. An adaptive balanced random walk sampling strategy is designed to balance node distributions across different domains dynamically and to reduce information redundancy. A type-aware attention mechanism is then used to learn semantic associations among nodes through relation-specific transformation matrices, so that feature differences across node types can be identified effectively. The obtained disentangled POI representations are then used for spatiotemporal encoding of user check-in sequences, and a self-attention mechanism is applied to aggregate users, temporal preference features. Finally, next POI recommendation is generated through a Softmax function. Results and Discussions Experiments on the Foursquare datasets from Tokyo and New York and the Sina Weibo dataset from Shanghai show that, compared with state-of-the-art baselines, the SR-GRL method achieves Recall@10 improvements of 2.22%\begin{document}$ \sim $\end{document} \begin{document}$ \sim $\end{document} \begin{document}$ \sim $\end{document} Conclusions Overall, the SR-GRL approach can balance the distributions of different node types dynamically and strengthen the modeling of complex semantic dependencies in heterogeneous contextual information.
Available online , doi: 10.11999/JEIT201066
Abstract:
To solve the problem of Group Repetition Interval (GRI) selection in the construction of the enhanced LORAN (eLORAN) system supplementary transmission station, a screening algorithm based on cross interference rate is proposed mainly from the mathematical point of view. Firstly, this method considers the requirement of second information, and on this basis, conducts a first screening by comparing the mutual Cross Rate Interference (CRI) with the adjacent Loran-C stations in the neighboring countries. Secondly, a second screening is conducted through permutation and pairwise comparison. Finally, the optimal GRI combination scheme is given by considering the requirements of data rate and system specification. Then, in view of the high-precision timing requirements for the new eLORAN system, an optimized selection is made in multiple optimal combinations. The analysis results show that the average interference rate of the optimal combination scheme obtained by this algorithm is comparable to that between the current navigation chains and can take into account the timing requirements, which can provide referential suggestions and theoretical basis for the construction of high-precision ground-based timing system.
To solve the problem of Group Repetition Interval (GRI) selection in the construction of the enhanced LORAN (eLORAN) system supplementary transmission station, a screening algorithm based on cross interference rate is proposed mainly from the mathematical point of view. Firstly, this method considers the requirement of second information, and on this basis, conducts a first screening by comparing the mutual Cross Rate Interference (CRI) with the adjacent Loran-C stations in the neighboring countries. Secondly, a second screening is conducted through permutation and pairwise comparison. Finally, the optimal GRI combination scheme is given by considering the requirements of data rate and system specification. Then, in view of the high-precision timing requirements for the new eLORAN system, an optimized selection is made in multiple optimal combinations. The analysis results show that the average interference rate of the optimal combination scheme obtained by this algorithm is comparable to that between the current navigation chains and can take into account the timing requirements, which can provide referential suggestions and theoretical basis for the construction of high-precision ground-based timing system.
