Email alert
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/JEIT260031
Abstract:
Objective Domain Adaptation Semantic Segmentation (DASS) is critical for remote sensing applications, including land-cover mapping, urban planning, and environmental monitoring. However, deep learning models often show severe performance degradation under domain shifts caused by imaging variation, geographic differences, and label-semantic heterogeneity. Conventional feature-alignment and generative adversarial network-based methods often fail to preserve semantic consistency. They are also sensitive to noisy supervision, especially when cross-domain gaps are large. This work aims to construct a robust DASS framework for semantically consistent image translation and reliable knowledge transfer. Methods A two-stage framework, termed Co-training Spatial-Guided DASS (CoSG-DASS), is proposed by integrating image translation and co-training. In the image-translation stage, a spatial information-guided latent diffusion model enhanced by ControlNet is designed. Semantic pseudo-labels and depth estimates are used as horizontal semantic and vertical spatial conditions to guide target-style image generation. To reduce the effect of noisy pseudo-labels, an Entropy-based Adaptive Guidance Intensity Module (EAGIM) is introduced. EAGIM estimates pixel-level confidence using information entropy and suppresses unreliable features. In the co-training stage, translated target-style images and unlabeled real target-domain images are used to train a segmentation model with a depth-guided segmentation head. Cross-entropy loss and adversarial loss are jointly used for optimization. Results and Discussions Extensive experiments are conducted on three cross-domain tasks. CoSG-DASS generates images that better match target-domain distributions. Quantitative results based on Fréchet Inception Distance (FID) show that the proposed method outperforms CycleGAN, UNI-Diff, and CRS-Diff in most settings (Table 1 ). Visual comparisons (Fig. 6 ) show that the method reduces edge blurring and category confusion. It also improves the separation of roads and vegetation and preserves small objects, such as vehicles. In the semantic segmentation stage, CoSG-DASS outperforms state-of-the-art domain adaptation methods. It improves mean Intersection over Union (mIoU) by 1.14%, 3.78%, and 2.49% on the cross-geographic task (Vaihingen IRRG→Potsdam IRRG), cross-imaging-mode task (Vaihingen IRRG→Potsdam RGB), and bidirectional label-semantic-heterogeneity tasks between DFC25 and LoveDA, respectively (Tables 2 ～4 ). Visual segmentation results (Fig. 7 ) confirm its strong boundary preservation and high accuracy in complex scenes. Ablation studies (Table 5 ) verify the contribution of the core components, including depth control, pseudo-label guidance, EAGIM, and the co-training strategy. Feature-distribution visualization based on Uniform Manifold Approximation and Projection (UMAP) further shows that CoSG-DASS reduces intra-class variation and increases inter-class separation after adaptation (Fig. 8 ). Conclusions CoSG-DASS alleviates domain shifts in remote sensing images through semantic-preserving diffusion-based translation and depth-guided co-training. It improves both image-translation quality and segmentation accuracy over existing methods. The proposed framework provides an effective solution for multi-source remote sensing interpretation. Future work will focus on extreme label-semantic heterogeneity and lightweight diffusion architectures.
Available online ,
doi: 10.11999/JEIT260536
Abstract:
Objective To address the complexity and limited modulation of existing reconfigurable metasurfaces integrating radiation and stealth, this study proposes a single-layer space-time-coding metasurface that dynamically switches between beam scanning and RCS reduction. By periodically modulating meta-atom states in the time domain, the phase responses of incident and harmonic waves are differentiated, enabling dynamic integration of radiation and stealth without complex feeding networks or multilayer structures. The design achieves precise beam scanning and excellent RCS reduction, providing a simple, low-cost solution for integrating radar communication and electromagnetic stealth. Methods This metasurface adopts a metal-dielectric-metal structure, with each meta-atom integrating a PIN diode. Electromagnetic simulations are conducted using CST Microwave Studio. At the optimal operating frequency, the meta-atom exhibits a 1-bit tunable reflection phase and a near-lossless co-polarized reflection amplitude. By feeding space-time-coding sequences into the metasurface model, dynamic control of beam scanning and RCS reduction is achieved. A prototype is fabricated using PCB technology, and its performance is validated through echo scattering measurements conducted with a vector network analyzer and a frequency offset option. Results and Discussions Simulations and measurements confirm that the proposed space-time-coding metasurface enables precise beam steering and excellent RCS reduction. At 10.0 GHz, the meta-atoms provide a 180° reflection phase difference and near-lossless amplitude. In radiation mode, harmonic beams are steered to angles of +2°, +14°, +32°, +46°, –15°, –28°, and –44°, with side-lobe levels for the ±3rd to fundamental harmonics suppressed to –10.59 dB. In scattering mode, the RCS of the fundamental echo is reduced by 11.56 dB compared to a copper plate. Additionally, a –10 dB RCS reduction is achieved from 9.9 to 10.1 GHz, peaking at 14.83 dB at 9.95 GHz. Experimental results validate the design’s integrated radiation-scattering capability and dynamic control flexibility. Conclusions This study presents a design method for a space-time-coding metasurface that enables dynamically integrated manipulation of radiation and scattering. By employing meta-atoms integrated with PIN diodes, the proposed method periodically modulates the operating states of the meta-atoms on a single-layer metasurface to achieve differentiated distributions of the wavefront phases of various harmonics. The proposed scheme significantly reduces design complexity and manufacturing costs. The successful realization of beam scanning and RCS reduction highlights the great potential of this technology in the integrated design of radar communication and electromagnetic stealth.
Available online ,
doi: 10.11999/JEIT251111
Abstract:
Objective Lip reading is a challenging yet important task in computer vision that aims to decode spoken language solely from visual lip movements. The task is difficult mainly because of inherent ambiguity in visual speech signals. On one hand, articulatory movements for different visemes can be highly subtle. For instance, lip displacement differences for confusable pairs such as /p/–/b/ and /m/–/n/ may be as small as 0.3～0.7 mm. These fine-grained spatial variations often fall below the effective resolution limits of conventional 3D convolutional neural networks. On the other hand, natural coarticulation in speech introduces temporal ambiguity, as mouth shapes may transiently blend multiple phonemes and make distinct visual units difficult to separate. These challenges are further aggravated by real-world factors such as uneven lighting and substantial inter-speaker differences in articulation. Current lip-reading models therefore often show limited ability to capture discriminative spatiotemporal features, which leads to suboptimal performance, particularly for phonemes with minimal visual differences. To address these issues, a robust lip-reading framework is developed to capture and exploit fine-grained spatiotemporal dependencies and improve recognition accuracy under diverse and realistic conditions. Methods To address the above limitations, a novel lip-reading framework, termed the Multi-dimensional Spatio-Temporal Enhancement Network (MSTEN), is proposed. The framework is designed to strengthen spatial and temporal representations through integrated attention mechanisms and advanced residual learning. It contains three core components that collaboratively model dependencies between spatial and temporal features, which are often insufficiently used in conventional architectures. The first component, the Self-adjusting Spatio-temporal Attention (SaSTA) module, adopts a self-adjusting mechanism that operates simultaneously across the height, width, and temporal dimensions. Query, key, and value tensors are generated through 1×1×1 3D convolutions, flattened across spatial and temporal dimensions, and used to compute attention weights by multiplying the query tensor with the transposed key tensor, followed by softmax normalization. The resulting attention map is multiplied by the value tensor and then combined with the original input through learnable parameters and a residual connection, thereby preserving contextual information and producing globally enhanced features. The second component, the Three-dimensional Enhanced Residual Block (TE-ResBlock), improves spatiotemporal feature extraction through temporal shift, multi-scale convolution, and channel shuffle. The temporal shift operation moves one quarter of the feature channels along the time axis to fuse information from adjacent frames without adding parameters. Multi-scale convolution adopts parallel branches with kernel sizes of 3×3, 3×1, 1×3, and 1×1 to capture features at different receptive fields. The outputs are concatenated and processed by channel shuffle to improve information exchange across groups, and four TE-ResBlocks are stacked to achieve progressive feature refinement. The third component, the Multi-dimensional Adaptive Fusion (MDAF) module, integrates spatial, temporal, and channel information through three submodules. These are a Channel Enhancement Module (CEM), which recalibrates features through max pooling, temporal convolution, and sigmoid activation, a Spatial Enhancement Module (SEM), which expands the receptive field through identity mapping and both standard and dilated convolution, and an Adaptive Temporal Capture Module (ATCM), which emphasizes dynamic movements through frame-difference features and temporal weight maps. The MDAF modules are inserted between TE-ResBlock stacks for iterative refinement. Finally, features extracted by the MSTEN front end are fed into a Densely Connected Temporal Convolutional Network (DC-TCN) back end, which consists of four blocks, each containing three temporal convolutional layers with dense connections, to model long-range phonological dependencies effectively. Results and Discussions The proposed framework is evaluated comprehensively on the widely used LRW and GRID datasets. The LRW dataset contains more than 500,000 video clips from over 1,000 speakers. The GRID dataset contains video clips from 34 speakers, each providing 1,000 utterances, with a total duration of 28 h. The proposed model achieves an accuracy of 91.18%, which is an absolute improvement of 2.82 percentage points over a strong ResNet18 baseline, demonstrating its substantial effectiveness. Ablation studies are further conducted to analyze the contribution of each key component. The results show clearly that each proposed module yields a meaningful performance gain. Specifically, the SaSTA module alone improves accuracy by 2.09%, which confirms the key role of global spatiotemporal attention. The TE-ResBlock increases accuracy by 1.73%, which verifies its effectiveness in multi-scale local feature extraction and inter-frame information fusion. The MDAF module provides a further 1.74% improvement, which highlights the benefit of adaptive multi-dimensional feature fusion, as shown in Table 2 . Conclusions This study advances lip reading through the proposed MSTEN front-end network. The framework is built on three main contributions. First, the SaSTA module proposes an effective mechanism for global context aggregation and performs multi-dimensional feature weighting across height, width, and temporal sequences. Second, the TE-ResBlock addresses central challenges in spatiotemporal modeling through the combined use of temporal displacement, multi-scale convolution, and enhanced channel interaction. Third, the MDAF module enables deep and coordinated integration of spatial, temporal, and channel information. Together, these components improve model performance substantially and achieve accuracies of 91.18% on the challenging LRW dataset and 97.82% on the GRID dataset. Ablation studies further confirm the individual and combined effectiveness of the proposed components. Future work will examine the extension of this framework to audio-visual speech recognition under noisy conditions and the development of domain adaptation strategies to improve robustness in low-resolution or resource-constrained scenarios.
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/JEIT260192
Abstract:
Objective In modern power inspection, the use of multiple Unmanned Aerial Vehicles (UAVs) for cooperative inspection is an efficient but challenging task. Existing multi-UAV path planning methods often have limited cooperative scheduling capability. They also fail to accurately capture the topological relationships among heterogeneous nodes, especially those between device nodes and charging stations under strict energy constraints. To address these limitations, this paper proposes Hierarchical Attention mechanism-based Path Planning for multi-UAV Inspection (HAPPI). The objective is to minimize the total flight distance of the UAV fleet while ensuring that all device nodes are inspected and all UAVs safely return to the base station under energy and visit-count constraints. Methods The multi-UAV power inspection problem is first formulated as a combinatorial optimization problem with energy constraints. It is then modeled within a Markov Decision Process (MDP) framework. To solve this problem, HAPPI adopts an encoder-decoder architecture with a customized hierarchical attention mechanism. The encoder uses a multi-level attention design to model three types of node relationships. Self-attention among device nodes is used to learn spatial proximity and visit-order preferences. Cross-attention between device nodes and charging stations is used to model energy supply-demand relationships. Self-attention among charging stations is used to explicitly capture the topological structure of the charging-station network. This hierarchical design enables the model to distinguish functional differences and dependencies among heterogeneous nodes. The decoder integrates the global graph embedding, the embedding of the last visited node, and the current remaining energy of the UAV to generate a context vector. A single-head attention mechanism is then used to compute compatibility scores for all candidate nodes. A masking strategy excludes infeasible nodes, including visited nodes, unreachable nodes, nodes that would prevent the UAV from reaching a charging station, and premature returns to the base station. The final node is selected from a probability distribution generated by softmax, which supports both greedy and sampling decoding strategies. The policy network is trained using reinforcement learning, and a baseline network is used to stabilize training. Policy-gradient optimization is used to minimize the expected total path length (Fig. 2 ). Results and Discussions Extensive simulations are conducted on three problem scales: T20C2, with 20 device nodes and 2 charging stations; T60C6, with 60 device nodes and 6 charging stations; and T100C10, with 100 device nodes and 10 charging stations. The training results show that HAPPI achieves faster convergence and a lower final cost than the baseline Attention Model (AM) and Heterogeneous Attention-based Deep Reinforcement Learning (HADRL) methods (Fig. 4 ). In the comprehensive performance comparison, HAPPI with sampling obtains the shortest total path lengths on T60C6 and T100C10, with values of 6.21 and 8.41, respectively. It outperforms five classical metaheuristic algorithms and two deep reinforcement learning baselines on these two larger-scale scenarios (Table 1 ). On T20C2, HADRL with sampling achieves the shortest path length, whereas HAPPI remains highly competitive. Overall, HAPPI reduces the total path length by approximately 12% on average compared with the baseline methods. The visualization results show that HAPPI generates routes with fewer route crossovers and a more balanced workload among UAVs, improving safety and efficiency (Fig. 6 and Fig. 7 ). The single-UAV path length distribution further confirms the superior load-balancing capability of HAPPI across all problem scales (Fig. 8 ). Conclusions This paper presents HAPPI, a hierarchical attention mechanism-based deep reinforcement learning method for cooperative path planning in multi-UAV power inspection scenarios with multiple charging stations. By explicitly modeling spatial relationships among device nodes, energy dependencies between device nodes and charging stations, and the internal topology of the charging-station network, HAPPI improves information aggregation and constraint satisfaction. Experimental results across different problem scales show that HAPPI achieves higher planning quality, greater computational efficiency, and stronger generalization than heuristic and learning-based comparison methods. Future work will extend this framework to multi-objective optimization that considers time, risk, and energy trade-offs, and will further validate the method using real-world inspection data.
Available online ,
doi: 10.11999/JEIT251099
Abstract:
Objective A central goal in modern cancer research is to determine why patients respond differently to the same therapy. This requires computational tools that combine genetic information with drug properties to predict treatment results, which is essential for advancing personalized oncology. Although existing methods have improved cancer drug response prediction, effective drug feature extraction and integration of multi-omics data from cell lines remain challenging. To address these issues, Graph Neural Networks (GNNs) have been increasingly used to process drug molecular graphs. In this study, a model based on a graph topology attention network is proposed to extract features from drug molecular graphs, and an attention mechanism is used to integrate multi-omics data. Methods In this study, a drug response prediction method based on Graph Topology Attention Network (GTAT) is proposed. The model integrates topological graph information to predict drug responses in cell lines. Drug SMILES strings are used to generate two different drug representations, and multi-omics data are incorporated to characterize cell lines (Fig. 1 ). For drug feature extraction, SMILES strings are first parsed to construct molecular graphs, which are then processed by GTAT. This network captures both topological information at the molecular graph level and atom-level features, thereby generating structured molecular representations. At the same time, Extended Connectivity Fingerprints are computed from the same SMILES strings and transformed into continuous feature vectors through a Multi-Layer Perceptron (MLP). The graph-based drug representation and the fingerprint-based representation are then concatenated to form a comprehensive drug feature vector. For cell line representation, multi-omics data are processed through omics-specific neural networks. The resulting features are fused through multi-head self-attention mechanisms, which enable the model to capture contextual interactions across omics modalities and generate an integrated cell line representation. Finally, the drug and cell line features are combined and fed into an MLP classifier to predict drug response results. The proposed model effectively integrates heterogeneous biological data sources and significantly improves prediction accuracy through multimodal learning and attention-based feature fusion. Results and Discussions The proposed method achieves competitive performance on both the GDSC and CCLE benchmark datasets (Table 2 ). Specifically, on the GDSC dataset, the proposed approach outperforms all competing methods across all four metrics, including AUC, AUPR, F1-score, and Accuracy. In particular, the AUPR is improved by approximately 1.92% compared with that of the second-best method, MOFGCN, which demonstrates an advantage in handling class imbalance. On the CCLE dataset, the proposed method still achieves the best performance in terms of AUC and Accuracy. Although its AUPR and F1-score are slightly lower than those of GADRP, the differences are minimal, and the method shows stronger overall discriminative ability, as reflected by AUC. These results validate the effectiveness and strong generalizability of the proposed method in drug sensitivity prediction tasks. The variation in AUPR and F1-score across datasets may be attributed to inherent differences in sample size and class distribution. The limited size of the CCLE dataset, combined with its specific class imbalance, with an approximately 4:1 ratio of resistant to sensitive samples, may restrict the model’s ability to fully learn the underlying data distribution, particularly for minority classes. In contrast, the GDSC dataset shows greater heterogeneity and a more pronounced class imbalance, approximately 8:1, which increases prediction difficulty and leads to lower performance on certain metrics. Conclusions Accurate prediction of drug response in cell lines remains a central challenge in precision medicine and has important implications for accelerating drug development and advancing personalized treatment. However, construction of a highly accurate predictive model that effectively integrates multi-source biological information remains difficult because of the complexity of drug molecular structures and the inherent heterogeneity of cell lines. To address this issue, a cell line drug response prediction model based on GTAT is proposed. In this model, GTAT is used to extract molecular graph features of drugs, which are then fused with molecular fingerprint features. Meanwhile, multi-omics features of cell lines are integrated through an attention mechanism. Experimental results demonstrate that the proposed model achieves superior performance compared with existing state-of-the-art benchmark methods on the employed datasets. This study provides a new perspective for cell line drug response prediction. Certain limitations remain, including the use of only three types of omics features for cell line representation and the effect of sample size on predictive performance. Future work will focus on integrating more diverse omics features, applying pre-trained large-scale models, and promoting clinical translation for personalized medicine.
Available online ,
doi: 10.11999/JEIT260263
Abstract:
Objective Satellite-Direct-to-Device (SD2D) systems that reuse terrestrial IMT spectrum may generate harmful downlink interference to incumbent IMT networks in neighboring administrations, particularly in cross-border deployments where SD2D downlinks overlap the receive bands of both IMT user equipment (UE) and IMT Base Stations (BSs). A practical coexistence methodology is therefore required to (i) translate IMT receiver protection criteria into explicit Power Flux Density (PFD) and Equivalent Power Flux Density (EPFD) constraints and (ii) validate these constraints using a dynamic simulation framework so that they can be converted into enforceable geographic coordination measures, such as minimum isolation distances. This study focuses on the dominant interference path, namely SD2D downlink interference to IMT receivers, and establishes a traceable workflow from deterministic protection limits to dynamic simulation validation and the corresponding minimum isolation distances. Methods A cross-border scenario is modeled in which Country A deploys an SD2D system and Country B operates a terrestrial IMT network. Two representative downlink frequencies, 1995 MHz and 2 190 MHz, are evaluated for two representative Starlink configurations, Starlink-1 and Starlink-2. The IMT network is modeled using ITU-R- and 3GPP-compliant parameters, with an I/N protection threshold of –6 dB and a target percentile κ (baseline κ=99.5%) for both IMT UEs and BSs. Satellite transmit antennas follow the ITU-R S.1528 reference pattern, IMT BS receive antennas follow the ITU-R F.1336 sector pattern, and IMT UEs are modeled with omnidirectional antennas. A back-lobe blockage model is incorporated into both satellite and BS antenna patterns to account for rear-side shielding. Signal propagation follows the ITU-R P.619 model, using free-space path loss as the conservative baseline, while an optional clutter-loss term is incorporated through a clutter-occurrence probability. Deterministic protection limits are derived by calculating the maximum permissible aggregate PFD for IMT UE protection and the maximum permissible aggregate EPFD for IMT BS protection. A dynamic simulation framework then validates these limits and searches for the required minimum isolation distances (Fig. 4 ). Co-channel beam isolation angles are optimized using the C/I Complementary Cumulative Distribution Function (CCDF), and a segmented search algorithm determines the minimum UE- and BS-side isolation distances together with the corresponding κ-percentile PFD/EPFD statistics. Results and Discussions The deterministic analysis yields a maximum permissible aggregate PFD of –102.72 dBW/m2/MHz at 2 190 MHz for IMT UEs and a maximum permissible aggregate EPFD of –129.53 dBW/m2/MHz at 1995 MHz for IMT BSs (Fig. 3 ). For Starlink-1, the C/I design criterion yields a minimum co-channel beam isolation angle pair of (12°, 12°) (Fig. 5 ). Dynamic simulation shows that, under the representative baseline configuration with an I/N threshold of –6 dB and κ=99.5%, the minimum isolation distances are 195 km for UE protection and 290 km for BS protection (Fig. 6 , Fig. 7 , and Table 4 ). The resulting coordination isolation distance is therefore 290 km, and the simulated κ-percentile PFD and EPFD agree with the deterministic protection limits, with a residual margin below 0.5 dB. For Starlink-2, the optimized co-channel beam isolation angles increase to (15°, 15°), and the corresponding minimum isolation distances increase to 272 km for UEs and 420 km for BSs under the same baseline configuration (Table 5 ). These baseline distances should be interpreted as representative values for the specified simulation configuration rather than unique, strictly converged results. Stability verification shows that, under different sampling intervals, simulation durations, and random seeds, the UE- and BS-side minimum isolation distances remain within 195～210 km and 290～300 km, respectively, for Starlink-1, and within 266～290 km and 370～420 km, respectively, for Starlink-2 (Table 7 ). Sensitivity analysis for Starlink-1 further indicates that the required minimum isolation distance is governed by the upper tail of the aggregate I/N distribution (Table 6 ). Increasing κ from 99.5% to 100% increases the UE- and BS-side minimum isolation distances from 195/290 km to 304/560 km. Clutter attenuation substantially reduces the UE-side minimum isolation distance, decreasing it to 173 km when the clutter-occurrence probability is 0.5, while producing little change in BS protection. Polarization reuse increases the UE- and BS-side minimum isolation distances to 222 km and 360 km, respectively, whereas increasing the number of co-channel beams to 16 increases the BS-side minimum isolation distance to 330 km. The minimum service elevation angle and the link establishment strategy are identified as the dominant operational factors. Changing the minimum service elevation angle from 10° to 35° changes the required UE- and BS-side minimum isolation distances from 340/460 km to 101/150 km, whereas replacing the Sat-MaxElevation strategy with the UE-MaxElevation strategy reduces them to 80/180 km. Conclusions The proposed workflow converts IMT receiver protection criteria into deterministic protection limits expressed as PFD and EPFD constraints and validates them using a dynamic simulation framework. Under an I/N threshold of –6 dB and κ=99.5%, the baseline and stability analyses jointly indicate representative UE- and BS-side minimum isolation-distance ranges of 195～210 km and 290～300 km for Starlink-1 and 266～290 km and 370～420 km for Starlink-2, rather than unique, strictly converged values. Sensitivity analysis further shows that κ only changes the statistical criterion used to extract tail events from the sample set, whereas clutter attenuation primarily benefits IMT UEs. In contrast, the minimum service elevation angle, polarization reuse, the number of co-channel beams, and the link establishment strategy reshape the worst-case interference geometry and can produce substantial, and sometimes non-monotonic, changes in the required minimum isolation distances. The proposed framework establishes a traceable link between IMT receiver protection criteria and enforceable border coordination measures.
Available online ,
doi: 10.11999/JEIT260189
Abstract:
Significance Remote Sensing Image-Text Retrieval (RS-TIR) connects large-scale Earth observation imagery with natural-language queries and has become an important interface for geospatial intelligence systems. Compared with conventional content-based retrieval, RS-TIR allows users to search for scenes, objects, spatial layouts, and functional regions through semantic descriptions rather than handcrafted visual cues. This capability is increasingly needed in natural resource monitoring, urban governance, disaster response, environmental assessment, and on-demand retrieval from rapidly growing satellite archives. However, RS-TIR remains challenging. Remote sensing imagery is captured from nadir or near-nadir perspectives, shows strong rotation invariance, and contains extreme scale variation, ranging from tiny vehicles to large airports. It also requires domain-specific semantic descriptions, such as land-use attributes, spatial distributions, and geoscientific relations. Meanwhile, high-quality image-text annotations remain limited relative to the scale of remote sensing data. These properties widen the cross-modal semantic gap between images and language and limit the generalization ability of traditional cross-modal retrieval methods. Against this background, this review examines how Vision-Language Foundation Models (VLMs) reshape RS-ITR through large-scale contrastive pre-training, stronger transferable representations, and more flexible multimodal interaction mechanisms. It also explains why remote sensing adaptation is needed and why a focused synthesis of architectures, datasets, alignment mechanisms, and future directions is timely for this field. Progress The technical development of RS-ITR is reviewed from three complementary perspectives. First, this review summarizes the domain-specific challenges that shape the task, including visually isotropic topology with extreme scale variation, professional and fine-grained textual semantics, and the compounded cross-modal semantic gap between overhead imagery and natural-language descriptions (Fig. 3 ). The overall survey structure is then presented to show the logical progression from task formulation to future challenges (Fig. 1 ). From a methodological perspective, RS-ITR has evolved from handcrafted visual descriptors and shallow semantic mapping to deep representation learning, and then to VLM-driven paradigms with stronger generalization and zero-shot transfer capability (Fig. 4 , Table 2 ). Early methods rely on color, texture, shape, and hash-based retrieval. However, they struggle to model high-level geospatial semantics and complex scene composition. Deep learning methods improve retrieval by learning joint embedding spaces, adopting dual-encoder or interaction-based architectures, and using multi-scale feature fusion and region-aware matching. These methods improve semantic consistency, but they still depend heavily on labeled data and often show limited robustness in open or cross-sensor scenarios. Second, this review summarizes the benchmark ecosystem used to evaluate these methods. Representative datasets range from small-scale test sets, such as Sydney-Caption and UCM-Caption, to mainstream benchmarks, such as RSICD and RSITMD, and recent large-scale training resources, such as RS5M and SkyScript (Table 1 ). These datasets show a clear transition from small manually annotated corpora to web-scale or automatically generated image-text pairs. This transition supports domain pre-training and large model adaptation. Third, this review analyzes the core VLM techniques that now drive progress in RS-ITR. The model spectrum and representative architecture families are systematically summarized, including contrastive dual-encoder models, multimodal interaction models, and remote sensing foundation models integrated with large language models (Fig. 5 , Fig. 6 , Table 3 ). Domain adaptation routes are further grouped into continued remote sensing pre-training, parameter-efficient transfer learning, adapter-based tuning, prompt learning, and instruction tuning. At the semantic alignment level, this review focuses on contrastive joint embedding, fine-grained multi-scale alignment, and the use of remote sensing priors, such as spatial topology and geolocation. Performance comparisons on RSICD and RSITMD show that remote sensing VLMs, especially RemoteCLIP, GeoRSCLIP, iEBAKER, and LRSCLIP, yield consistent gains in mean Recall (mR) and overall retrieval robustness (Table 4). In parallel, this review tracks the extension of retrieval capability into unified multi-task remote sensing models, in which retrieval, grounding, segmentation, and reasoning begin to share a common multimodal representation space. Conclusions Several conclusions are drawn from the comparative analysis. First, VLMs establish a dominant paradigm for RS-ITR because they narrow the cross-modal semantic gap and improve transferability across datasets and scenes. Second, no single architecture is universally optimal. Dual-encoder models remain attractive for large-scale retrieval because of their efficiency, whereas interaction-based or instruction-enhanced models provide finer semantic alignment at a higher computational cost. Third, domain adaptation is indispensable. Continued pre-training on remote sensing image-text corpora, parameter-efficient tuning, and prompt-based adaptation consistently outperform direct reuse of internet-trained VLMs. This finding indicates that remote sensing imagery differs too strongly from natural-image distributions for generic pre-training alone to be sufficient. Fourth, the most effective recent methods do not improve performance through scale alone. They also exploit remote sensing-specific information, including multi-scale structures, foreground objects, explicit keyword reasoning, and spatial priors. Finally, this review shows that the field is shifting from isolated retrieval models toward more general geospatial multimodal systems. Retrieval is no longer treated only as a matching task. It is also becoming a key capability that supports question answering, instruction following, knowledge augmentation, and coordinated reasoning in remote sensing applications. Prospects Future research is expected to advance in four closely related directions. The first direction is the unified representation of multi-source heterogeneous data, especially the integration of optical imagery with Synthetic Aperture Radar (SAR), hyperspectral data, thermal infrared observations, and multi-temporal acquisitions. The second direction is knowledge-enhanced retrieval, in which geospatial priors, land-use rules, remote sensing terminology, and external knowledge bases are incorporated into multimodal alignment and retrieval-augmented reasoning. The third direction is lifelong and open-world learning. Real deployment requires models to remain reliable under seasonal variation, sensor updates, regional domain shifts, cloud contamination, and newly emerging categories, while avoiding catastrophic forgetting. The fourth direction is efficiency and deployability. Practical remote sensing systems often operate under tight computational budgets. Therefore, lightweight tuning, sparse computation, token reduction, model compression, and on-orbit and edge inference will become increasingly important. Interactive and explainable retrieval is also likely to gain importance. It allows analysts to refine queries through dialogue and inspect the image regions or semantic cues that support retrieval decisions. Overall, continued progress in data construction, domain adaptation, semantic alignment, and efficient multimodal modeling is expected to make RS-ITR a more robust infrastructure capability for Earth observation applications.
Available online ,
doi: 10.11999/JEIT260140
Abstract:
Objective Side-Channel Attack (SCA) is one of the primary threats to the physical security of cryptographic chips, and deep learning methods for secret key recovery have attracted considerable attention in the field of SCA. However, existing deep learning-based side-channel attack methods have limited capability to focus on critical leakage intervals during feature extraction, particularly for long traces with high-dimensional noise. Therefore, irrelevant background noise interferes with feature extraction, leading to reduced attack efficiency and slower convergence of Guessing Entropy (GE). To address these limitations, a deep side-channel attack method integrating the Convolutional Block Attention Module (CBAM) and triplet loss is proposed to improve the extraction of weak leakage features under complex noise conditions and enhance secret key recovery efficiency. Methods CBAM is embedded into a Convolutional Neural Network (CNN) to construct an adaptive feature extraction network. CBAM consists of a Channel Attention Module (CAM) and a Spatial Attention Module (SAM). CAM adaptively recalibrates feature-channel weights to emphasize leakage-related features with a high Signal-to-Noise Ratio (SNR), whereas SAM identifies Points Of Interest (POI) in the temporal domain and suppresses background noise outside the leakage intervals. After attention-based feature refinement, triplet loss is adopted as the optimization objective to optimize the distribution of embedding features, encouraging compact intra-class clusters while maximizing inter-class separation in the embedding space. Finally, a multivariate Gaussian template attack is performed using the optimized embedding features to recover the secret key. The overall framework is illustrated in (Fig. 2 ). Results and Discussions The proposed method is evaluated on two public benchmark datasets, ASCAD and AES_HD, using GE and the minimum number of attack traces required for GE to converge to 1 (\begin{document}$ {T}_{{\mathrm{GE0}}} $\end{document} Table 2 and Fig. 3 ). On the low-SNR AES_HD dataset, the proposed method requires 1,219 attack traces, outperforming MHA and NLS while maintaining smooth and stable GE convergence (Table 2 and Fig. 4 ). Furthermore, desynchronization experiments demonstrate that the proposed method maintains accurate localization of effective leakage points under severe desynchronization noise, indicating strong robustness to time-domain jitter (Table 3 ). Ablation studies further verify the synergistic effect of the proposed architecture and confirm the effectiveness of its core components (Table 4 ). Conclusions A deep side-channel attack method integrating CBAM and triplet-loss-based metric learning is proposed. The CBAM module enables the network to adaptively focus on leakage-related features, improving feature extraction over conventional CNN-based methods. Triplet loss enhances the discriminability of embedding features, thereby improving template matching accuracy. Experimental results on the ASCAD and AES_HD datasets demonstrate that the proposed method substantially reduces the number of attack traces required for successful secret key recovery and accelerates GE convergence. The proposed method consistently outperforms existing mainstream approaches under fixed-key, random-key, and low-SNR conditions. Future work will focus on improving robustness under more severe desynchronization conditions and enhancing generalization in small-sample scenarios.
Available online ,
doi: 10.11999/JEIT260267
Abstract:
Objective WiFi-based indoor passive human tracking has attracted increasing attention for smart home applications because Internet of Things (IoT) devices are widely interconnected through existing WiFi infrastructures. Existing approaches estimate the Angle of Arrival (AoA) or Time of Flight (ToF) of the target-reflected path to achieve tracking. However, these methods are fundamentally limited by the difficulty of separating weak target-reflected signals from dense multipath propagation when commercial WiFi devices provide only a small antenna array and limited bandwidth under existing communication protocols. Therefore, Doppler Frequency Shift (DFS)-based approaches that exploit multiple links have become more practical. Dead Reckoning (DR) is widely adopted in these methods, but dynamic link selection for signal fusion remains challenging. Furthermore, the performance of DR-based methods degrades substantially when device-location errors are present. Existing methods also generally employ a single-transmitter–multi-receiver architecture, in which Access Point (AP) broadcasts downlink WiFi signals to STAtions (STA). It complicates data aggregation and limits the use of multiple receive antennas of AP's. To address the challenges of multi-link signal fusion and device-location uncertainty, a Particle Filter (PF)-based method is proposed. Furthermore, a multi-transmitter–single-receiver architecture is adopted, in which multiple STAs transmit uplink packets to a single AP. This architecture simplifies data aggregation while exploiting the AP’s multi-antenna capability. Methods In the IEEE 802.11 standard, the wireless channel is estimated at the receiver using pilot signals embedded in WiFi packets, and Channel State Information (CSI) is continuously obtained. Human motion perturbs multipath propagation and produces time-varying changes in CSI. Therefore, CSI serves as the primary sensing signal because it implicitly captures target motion. In practice, raw CSI is affected by amplitude and phase impairments. Accordingly, signal preprocessing is first performed before Doppler extraction. Subsequently, multi-link DFS measurements are fused using PF, in which the target state is sequentially propagated and updated according to likelihoods derived from DFS measurements. This probabilistic framework naturally enables dynamic multi-link fusion because unreliable links receive lower weights rather than being deterministically discarded. Moreover, the effects of device-location errors are incorporated into the measurement-noise model, improving robustness to device-location uncertainty. When prior knowledge of device locations is unavailable, device self-localization is first performed by jointly estimating the Line-of-Sight (LoS) AoA and ToF between the AP and each STA. Preliminary device self-localization experiments (Fig. 5 ) demonstrate a median STA localization error of approximately 0.56 m (Table 1 ), providing reliable initialization for the subsequent tracking algorithm. Results and Discussions A prototype system is implemented using a multi-transmitter–single-receiver architecture with commercial Intel AX200/AX201 WiFi cards. CSI is collected over a 20 MHz channel centered at 5.24 GHz using PicoScenes, with IEEE 802.11ac packets transmitted at 100 or 200 Hz. Each CSI sample contains measurements from 57 subcarriers. Meanwhile, Fine Time Measurement (FTM) is performed on Channel 11 at 2.4 GHz using iw and hostapd. In the prototype system, each STA transmitter is equipped with a single antenna, while the AP receiver employs two antennas. Experiments are conducted using one AP and three or four STAs (Fig. 7 ). WiTraj and PITrack, both based on DR, are used as baseline methods. When device self-localization is required, the proposed method achieves a median tracking error of 0.47 m, compared with 1.78 m for WiTraj and 1.77 m for PITrack (Fig. 9 ), representing an accuracy improvement of approximately 70% over conventional DR-based methods. Error-injection experiments with random device-location perturbations show that, unlike conventional DR-based methods, which are highly sensitive to device-location errors, the proposed method remains robust to device-location uncertainty (Fig. 10 ). Additionally, computational complexity analysis shows that, although execution time increases with the particle count (Table 3 ), tracking performance reaches a stable level beyond a moderate particle count (Fig. 11 ). Therefore, real-time operation can be achieved by selecting an appropriate particle count. Finally, experiments in complex environments demonstrate that the proposed method consistently achieves higher tracking accuracy than the baseline methods (Fig. 12 ). Conclusions A PF-based method is proposed for multi-link collaborative passive human tracking using commercial WiFi devices. By probabilistically fusing measurements from multiple links, robust tracking is achieved even when device-location errors are present. When device locations are unavailable, device self-localization is achieved by combining CSI-based estimation with FTM measurements, providing the geometric information required for tracking. A prototype system based on a multi-transmitter–single-receiver architecture is developed to simplify data aggregation while exploiting the AP’s multi-antenna capability. Experimental results demonstrate that the proposed method achieves sub-0.5 m median tracking error when device-location errors are present, representing an improvement of approximately 70% over conventional DR-based methods. The proposed method also exhibits strong robustness to device-location uncertainty and supports real-time implementation when an appropriate particle count is selected. Future work will extend the framework to multi-person tracking and evaluate its performance under more realistic deployment conditions.
Available online ,
doi: 10.11999/JEIT260327
Abstract:
Objective Oracle Bone Character (OBC) recognition is challenged by an extreme long-tailed distribution and substantial intra-class variation. Conventional deep learning methods are often dominated by head classes, whereas existing approaches tend to overfit tail classes or fail to account for differences in learning difficulty across classes. To address these limitations, a two-stage decoupled learning framework is proposed to improve the recognition of tail and difficult classes while preserving the discriminative capability of head classes. Methods The proposed framework decouples feature representation learning from classifier optimization. In the first stage, the backbone network is trained using a mixed data augmentation strategy that combines CutMix and RandAugment with Label-Distribution-Aware Margin (LDAM) loss to learn robust feature representations and alleviate the effect of intra-class variation. In the second stage, the backbone network is frozen, and only the classifier is optimized. An adaptive difficulty sampling strategy is proposed to dynamically assign sampling weights according to historical and current class-level training difficulty. The classifier is further optimized using a Class-Balanced LDAM (CBL) loss, which combines class-balanced weighting with LDAM to refine decision boundaries for long-tailed classification. Results and Discussions Experiments on the highly imbalanced OBC306 dataset demonstrate that the proposed method achieves an overall accuracy of 94.34% and an average class accuracy of 89.89%. Compared with the Inception-v4 baseline, the proposed method improves the average class accuracy by 19.61%. Comparisons with representative long-tailed OBC recognition methods further demonstrate superior overall performance. Comprehensive ablation studies verify the effectiveness of the mixed data augmentation strategy and the adaptive difficulty sampling strategy in improving the recognition of rare and difficult characters. Parameter sensitivity analysis and qualitative error analysis further confirm the robustness and effectiveness of the proposed framework. Conclusions The proposed two-stage decoupled learning framework effectively addresses long-tailed OBC recognition by balancing the learning priorities of head, tail, and difficult classes. The mixed data augmentation strategy improves feature robustness, whereas the adaptive difficulty sampling strategy and the Class-Balanced LDAM loss jointly optimize classifier learning and refine decision boundaries without degrading head-class recognition performance. The proposed framework provides an effective solution for the digital recognition of Oracle Bone Characters and offers technical support for low-resource ancient character recognition.
Available online ,
doi: 10.11999/JEIT260283
Abstract:
Objective Filter circuits are widely used in modern control and signal-processing systems for noise suppression and signal integrity enhancement. Conventional Resistor-Capacitor (RC) filters are widely applied, but their fixed parameters limit adaptability and miniaturization in emerging molecular and nanoscale computing platforms. To address these limitations, DNA Strand Displacement (DSD) technology is integrated with memristor theory to develop tunable multistable molecular filter circuits. This study aims to design and validate first- and second-order low-pass filter circuits based on the dynamic response and state-dependent behavior of a DSD-based memristor. The proposed filters are designed to improve frequency selectivity, parameter adaptability, and system stability compared with traditional filter architectures. This approach is intended for molecular signal processing, integrated biocircuits, and adaptive filtering systems that require compact size and reconfigurability. Methods The method consists of four stages. First, core DSD reaction modules, including sine, cosine, integration, addition, and multiplication modules, are designed to construct a programmable multistable memristor model. Second, square-wave and sinusoidal input signals are generated through DSD reactions to evaluate the memristor response under different frequencies and amplitudes. Third, the memristor is embedded into low-pass filter structures to construct first- and second-order DSD-based memristor filter circuits. Fourth, simulations are performed using Visual DSD for molecular dynamics analysis and MATLAB for circuit-level analysis. Circuit performance is evaluated using transfer functions, Nyquist plots, Bode diagrams, and time-domain comparisons with classical RC filters. This combined simulation strategy verifies both molecular feasibility and circuit functionality. Results and Discussions The DSD-based memristor exhibits multistable behavior and converges to six stable equilibrium points under different initial conditions (Fig. 8 ). Its hysteresis characteristics further confirm the state-dependent memory behavior of the designed molecular memristor (Fig. 7 ). The first-order DSD-based memristor filter circuit provides stable attenuation for square-wave and sinusoidal input signals. Its output amplitudes are consistently higher than those of the traditional RC filter across the tested frequencies (Table 3 ). The second-order DSD-based memristor filter circuit further reduces signal delay and improves stability, especially under high-frequency inputs (Table 4 ). Frequency-response analyses show that the cutoff frequency can be dynamically tuned by adjusting DSD reaction rates and initial concentrations (Figs. 9 and 11 ). Time-domain simulations further confirm the filtering performance of the first- and second-order circuits (Figs. 10 and 12 ). Reliability analysis indicates that lower initial copy numbers increase stochastic molecular noise, whereas higher initial copy numbers make the output distribution closer to the deterministic response and improve the probability of successful filtering. These results verify the feasibility of DSD-memristor integration for adaptive molecular filtering. Conclusions A DSD-based memristor with multistable characteristics and its corresponding first- and second-order low-pass filter circuits are designed and validated. Compared with traditional RC architectures, the proposed filters show improved output stability, parameter tunability, and frequency adaptability. By combining DSD technology with memristor theory, this study provides a reconfigurable molecular-scale filtering framework for signal-processing applications. The results provide a basis for future work on adaptive molecular circuits, intelligent filtering, and nanoelectronic system design. Further studies should focus on experimental validation, real-time tuning strategies, sequence optimization, anti-interference design, signal amplification, and circuit integration.
Available online ,
doi: 10.11999/JEIT260024
Abstract:
Objective To address B5G/6G demands for high spectral efficiency and transmission reliability, non-orthogonal Prolate Spheroidal Wave Function (PSWF) modulation has drawn extensive research interest for its strong time-frequency energy concentration. However, severe mutual interference in PSWF signal multiplexing degrades conventional detection performance in complex channels. Limited by ideal channel assumptions or single-modal feature extraction, existing methods cannot fully exploit signal temporal-spatial information and lack adaptability to dynamic interference. This work proposes an intelligent detection architecture with adaptive temporal-spatial feature fusion (ATSFF) for accurate, robust non-orthogonal PSWF signal detection. Method A dual-path parallel framework extracts complementary features from time and transform domains. A Gated Recurrent Unit (GRU) extracts deep temporal features and captures long-range dependencies from 1D received signals. The other path converts 1D signals into 2D representations via the Gramian Angular Difference Field (GADF), and extracts hierarchical spatial features using ResNet50. An adaptive probability-weighted fusion mechanism dynamically adjusts feature contributions based on sub-network prediction uncertainty to generate robust detection outputs. Results and Discussion Simulations on a 32-class non-orthogonal PSWF dataset (Fig. 2 ) show the proposed ATSFF outperforms conventional coherent detection, cross-term suppression detection, Approximate Message Passing Interleave Division Multiple Access (AMP-IDMA) and Temporal Sparse Bayesian Learning Least Squares (TMSBL-LS) across the full Signal-to-Noise Ratio (SNR) range. t-SNE visualization (Fig. 4 ) confirms better inter-class separability and intra-class compactness of fused features. At a bit error rate of 4×10–5, it achieves 0.2 dB gain (Fig. 6 ) over coherent detection. ATSFF has higher network overhead and weaker real-time performance than linear detection, yet it supports GPU batch inference with fixed single-sample computation cost, suiting accuracy-sensitive communication scenarios. Conclusions Targeting interference issues in non-orthogonal PSWF signal detection, this paper presents an adaptive temporal-spatial feature fusion detection method. It realizes dual-modal feature extraction via GRU and ResNet50, and adopts a prediction-uncertainty-based adaptive weighted fusion mechanism to integrate dual-path strengths, significantly improving detection accuracy and robustness. Simulations validate its superior performance and stable channel adaptability, providing an efficient solution and design reference for intelligent detection of non-orthogonal PSWF signals.
Available online ,
doi: 10.11999/JEIT260149
Abstract:
Objective Supply-demand imbalances in modern distribution networks are intensified by the increasing penetration of distributed renewable energy and frequent extreme high-temperature events. Large-scale Inverter Air-Conditioning (IAC) clusters can be aggregated as virtual energy storage resources for Demand Response (DR), providing an effective way to improve grid flexibility. However, existing dispatch strategies are often limited by the curse of dimensionality. Conventional penalty-function-based soft constraints also fail to strictly satisfy aggregate power equality constraints and may introduce steady-state errors. This paper develops an optimization framework in which grid-side power commands are accurately tracked while user thermal discomfort is reduced and fairness among heterogeneous users is maintained. Methods A multi-objective optimization framework based on an Equivalent Thermal Parameter (ETP) model is established to describe the thermodynamic states of heterogeneous buildings. To balance collective comfort and individual fairness, a composite fitness function is designed by integrating a weighted mean-squared error term, a fairness variance term, and a maximum violation suppression term. To eliminate the steady-state errors of traditional penalty-based methods, a Spatial-Fitness Adaptive Particle Swarm Optimization (SFA-PSO) algorithm is proposed. A geometric constraint surface projection mechanism maps particles strictly onto the power-conservation hyperplane, thereby satisfying the aggregate power equality constraint. In addition, the learning factors are dynamically adjusted through a Spatial-Fitness Adaptive (SFA) strategy. This strategy measures the mismatch between a particle’s fitness rank and spatial distance rank, which helps prevent premature convergence in high-dimensional search spaces. Results and Discussions Extensive continuous scheduling simulations are conducted in a complex dynamic environment. The environment includes multi-source thermal disturbances, a bidirectional communication packet loss rate of 1%, and Part Load Ratio (PLR) values of 20%, 50%, and 80%. First, ablation experiments confirm that constraint surface projection guarantees power tracking accuracy. Traditional penalty-based methods, such as Penalty Particle Swarm Optimization (Penalty-PSO), produce steady-state power deviations of approximately 10^-1 kW. By contrast, SFA-PSO limits aggregate power tracking errors to within 10^-9 kW (Fig. 3 ). The SFA strategy also prevents the premature convergence observed in Physical Particle Swarm Optimization (Phy-PSO). It enables continuous fitness reduction, especially in low-load scenarios with narrow feasible regions (Fig. 4 ). This improvement is attributed to the dynamic evolution of the learning factors. The cognitive factor remains high at the early stage to promote global exploration. It then decreases as the social factor increases, which strengthens local exploitation and improves convergence precision (Fig. 5 ). Second, continuous dynamic scheduling performance is evaluated through a 6-hour simulation during the peak load period from 12:00 to 18:00. The dispatch interval is 5 min, yielding 72 decision steps. Under tight peak-load constraints, Genetic Algorithm (GA) and Whale Optimization Algorithm (WOA) show severe power-limit violations because their population update rules do not cooperate well with the projection mechanism. By contrast, SFA-PSO maintains strict constraint satisfaction (Fig. 7 ). SFA-PSO remains at the lowest fitness level throughout the real-time evolution curves, indicating strong robustness against environmental thermal noise and uplink and downlink communication packet loss (Fig. 8 ). Quantitatively, compared with eight baseline algorithms, including Social Learning Particle Swarm Optimization (SLPSO), Competitive Swarm Optimizer (CSO), and Dynamic State Cluster-Based Particle Swarm Optimization (DSCPSO), SFA-PSO achieves the best overall performance. It obtains an average fitness of 904, a minimum fitness of 243, and the lowest standard deviation of 551 (Table 2 ). Finally, scalability analyses across cluster sizes from 100 to 1,000 nodes further validate the high-dimensional optimization capability of SFA-PSO. In all scale scenarios, SFA-PSO shows the strongest optimization capacity. It achieves rapid initial descent within the first 20 iterations and maintains continuous exploration in later stages (Fig. 9 ). Although the projection and SFA mechanisms increase computational time by 30% to 50% compared with basic Particle Swarm Optimization (PSO) (Fig. 6 ), the absolute optimization time remains stable at approximately 1.5 seconds even for a 1,000-node cluster (Fig. 9 ). This computational overhead is acceptable for minute-level control cycles and meets the real-time dispatch requirements of modern smart grids. Conclusions The proposed SFA-PSO algorithm effectively addresses the steady-state error of traditional soft-constraint methods in aggregate power control. By ensuring accurate tracking of dispatch commands and mitigating high-dimensional search traps, it provides a robust and scalable solution for flexible scheduling of large-scale IAC loads in smart grids. It also maintains a practical balance between grid-side regulation and user-side comfort. The method still has limitations. The constraint projection mechanism depends on the host algorithm, which restricts cross-algorithm generalization. High-precision tracking also increases computational cost. Future work will focus on adaptive constraint handling and lightweight algorithm design. Coordinated scheduling for heterogeneous loads, such as electric vehicles and energy storage, will also be investigated.
Available online ,
doi: 10.11999/JEIT260171
Abstract:
Objective In Rate-Splitting Multiple Access (RSMA) systems, the achievable common-stream rate is limited by the user with the weakest channel quality. This constraint reduces scalability, robustness, and user fairness in dense 6G networks. Existing cooperative RSMA architectures partly reduce this bottleneck, but they remain constrained by fixed channel conditions and limited interference management. To address these issues, this paper proposes a dual Reconfigurable Intelligent Surface (RIS) cooperative RSMA system. Two cooperatively deployed RISs create additional controllable propagation paths through cascaded double reflection. The objective is to maximize the system sum rate by jointly optimizing Base Station (BS) Beamforming (BF), Rate Splitting (RS) strategies, and dual-RIS phase configurations, thereby improving spectral efficiency, robustness, and user fairness under users’ Quality of Service (QoS) constraints. Methods A tractable system model is developed for the dual-RIS cooperative RSMA system. The model captures cascaded multi-link channels, multi-node channel structures, and interference coupling. Based on this model, a joint optimization problem is formulated to maximize the system sum rate by optimizing BS BF, RS strategies, and the discrete phase shifts of both RISs. Because of strong variable coupling and non-convexity, a low-complexity Alternating Optimization (AO) algorithm is designed. The original problem is decomposed into three subproblems: BS-side RIS phase optimization, user-side RIS phase optimization, and BS BF optimization. Semidefinite Relaxation (SDR) and Successive Convex Approximation (SCA) are used to transform these subproblems into tractable convex forms, which are then solved iteratively with fast convergence. Results and Discussions Simulation results verify the effectiveness of the proposed dual-RIS cooperative RSMA system. The proposed AO algorithm converges within six iterations under different numbers of RIS reflecting elements, and it reaches 97.8% of the steady-state sum rate within three iterations when M = 170 (Fig. 3 ). Compared with SOPS and RPS, the proposed phase-configuration scheme obtains 10.6% and 31.8% sum-rate gains when M = 190, respectively (Fig. 4 ). The proposed RSMA scheme also outperforms NOMA and SDMA by 10.0% and 14.6%, respectively (Fig. 5 ). Under M = 160 and b = 4, dual-RIS cooperation provides an 11.9% sum-rate gain over the single-RIS scheme, and its performance is close to the CPS upper bound (Fig. 6 ). Balanced allocation of reflecting elements between the two RISs further improves the sum rate (Fig. 7 ). The proposed BF strategy also outperforms ZF and RBF, achieving 33.2% and 336.5% gains at a transmit power of 30 dBm, respectively (Fig. 8 ). Under different RIS cooperation modes, the proposed joint optimization scheme achieves the best overall performance (Fig. 9 ). These results show that dual-RIS cooperative RSMA improves common-stream decoding, interference suppression, robustness, and user fairness. Conclusions This paper investigates a dual-RIS cooperative RSMA communication system. The proposed architecture improves common-stream decoding while mitigating complex interference. To maximize the system sum rate, BS BF vectors, RS vectors, and the discrete phase matrices of two RISs are jointly optimized. A low-complexity AO algorithm based on SDR and SCA is developed to solve the strongly coupled non-convex problem. Simulation results show that the proposed dual-RIS cooperative RSMA scheme achieves clear sum-rate gains over advanced benchmark schemes. Compared with the single-RIS mode and SDMA, it obtains 11.9% and 14.6% rate gains, respectively, while improving system robustness and user fairness.
Column
- Cover
- Contents
- Excellence Action Plan Leading Column
- Special Topic on Intelligent Sensing and Processing Technology in Complex Electromagnetic Environments
- Radar, Sonar, Navigation and Array Signal Processing
- Wireless Communication and Internet of Things
- Cryption and Network Information Security
- Image and Intelligent Information Processing
- Circuit and System Design
Display Method:
2026, 48(6): 2317-2326.
doi: 10.11999/JEIT251381
Abstract:
Objective Reconfigurable Intelligent Surfaces (RISs) have strong potential to improve coverage and Spectral Efficiency (SE) in future wireless networks. However, when RISs are applied to wideband Multiple-Input Single-Output Orthogonal Frequency Division Multiple Access (MISO-OFDMA) systems, their practical benefits are limited by two key challenges. First, RIS reflection coefficients may not match the frequency-selective channel conditions across all subcarriers. Second, subcarrier allocation, Base Station (BS) active beamforming, and RIS passive beamforming are strongly coupled. These challenges become more serious in multi-group multicast scenarios, where shared data streams increase inter-group interference. Therefore, this article proposes a Graph Neural Network (GNN)-driven optimization framework to maximize the system SE through joint active beamforming, passive beamforming, and subcarrier allocation. Methods To address the optimization difficulty caused by the strong coupling among subcarrier allocation, BS active beamforming, and RIS passive beamforming, this work develops a model-driven GNN optimization framework. The objective is to maximize the system SE. First, a complete system model containing the BS, RIS, and multi-group multicast users is established (Fig 1 ). The formulation includes practical constraints, such as the BS transmit power limit, the unit-modulus constraint of RIS elements, and the binary constraint on subcarrier allocation. To satisfy the multicast requirement, the SE of each group is defined as the minimum SE among all users in that group. This definition further increases the non-convexity of the optimization problem.The first component of the proposed network, GNN1 (Fig 3 ), contains an initialization layer and a message-update layer. For each subcarrier \begin{document}$ n\in \mathcal{N} $\end{document} \begin{document}$ \left\{{\mathbf{H}}_{k,n},k\in \mathcal{K}\right\} $\end{document} Fig. 4 ) and RIS-level aggregation (Fig. 5 ) are then performed. GNN2 (Fig 6 ) takes the subcarrier-wise embeddings generated by GNN1 as input and constructs an expanded graph with group nodes (Fig. 7 ) and an RIS node (Fig. 8 ). By aggregating messages among subcarrier nodes, group nodes, and the RIS node, GNN2 fuses cross-subcarrier information and captures the global coupling among system components. Based on the integrated representation, GNN2 outputs the BS active beamforming matrix and RIS passive beamforming vector. Output-layer normalization is used to satisfy the physical constraints. Finally, given the beamforming parameters, subcarrier allocation is performed using the maximum-SE criterion. The learning objective is defined as maximizing the total SE. Results and Discussions The proposed GNN algorithm consistently outperforms all random benchmark schemes, including APG-randAllocate, APG-randActive, and APG-randPassive, across the full transmit power range from 0 to 20 dBm. This advantage indicates that the proposed method can dynamically handle subcarrier allocation and joint active and passive beamforming optimization. It also maintains stable and superior performance under large transmit-power variations. Overall, the system SE of all schemes increases monotonically with BS transmit power because higher transmit power improves the received signal-to-noise ratio and increases the achievable rate. Compared with the benchmark methods, the GNN adaptively coordinates BS active beamforming and RIS passive beamforming at different power levels and better uses the reflection gain provided by the RIS. Therefore, the GNN maintains a consistent performance advantage across the full power range. Even in the high-power region, it outperforms APG and LAO, which further verifies its robustness (Fig. 10 ).When the number of RIS elements varies, the GNN maintains a clear performance advantage over both APG and LAO. In general, the system SE increases with the number of RIS elements because a larger RIS provides higher array gain and improves the equivalent channel conditions. According to the numerical results, the proposed GNN achieves a spectral efficiency of 2.066 5 bit/(s·Hz), which is approximately 7.46% and 3.79% higher than those of LAO and APG, respectively. Meanwhile, the average computational time of the GNN is only about 0.007 5 s, which is approximately 4% of that required by the benchmark methods. These results demonstrate that the proposed GNN effectively uses the performance gain provided by RIS scaling and achieves a good balance between system performance and computational complexity (Fig. 11 and Table 2 ).The relationship between system SE and the number of user groups is then examined under fixed settings for the number of transmit antennas and users. The generalization ability of the proposed GNN is further evaluated by training the model at a fixed transmit power and testing it over a wide transmit power range from 0 to 20 dBm. The training and testing curves almost overlap, indicating that the proposed GNN generalizes well to unseen transmit power levels. Across the full power range, the GNN consistently outperforms the LAO and APG benchmarks, further confirming its robustness and adaptability under different transmission conditions (Fig. 13 ). Conclusions For the RIS-assisted MISO-OFDMA system, this paper formulates a joint optimization problem for subcarrier allocation, BS active beamforming, and RIS passive beamforming to maximize the system SE. A model-driven GNN method is proposed to solve this problem. Comparative experiments with benchmark algorithms are conducted to validate the proposed method. The results demonstrate that the proposed GNN algorithm consistently outperforms LAO and APG in overall performance. It also exhibits strong robustness under different numbers of user groups and transmit power settings, which supports its potential for practical deployment in complex engineering scenarios.
2026, 48(6): 2327-2340.
doi: 10.11999/JEIT250884
Abstract:
Objective Direction Of Arrival (DOA) estimation is affected by low Signal-to-Noise Ratios (SNR), the coexistence of Circular Signals (CSs) and Non-Circular Signals (NCSs), and multiple forms of array imperfections. Conventional subspace-based estimators exhibit model mismatch in such environments and show reduced accuracy. Although neural-network methods provide data-driven alternatives, the effective use of the distinctive statistical properties of NCSs and the maintenance of robustness against diverse array errors remain insufficiently addressed. The objective is to design a DOA estimation algorithm that operates reliably for mixed CSs and NCSs in the presence of array imperfections and provides improved estimation accuracy in challenging operating conditions. Methods A robust DOA estimation algorithm is proposed based on an improved Vision Transformer (ViT) model. A six-channel image-like input is first constructed by fusing features derived from the covariance matrix and pseudo-covariance matrix of the received signal. These channels include the real component, imaginary component, magnitude, phase, magnitude ratio reflecting the NCS characteristic, and the phase of the pseudo-covariance matrix. A gradient-masking mechanism is introduced to adaptively fuse core and auxiliary features. The ViT architecture is then modified: the standard patch-embedding module is replaced with a convolutional layer to extract local information, and a dual-class-token attention mechanism, placed at the sequence head and tail, is designed to enhance feature representation. A standard Transformer encoder is used for deep feature learning, and DOA estimation is performed through a multi-label classification head. Results and Discussions Extensive simulations are carried out to assess the proposed algorithm (6C-ViT) against MUSIC, NC-MUSIC, a Convolutional Neural Network (6C-CNN), a Residual Network (6C-ResNet), and a MultiLayer Perceptron (6C-MLP). Performance is evaluated using Root Mean Square Error (RMSE) and angular estimation error under different operating conditions. Under single-source scenarios with low SNR and no array errors, 6C-ViT achieves near-zero RMSE across most angles and shows minor edge deviations (Fig. 2 ). It maintains the lowest RMSE across the SNR range from –20 dB to 5 dB (Fig. 3 ), indicating good generalization to unseen SNR levels. In dual-source scenarios containing mixed CS and NCSs under array errors, 6C-ViT shows clear advantages. Its estimation errors fluctuate slightly around zero, whereas competing techniques present larger errors and pronounced instabilities, especially near array edges (Fig. 4 ). Its RMSE decreases steadily as SNR increases and reaches below 0.1° at high SNR, while traditional approaches saturate around 0.4° (Fig. 5 ). Robust behavior is further observed across different numbers of signal sources (K = 1, 2, 3) and snapshot counts (100 to 2 000). 6C-ViT preserves high accuracy and stability under these variations, whereas other methods show marked degradation or instability, most evident at low snapshot counts or with multiple sources (Fig. 6 ). When evaluated using unknown modulation types, including UQPSK with a non-circularity rate of 0.6 and 64QAM, under array errors, 6C-ViT continues to produce the lowest RMSE across most angles (Fig. 7 ), demonstrating strong generalization capability. Ablation studies (Fig. 8 ) confirm the contributions of the six-channel input, the gradient masking module, the convolutional embedding, and the dual class token mechanism. The complete configuration yields the highest accuracy and the most stable performance. Conclusions Strong robustness is demonstrated in complex scenarios that contain mixed CS and NCSs, multiple array imperfections, low SNR, and closely spaced sources. By fusing multi-dimensional features of the received signal and using an enhanced Transformer architecture, the algorithm attains higher estimation accuracy and improved generalization across different signal types, error conditions, snapshot counts, and noise levels compared with subspace- and neural-network-based baselines. The method provides a reliable DOA estimation solution for demanding practical environments.
2026, 48(6): 2341-2351.
doi: 10.11999/JEIT251355
Abstract:
Objective Radar signal sorting is a critical step in electronic reconnaissance and battlefield situational awareness. It is used to accurately separate interleaved pulse streams in complex electromagnetic environments. Although multi-station cooperative reconnaissance systems provide spatial diversity gains that can mitigate the parameter ambiguity and aliasing problems of single-station systems, their practical deployment faces major challenges. Traditional centralized processing architectures require massive volumes of raw Pulse Description Word (PDW) data to be transmitted to a central server. This requirement leads to prohibitive communication bandwidth costs and increases the risk of leakage of sensitive electromagnetic spectrum intelligence. In addition, because stations are geographically distributed and differ in antenna scanning patterns, the data collected at different stations often show significant Non-Independent and Identically Distributed (Non-IID) characteristics. Such heterogeneity reduces the generalization ability of local models trained on isolated data islands. To resolve the conflict between data isolation and the need for collaborative intelligence, a multi-station collaborative radar signal sorting method is proposed based on a Federated Learning (FL) framework. Collaborative model training is enabled without exchange of raw data, so that data privacy is preserved, communication overhead is reduced, and sorting robustness is improved in heterogeneous and noisy battlefield environments. Methods A centralized federated sorting framework is constructed to coordinate multiple reconnaissance stations. The method contains three main components: feature preprocessing, a lightweight local temporal model, and a heterogeneity-aware aggregation strategy. First, in data preprocessing, the raw PDW parameters, including TOA, CF, and PW, are normalized to address substantial differences in scale. Specifically, TOA is transformed into first-order differential values to extract Pulse Repetition Interval (PRI) information, which prevents numerical saturation and captures periodic patterns effectively (Fig. 3 ). Second, a local time-series sorting model is designed for the resource constraints of edge devices. A bidirectional Long Short-Term Memory (LSTM) network is used as the backbone to capture long-range dependencies and dynamic patterns in pulse sequences from both forward and backward directions. To accelerate convergence and prevent gradient vanishing, residual connections are added to fuse static and dynamic features. The extracted features are then mapped to the radiation source category space through a cascaded linear classification layer. Third, to address model drift caused by Non-IID data, including feature distribution shift and label distribution shift, a new aggregation strategy is proposed based on parameter decomposition and proximal regularization. Model parameters are decoupled into a feature extractor and a classifier. During federated aggregation, only the parameters of the generic feature extractor are uploaded and globally averaged, whereas the personalized classifier parameters are retained locally to adapt to the class distribution of each station. Furthermore, a proximal regularization term is added to the local loss function (Eq. 20). This constraint limits the deviation of local updates from the global model and ensures that the optimization direction does not diverge substantially because of local data heterogeneity, thereby improving the stability and convergence speed of the global model. Results and Discussions Extensive simulation experiments are conducted on core datasets with 3 stations and 5 radars, and on extended datasets with 9 stations and 12 radars, including complex modulation patterns such as jitter, sliding, and staggering. Quantitative analysis shows that the proposed method achieves sorting performance comparable to that of Centralized Learning (CL). On the core dataset, the Precision, Recall, and F1-score of the proposed method reach 96.51%, 96.35%, and 96.42%, respectively, exceeding those of FedAvg by approximately 0.67% in F1-score. On the more challenging extended dataset, the performance advantage becomes more significant, with an F1-score improvement of 3.86% over FedAvg (Table 4 ). These results indicate that the parameter decomposition strategy effectively balances common feature learning with personalized decision-making. Analysis by class further shows that, for categories that are difficult to distinguish, such as Radar 7 and Radar 10, the proposed method improves recognition accuracy by up to 15% and 6%, respectively, compared with FedAvg (Fig. 7 and Fig. 8 ). Robustness tests further demonstrate the adaptability of the method. When the number of participating stations increases from 3 to 9 (Fig. 9 ), the F1-score rises steadily from 73.53% to 83.75%. This result confirms that enlarging node scale in the FL framework produces collaborative gains through more diverse samples and reduced geographic statistical heterogeneity, which substantially improve model generalization and robustness. Under severe class skew conditions, the method maintains an F1-score above 80% on the core dataset (Fig. 10 and Fig. 11 ). Furthermore, under extreme electromagnetic conditions characterized by high pulse loss rates of 70% and spurious pulse rates of 70%, the model maintains sorting performance above 75%, which demonstrates strong robustness against noise and interference (Fig. 12 ). Conclusions An FL-based framework is proposed for multi-station collaborative radar signal sorting to address data privacy and transmission constraints in distributed reconnaissance. By integrating a lightweight LSTM with a heterogeneity-aware aggregation mechanism, the method effectively captures temporal pulse features and mitigates model drift caused by Non-IID data. Experimental results verify that the approach achieves accuracy comparable to that of centralized methods and shows superior robustness under label skew and severe data degradation, including high pulse loss and spurious pulse rates. This study provides a privacy-preserving and efficient solution for intelligent signal processing in distributed electronic warfare systems.
2026, 48(6): 2352-2363.
doi: 10.11999/JEIT250697
Abstract:
Objective The verification and validation of radar target detection algorithms, especially for maritime surveillance, require high-fidelity synthetic sea clutter data. However, realistic sea clutter generation under high sea-state conditions, such as Sea State 4 and above, remains challenging because sea clutter is nonstationary and non-Gaussian. Traditional statistical models often fail to capture complex time-frequency characteristics, particularly when direct measurements are difficult to obtain. To address small-sample sea clutter generation, this study proposes a physics-constrained deep learning framework with adaptive transfer learning. The aim is to generate high-quality synthetic sea clutter data that closely match measured data and provide a reliable data basis for the development and testing of advanced radar systems. Methods The proposed framework integrates a Complex Variational Autoencoder Wasserstein Generative Adversarial Network (CVAE-WGAN) with transfer learning for high-fidelity sea clutter generation under small-sample conditions. The model operates in the complex domain and jointly processes in-phase and quadrature components, preserving signal orthogonality and phase relationships. An Amplitude-Phase Attention (APA) module is designed to enhance joint amplitude-phase modeling, and complex residual blocks are used to improve gradient propagation and training stability. A physics-constrained loss system, including time-frequency ridge loss and weighted Doppler-band loss, is developed to guide the generated data toward the physical characteristics of sea clutter. To reduce dependence on target-domain samples, an adaptive transfer learning mechanism based on Kullback-Leibler Divergence (KLD) is used to dynamically adjust fine-tuning and support knowledge transfer across different sea-state scenarios. Results and Discussions The proposed CVAE-WGAN framework is evaluated on measured sea clutter datasets and shows strong performance in synthetic data generation. In the source domain, namely Sea State 4, the generated data match the measured data in amplitude statistics, temporal correlation, spectral characteristics, and joint time-frequency characteristics. The probability density function-cosine similarity (PDF-CS) is 0.872 (Fig. 5 , Table 1 ), the autocorrelation function-cosine similarity (ACF-CS) is 0.938 2 (Fig. 7 , Table 1 ), and the spectrum root mean square error (SPEC-RMSE) is 4.537 9 dB (Fig. 6 , Table 1 ). The joint time-frequency accuracy reaches 95.31% when |z|≤1.5 (Fig. 8 , Table 1 ). The adaptive transfer learning mechanism is further validated by transferring the pretrained model to Sea State 5 using only 20% of the target-domain samples. The generated clutter maintains a close fit to the measured amplitude distribution (PDF-CS = 0.844 8) (Fig. 9 , Table 3 ) and shows good autocorrelation characteristics (ACF-CS = 0.955 7) (Fig. 10 , Table 4 ). Its spectral characteristics are also close to those of the measured data, with SPEC-RMSE = 4.498 2 dB under the 20% target-domain setting (Fig. 11 , Table 4 ). The joint time-frequency accuracy reaches 77.8% when |z|≤1.0 and 92.18% when |z|≤1.5 (Fig. 12 , Table 4 ). Ablation experiments show that the APA module is essential for joint amplitude-phase modeling. Removing this module reduces PDF-CS by 17.3% and increases SPEC-RMSE by 25.9% (Table 2 ). The sample-size sensitivity test shows that the method remains stable when the target-domain sample proportion is no less than 15%, with PDF-CS > 0.6 and joint time-frequency accuracy > 82% when |z|≤1.5 (Table 4 ). These results indicate that the proposed method is suitable for sea clutter generation in data-scarce scenarios. Conclusions This study presents a framework for high-fidelity sea clutter generation under small-sample conditions. The framework combines CVAE-WGAN, physics-constrained learning, and adaptive transfer learning. Guided by the proposed loss system, the model effectively captures both the statistical and physical characteristics of high sea-state clutter. The KLD-based adaptive transfer strategy improves cross-sea-state adaptability and supports high-quality data generation with limited target-domain samples. The proposed framework provides a reliable and scalable data basis for the development and testing of radar anti-interference algorithms. Future work will further optimize the method for extremely scarce samples and extend it to other nonstationary radar signal scenarios.
2026, 48(6): 2364-2373.
doi: 10.11999/JEIT251017
Abstract:
Objective Supervised short-wave signal modulation recognition methods generally assume identical distributions between source-domain training data and target-domain test data. Short-wave channels are affected by ionospheric variation, which creates substantial distribution discrepancies across domains and reduces model performance. Deployment on unmanned edge platforms is further restricted by limited computational resources, scarce labeled samples, and data-privacy requirements. This study proposes a lightweight recognition method based on source-model transfer that enables privacy-preserving model adaptation without access to source-domain data. Methods A Multi-Modal Source-Model Transfer Framework (M-SMOT) is developed. It applies information-maximization loss and a self-supervised pseudo-labeling strategy to support model adaptation without revisiting source-domain data. The method achieves cross-channel recognition of short-wave modulation signals with reduced computational cost while maintaining data privacy. Multi-modal information—including in-phase/quadrature (I/Q) components, amplitude-phase (AP) characteristics, and spectral features—is fused to exploit complementary representations and improve robustness under complex channel variation. Results and Discussions Experiments show that the proposed method consistently outperforms the Source-Only baseline across six cross-channel scenarios, with accuracy gains from 0.31% to 10.81% (Table 1 ). In few-shot adaptation, average recognition accuracies reach 98.3% and 96% of the full-sample baseline when target-domain samples are reduced to 10% and 1%, respectively (Fig. 12 ). Ablation studies confirm the effectiveness of the self-supervised pseudo-labeling module (Fig. 16 ) and the multi-modal fusion strategy (Fig. 17 ). The lightweight design is verified by zero source-data storage, a peak memory footprint of 6.00 MB, and convergence within one fine-tuning epoch (Table 2 ). These findings show that the method mitigates domain discrepancies and protects privacy under resource-limited conditions. Conclusions The M-SMOT method integrates data-privacy protection, source-model adaptation, few-shot generalization, and low resource consumption. It provides a practical solution for cross-channel modulation recognition in short-wave communication and is suited for deployment on resource-constrained edge devices.
2026, 48(6): 2374-2383.
doi: 10.11999/JEIT251023
Abstract:
In dense and structurally complex urban environments, Unmanned Aerial Vehicle (UAV) swarm radiation source localization is affected by signal attenuation, multipath propagation, and building obstructions. To address these limitations, a dynamic formation-optimization method for UAV swarms is proposed. By improving the geometric configuration of the swarm, the method reduces path loss and interference, which strengthens localization accuracy. Received signal strength is used to evaluate signal quality in real time and supports adaptive formation adjustments that improve propagation conditions. Geometric dilution of precision and root mean square error metrics are integrated to refine swarm geometry and improve distance-estimation reliability. Simulation results show that the proposed method converges faster and improves localization accuracy in complex urban environments, reducing errors by more than 80 percent. The method adapts to environmental variation and demonstrates strong robustness and practical value. Objective UAV swarm localization and formation control in urban environments are affected by obstacles, signal attenuation, and rapid variation in the surroundings that reduce the reliability of conventional methods. This study proposes a radiation source localization approach that integrates the Received Signal Strength Indicator (RSSI) with dynamic formation adjustment to improve localization accuracy and strengthen system robustness in complex urban scenarios. RSSI is used once in full form, then referenced consistently. Methods The method uses RSSI measurements to estimate the distance to the radiation source and adjusts UAV swarm formation in real time to reduce localization errors. These adjustments are based on feedback that reflects relative positions, signal strength, and environmental variation. Localization accuracy is strengthened through a multi-sensor fusion strategy that integrates GPS, IMU, and depth-camera data. A data-quality assessment mechanism evaluates signal reliability and triggers formation adaptation when the signal drops below a predefined threshold. This optimization process reduces positioning errors and improves system robustness. Results and Discussions Simulation experiments in an ROS-based environment were conducted to evaluate the UAV swarm localization method under urban obstacles and multipath conditions. The swarm began in a hexagonal formation and adjusted its geometry according to environmental variation and localization confidence (Fig. 3 , Fig. 4 ). As shown in Fig. 5, localization errors fluctuated during initialization but converged to below 1 m after 150 s. Formation comparisons (Fig. 6 ) showed that symmetric structures such as hexagonal and triangular formations maintained errors below 0.5 m, whereas asymmetric formations (T and Y shape) produced deviations up to 4.9 m. Further comparisons (Fig. 7 ) showed that traditional RSSI saturated near 15 m, direction of arrival fluctuated between 5 and 14 m, and time difference of arrival failed due to synchronization problems. The proposed method achieved sub-meter accuracy within 60 s and remained robust throughout the mission. These findings indicate that combining RSSI-based distance estimation with dynamic formation adjustment improves localization accuracy, convergence speed, and adaptability under complex environmental conditions. Conclusions This study addresses UAV swarm localization in complex urban environments by integrating RSSI-based distance estimation, dynamic formation adjustment, and multi-sensor fusion. ROS-based simulations show that: (1) localization errors converge rapidly to sub-meter levels, reaching below 1 m within 150 s under non-line-of-sight conditions; (2) symmetric formations such as hexagonal and triangular configurations outperform asymmetric ones and reduce errors by up to 67 percent compared with fixed Y-shaped formations; and (3) relative to traditional RSSI, direction of arrival, and time difference of arrival approaches, the proposed method shows faster convergence, higher stability, and stronger robustness.
2026, 48(6): 2384-2395.
doi: 10.11999/JEIT251144
Abstract:
Objective Blind Source Separation (BSS) of single-channel co-channel overlapped communication signals remains challenging in non-cooperative reception. Conventional multi-channel methods are not applicable because of antenna limitations. Existing deep learning methods also show limited long-sequence modeling ability, high computational cost, and reduced performance for signals with small carrier frequency offsets. These limitations restrict the practical use of BSS techniques in dense electromagnetic environments. An efficient and robust framework is therefore needed to capture long-range temporal dependencies while maintaining computational feasibility. Methods S4-UNET integrates the U-NET encoder-decoder framework with the Structured State Space sequence model (S4). A Temporal State Enhancement Module (TSEM) is designed as the backbone block of both the encoder and decoder. It extracts local temporal features through residual learning. To model long-range dependencies, S4 is embedded in the odd-numbered stages of the encoder. This design captures global temporal correlations with near-linear computational complexity. S4 converts sequence modeling into a state-space evolution process and uses the Fast Fourier Transform (FFT) for efficient convolution. Skip connections and the Gated Linear Unit (GLU) are used to preserve fine-grained local details. Multi-scale feature fusion is achieved through skip connections between corresponding encoder and decoder stages. Signal resolution is then progressively restored by interpolation-based upsampling. The model also adaptively tokenizes feature maps in the temporal or channel dimension according to feature scale, which improves sequence representation. Results and Discussions Experiments are conducted on simulated datasets with small carrier frequency offsets, including same-modulation mixtures, mixed-modulation mixtures, and different-bandwidth mixtures. Public benchmark datasets and a measured dataset collected using hardware are also used. Quantitative results and visualizations (Fig. 3 , Fig. 5 , Table 5 ) show that S4-UNET consistently outperforms representative deep learning baselines, including ConvTasNet and CTDCRN, and the classical Time-Delay Embedding Independent Component Analysis (TDE-ICA) algorithm across different signal lengths and modulation schemes. The model maintains robust separation fidelity under randomly distributed carrier frequency offsets and initial phase differences (Table 3 ), confirming its strong generalization ability. Ablation and sensitivity analyses (Table 6 , Table 7 , Table 8 ) show that placing S4 in the odd-numbered encoder stages, using suitable convolutional stride settings, and adopting GLU jointly support a favorable balance between separation accuracy and computational efficiency. The model also maintains competitive inference latency while processing both long and short sequences, indicating its practical value. Conclusions S4-UNET addresses the main challenges of single-channel co-channel BSS by combining multi-scale convolutional feature extraction with efficient state-space long-sequence modeling. It achieves superior separation performance, strong robustness to small carrier frequency offsets, and good generalization across different data domains. The present work focuses on dual-source mixtures. Its modular architecture provides a basis for future extensions to mixtures with an unknown number of sources by integrating source number estimation and iterative cancellation strategies.
2026, 48(6): 2396-2404.
doi: 10.11999/JEIT251134
Abstract:
Objective Lightning identification is essential for lightning observation, location, warning, and disaster prevention. Large volumes of Low-Frequency/Very-Low-Frequency (LF/VLF) Lightning Electromagnetic Pulse (LEMP) waveform data require automatic and accurate classification methods. Deep learning has been widely used for feature extraction and classification, providing a feasible approach for LEMP waveform identification. However, anthropogenic electromagnetic interference and natural LEMP signals often overlap in the time and frequency domains. Their waveform features are also complex and diverse, which limits the accuracy and generalization ability of existing identification algorithms. Therefore, a more efficient deep learning model is required to distinguish LEMP signals from non-lightning electromagnetic signals. Methods This paper proposes a Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM) deep neural network model that integrates multi-scale residual convolution and cross-layer feature fusion. The model is designed for binary classification of LEMP and non-lightning electromagnetic signals and enables accurate diagnostic identification of LEMP signals. Using observational data from an LF/VLF lightning magnetic-field detection system, a multi-scale residual network is first used to extract multidimensional features from electromagnetic waveforms layer by layer. The time-domain features output by each convolutional layer are then organized into a cross-layer time-domain feature sequence according to network depth. This sequence is input into the LSTM module for adaptive weighted fusion. This mechanism uses the sequence modeling ability of LSTM to learn the relative importance of features at different hierarchical levels, rather than to model the temporal dynamics of the original waveform. Results and Discussions The proposed CNN-LSTM model achieves a precision of 100%, a recall of 99.82%, an F1-score of 99.91%, and an accuracy of 99.89%. It obtains the best performance across all evaluation metrics. The model effectively identifies LEMP samples and reduces the misclassification of non-lightning samples. The Bayes classifier achieves high precision (93.14%), but its recall is relatively low (80.14%). The Support Vector Machine (SVM) model improves on the Bayes classifier across all metrics, but it remains inferior to the proposed CNN-LSTM model. The MultiLayer Perceptron (MLP) and K-Nearest Neighbor (KNN) models also show limitations in precision, recall, and accuracy compared with CNN-LSTM. The Decision Tree (DT) model obtains reasonable results, but its precision and recall are lower than those of MLP and KNN, with a recall of only 88.01%. These results indicate that CNN-LSTM has clear advantages in LEMP waveform identification. This improvement is mainly attributed to the multi-scale residual CNN module, which automatically extracts low-level local features from raw waveform data. Additionally, the LSTM-based adaptive weighted fusion mechanism is applied to feature sequences from different network layers. As a feature integration tool across network depths, its input is an inter-layer feature sequence rather than an original waveform time series. This design improves the flexibility and discriminative ability of feature fusion, enables the model to learn the relative importance of features at different network depths, and supports effective aggregation of discriminative features. A confusion matrix was also generated to evaluate classification performance on the test set. Overall, comparison with baseline models confirms the superiority of the proposed model for LEMP waveform identification. Conclusions The CNN-LSTM model effectively identifies LEMP samples and reduces the misclassification of non-lightning samples. Compared with baseline models, it shows excellent identification performance in the binary classification of LEMP and non-lightning electromagnetic signals. The results also verify the effectiveness of convolutional feature extraction and LSTM-based cross-layer feature fusion for LEMP waveform identification.
2026, 48(6): 2405-2417.
doi: 10.11999/JEIT250935
Abstract:
Objective Target detection in nonzero-mean clutter is a central problem in radar and hyperspectral imaging systems. Conventional detectors that assume zero-mean clutter show reduced effectiveness when clutter exhibits nonzero-mean characteristics caused by environmental conditions or interference. The objective of this work is to design adaptive detectors that remain robust in nonzero-mean clutter and to analyze their statistical performance under signal-mismatch conditions. Methods Three adaptive detectors are derived based on the Generalized Likelihood Ratio Test (GLRT), the Rao test, and the Wald test. The detectors are formulated to accommodate unknown clutter mean and covariance matrix by estimating these quantities from training samples. A generalized signal-mismatch scenario is examined, in which the actual signal steering vector may deviate from the nominal vector. Analytical expressions for the Probability of Detection (PD) and False Alarm (PFA) are obtained for each detector to assess statistical performance. Results and Discussions Analytical expressions for the probabilities of detection and false alarm of the three detectors are validated through Monte Carlo simulation. All detectors maintain the Constant False Alarm Rate (CFAR) property. The amplitude characteristic of nonzero mean does not directly affect detection performance; rather, it exerts an effect through the loss factor of the output Signal-to-Clutter Ratio (SCR) and the Degrees Of Freedom (DOFs) of the detectors’ statistical distributions. Numerical evaluations using simulated and real data show that the proposed detectors achieve better performance than conventional approaches. Conclusions The three CFAR adaptive detectors based on the GLRT, the Rao test, and the Wald test are effective for target detection in nonzero-mean clutter. The nonzero mean of clutter affects detection performance in two ways: it reduces the optimal output signal-to-clutter ratio of the detectors and decreases the DOF of their statistical distributions. Simulated data show that, when there is no signal mismatch, the GLRT-NMC detector provides the highest PD. When measured data are used and no signal mismatch is present, either the Rao-NMC or Wald-NMC achieves a higher PD than the GLRT-NMC. Under signal-mismatch conditions, for both measured and simulated data, the Rao-NMC demonstrates the best sensitivity to mismatch, whereas the Wald-NMC exhibits the strongest robustness.
2026, 48(6): 2418-2428.
doi: 10.11999/JEIT251247
Abstract:
Objective With the rapid growth of wireless devices and communication services, spectrum resources have become increasingly scarce. Spectrum sensing, as a fundamental function of cognitive radio, enables dynamic spectrum access and improves spectrum utilization efficiency. However, conventional spectrum sensing methods based on circular signal assumptions cannot effectively detect noncircular signals. In addition, some detectors designed for noncircular signals show degraded performance under low Signal-to-Noise Ratio (SNR) or limited sample conditions. To address these limitations, a nonparametric spectrum sensing scheme based on the Weighted Pearson Correlation Coefficient (WPCC) is proposed. The scheme applies a linear fusion strategy to the real-valued composite coherence matrix, which captures the second-order statistical characteristics of noncircular signals. Methods The WPCC detector constructs a real-valued composite observation vector and computes the corresponding composite coherence matrix. Pearson Correlation Coefficients (PCCs) are extracted from this matrix to characterize the statistical properties of noncircular signals. The first two product moments of squared sample PCCs are derived, and optimal fusion weights are obtained based on the deflection coefficient. The true PCCs are approximated by their sample estimates to obtain data-driven fusion weights that do not require prior knowledge of sensing channels. These weights are then linearly combined with the squared sample PCCs to construct the WPCC test statistic, thereby exploiting the spatial diversity of sensing antennas. The final decision is made by comparing the WPCC statistic with a sensing threshold determined by the specified false alarm probability. Specifically, a WPCC value below the threshold indicates the null hypothesis of an idle frequency band, whereas a value above the threshold indicates the alternative hypothesis that the frequency band is occupied by primary users. Results and Discussions Simulation experiments evaluate the sensing performance of the proposed nonparametric WPCC-based method (Algorithm 1) in terms of sensing probability, deflection coefficient, Receiver Operating Characteristic (ROC) curve, and Area Under the Curve (AUC), with comparisons to NCLMPIT, NCAGM, NCHDM, and NCJT. The numerical results show that the proposed method outperforms the compared detectors under various simulation conditions. In particular, the WPCC detector achieves the highest sensing probability and exhibits superior performance at low false alarm probabilities of 0.05 (Fig. 2 ), 0.01 (Fig. 3(a) ), and 0.005 (Fig. 3(b) ), with sample sizes not exceeding 100. In addition, the proposed method shows clear advantages under different numbers of antennas (Fig. 4 ), different noise variance conditions (Fig. 5 ), and different levels of correlation strength (Fig. 6 ). The applicability of the WPCC method to circular signals is also demonstrated by its high sensing probability for QPSK and 16PSK signals (Fig. 7 ). The superior overall performance of the proposed detector is further confirmed by higher deflection coefficient curves and ROC curves (Figs. 8, 9 ). The largest AUC values quantitatively demonstrate its overall optimality among all considered methods (Table 1 ). These results indicate strong robustness under low SNR and small-sample conditions. Conclusions A Pearson correlation fusion sensing method for noncircular signals is proposed based on the real-valued composite covariance representation and the Locally Most Powerful Invariant Test (LMPIT) framework. By combining optimal fusion weights derived from sample PCCs with a linear weighting scheme, the method fully exploits second-order statistical information. It enhances strongly correlated components while suppressing weak correlations and noise interference. Analytical expressions for the false alarm probability and sensing threshold are derived. Both theoretical analysis and simulation results show that the proposed method achieves superior performance compared with existing noncircular signal sensing methods in terms of sensing probability, deflection coefficient, ROC curve, and AUC.
2026, 48(6): 2429-2437.
doi: 10.11999/JEIT251383
Abstract:
Objective Direction-of-Arrival (DOA) estimation often fails under mixed non-Gaussian noise, particularly for coherent sources. To address this problem, a Multi-Channel Feature Comprehensive Fusion Network (MCFCF) based on the Maximum Correntropy Criterion (MCC) is proposed. By integrating Multi-Scale Feature Extraction (MSFE) and Spatial-Frequency Attention Fusion (SFAF), the framework performs joint spatial-frequency calibration. MCC is used instead of Mean Squared Error (MSE) to improve robustness against impulsive outliers. Experiments show that MCFCF outperforms Multiple Signal Classification (MUSIC), Estimation of Signal Parameters via Rotational Invariance Techniques (ESPRIT), and state-of-the-art deep networks in accuracy and super-resolution performance, especially under low signal-to-noise ratio (SNR) and coherent-source conditions. Methods The proposed MCFCF framework is designed for DOA estimation in mixed non-Gaussian noise and coherent-source environments. Forward-Backward Spatial Smoothing (FBSS) is used for rank recovery. MSFE and SFAF are then combined for physical manifold calibration through multi-scale feature extraction and spatial-frequency attention. The MCC loss function is adopted to form a statistically robust optimization objective against impulsive outliers. Results and Discussions Experimental analysis verifies the performance of MCFCF in complex environments (Figs. 5 -8 ). Under mixed noise with strong impulsive and heavy-tailed components, traditional subspace methods fail because of subspace swapping, with root mean square error (RMSE) values exceeding 10°. By contrast, MCFCF maintains an RMSE of approximately \begin{document}$ {1.5}^{\circ } $\end{document} Fig. 5 ). Adaptability tests across heterogeneous noise distributions show strong distribution invariance (Fig. 6 ). In the white-noise benchmark, MCFCF compensates for FBSS-induced aperture loss and finite-sample variance, outperforming MUSIC in the low-SNR region. Mechanism analysis indicates that the multi-scale receptive fields of MSFE improve stability, while the amplitude-phase decoupling of SFAF improves estimation precision (Figs. 7 ～8 ). In addition, the MCC-based soft-truncation mechanism suppresses heavy-tailed impulsive interference at the optimization level, thereby improving global robustness. Conclusions An MCC-based MCFCF framework is proposed to address DOA estimation failures in mixed non-Gaussian noise and coherent-source environments. FBSS is used for rank recovery, and MSFE and SFAF are integrated for manifold calibration. The MCC loss further suppresses impulsive outliers. The results confirm that the proposed method achieves higher accuracy and stronger distribution invariance at –15 dB SNR than MUSIC, ESPRIT, and state-of-the-art deep learning baselines. These findings demonstrate its robustness in hostile electromagnetic environments.
2026, 48(6): 2438-2447.
doi: 10.11999/JEIT251371
Abstract:
Objective Communication and sensing systems are moving toward higher frequency bands, larger antenna arrays, and smaller hardware. Their hardware architectures, channel characteristics, and signal processing methods are therefore becoming increasingly similar. This trend supports Integrated Sensing And Communication (ISAC), in which joint estimation of channel and sensing target parameters has become a key research topic. Existing studies have achieved joint estimation of these two parameter categories within a unified tensor framework, but two limitations remain. First, most studies focus on parameter estimation and do not further convert multidimensional estimates into accurate localization of Scatterer Points (SPs), the Mobile Transmitter (MT), and sensing targets. This limitation prevents a complete spatial characterization of the wireless propagation environment. Second, the fusion of channel and sensing target parameter information has received limited attention, which restricts further improvement in parameter estimation and localization accuracy. Methods To address channel/sensing target parameter estimation and localization in millimeter-Wave (mmWave) Multiple-Input Multiple-Output (MIMO) ISAC systems, this paper proposes a tensor decomposition algorithm based on information fusion. First, a unified fourth-order PARAllel FACtor (PARAFAC) model is constructed at the Base Station (BS) for uplink channel and sensing target parameter estimation. To reduce computational complexity, the fourth-order tensor model is transformed into a third-order form, and the Trilinear Alternating Least Squares (TALS) method is used to estimate three factor matrices. The special structure of one factor matrix is then exploited. A closed-form decomposition is used to decouple the coupled factor matrix, and Angle of Departure (AoD), Angle of Arrival (AoA), time delay, Doppler shift, and coefficients are extracted from the four estimated factor matrices. Based on these estimates, the MT, SPs, and sensing targets are localized separately using geometric relationships. The estimation accuracy of SPs is further improved by fusing Doppler shift and position information from SPs and sensing targets. The Cramér-Rao Bound (CRB) is derived as a theoretical performance benchmark for the five types of parameters. Results and Discussions The first simulation experiment shows that the proposed algorithm and the Optimized Quadrilinear Alternating Least Squares (Op-QALS) algorithm outperform the Co-SVD-BALS algorithm in terms of Root Mean Square Error (RMSE) for channel/sensing target parameter estimation and localization (Fig. 2 , Fig. 3 , Fig. 4 ). With information fusion, the proposed algorithm achieves the best Doppler shift and position estimation performance for SPs (Fig. 2(d) , Fig. 4(a) ). This advantage occurs because the proposed algorithm and Op-QALS fully exploit the multidimensional structure of the received signal. The fusion operation further improves the estimation capability of the proposed algorithm, whereas Co-SVD-BALS accumulates errors during stepwise factor matrix estimation. In terms of Average Processing Time (APT), the proposed algorithm requires slightly more time for localization than Co-SVD-BALS, but far less time than Op-QALS (Table 1 and Table 2 ). The proposed algorithm therefore achieves accurate parameter estimation and localization at a reasonable computational cost. The second simulation experiment shows that, under two Signal-to-Noise Ratio (SNR) levels, the localization accuracy of all algorithms improves as \begin{document}$ K $\end{document} Fig. 5 ). The fusion operation does not substantially increase the APT of the proposed algorithm (Fig. 5(d) ). The third simulation experiment indicates that increasing \begin{document}$ {M}_{\mathrm{RE}}\left(M_{\mathrm{RE}}^{\mathrm{s}}\right) $\end{document} \begin{document}$ N $\end{document} Fig. 6 ). Conclusions This paper proposes an information fusion algorithm for channel/sensing target parameter estimation and localization within a unified tensor framework. By exploiting the Vandermonde structure of a factor matrix, the proposed algorithm preserves estimation accuracy while reducing computational complexity. The fusion operation further improves SP parameter estimation and localization without a substantial increase in computational overhead. Future work will extend the algorithm to more general array configurations and examine higher-order tensor processing for multi-BS cooperation and multi-user access scenarios.
2026, 48(6): 2448-2458.
doi: 10.11999/JEIT251325
Abstract:
The seamless integration of active circuitry and antennas can effectively improve link performance and system integration. At present, active-integrated antennas are mainly designed by adjusting the antenna impedance while maintaining the desired radiation characteristics to achieve direct matching with active transistors. However, the effect of the antenna’s complex impedance on the phase response of the active channel, as well as its potential application in active-integrated phased arrays, has not been thoroughly studied. This paper proposes a channel phase self-compensation method for active-integrated arrays. For each active channel, the active transistor is directly integrated with the radiating element, where the load impedance at the transistor drain is matched to the input impedance of the antenna element. Under a constant active gain, the required complex load impedance is solved to establish an explicit mapping between the phase response of each active channel and its corresponding load impedance. According to the phase-shift requirements among array channels, appropriate load impedances are selected as the input impedances of the corresponding radiating elements. This approach applies a predefined phase distribution to each channel without using external phase-shifting structures. It can control the initial beam direction or compensate for the path difference between elements in conformal arrays. An active-integrated phased-array antenna with a preset beam direction is designed as a demonstration example to verify the effectiveness of the proposed method. The method provides an efficient design approach for next-generation active-integrated arrays. Objective In the traditional design approach, active circuit channels and antenna arrays are matched to 50 Ω before interconnection. This configuration occupies considerable physical space and limits system-level integration. In addition, insertion loss in passive matching networks and mismatch loss at the interconnections reduce overall link performance. Direct co-integration of active circuitry and antenna elements can address these limitations. However, multi-channel active-integrated antenna arrays often require one or multiple superimposed phase distributions across the channels to satisfy different application requirements, such as initial beam offset in fuze systems, wavefront compensation in conformal active phased arrays, and wide-angle beam scanning. These phase gradients are typically realized through backend phase-shifting networks. In this work, the complex impedance characteristics of the antenna are adjusted when it is directly integrated with the active circuitry. The phase response of the active-integrated channels can therefore be tuned within a certain range without using complex matching networks or additional phase shifters. This strategy reduces the complexity and performance requirements of the backend phase-shifting network. The advantages are more evident in millimeter-wave, high-frequency, and terahertz systems, where the available phase-shift range of phase shifters is limited. Methods Phase self-compensation of the active channels is achieved through the direct integration of the active transistor and the radiating element. In this configuration, the drain output of the transistor is directly connected to the input of the radiating element, and impedance transformation is realized within the antenna element. The proposed method includes three main steps. (1) The active transistor is first modeled as a two-port network. By evaluating the antenna element’s complex impedance as the load on different constant-gain circles, the mapping between the phase response of the active channel and the load impedance is established. The achievable phase-shift range of the active channel is then determined. (2) According to the required phase-shift distribution among the array channels, suitable combinations of active gain and corresponding complex load impedances (not unique) are selected. These combinations are not unique. (3) The realizability of the selected impedances is examined according to the characteristics of the radiating element. The impedance values with the highest feasibility are implemented by optimizing the radiating element, which includes fine adjustment of its geometry and feed position to meet the target impedance. When the radiating element is modified, particularly for circularly polarized elements, desirable radiation characteristics must also be preserved, including good axial ratio and beam-scanning performance. Results and Discussions The proposed phase self-compensation mechanism enables the array to achieve initial beam pointing and compensate for path-length differences caused by special array geometries, such as conformal or curved surfaces, without using additional phase-shifting structures. Therefore, the performance requirements of the backend phase-shifting network in active phased arrays can be reduced. To verify the effectiveness of the proposed method, a 1×4 circularly polarized active-integrated linear array (Fig. 9 ) is designed and demonstrated. Based on channel-level impedance calculations (Fig. 6 ) and an analysis of the antenna-element impedance characteristics (Fig. 8 ), a phase gradient of 38° between adjacent channels is synthesized and applied to the circularly polarized active-integrated array. Without degrading the circular polarization performance and without external phase-shifting circuitry, the initial beam direction of the active-integrated phased array is shifted to the desired angle of θ0 = 12° (Fig. 13 ). The phase self-compensation design does not degrade the beam-scanning capability of the array. After an additional phase gradient is applied for beam steering, the array achieves a scanning range of up to 50°. The gain reduction remains within 2 dB relative to the initial pointing direction, and the axial ratio remains below 4 dB throughout the scanning range. Conclusions Within the framework of active-integrated arrays, this work uses the phase-tuning effect produced by the complex impedance at the antenna port when the radiating element is directly matched to the active transistor. A desired phase-gradient distribution can therefore be synthesized among the channels of an active-integrated phased array within an achievable range. This capability enables compensation for required phase distributions, such as preset beam direction and path-length equalization in conformal-array applications, without relying on additional phase shifters. Therefore, the complexity and performance requirements of the backend phase-shifting circuitry are reduced. The effectiveness of the proposed method is validated through a multi-channel circularly polarized active-integrated phased-array prototype with a preset beam direction. Both full-wave simulations and experimental measurements confirm that the phase self-compensation mechanism provides the required initial beam pointing while preserving beam-scanning capability and polarization performance. This study provides a new approach for the design of high-efficiency next-generation active-integrated phased arrays.
2026, 48(6): 2459-2475.
doi: 10.11999/JEIT250715
Abstract:
Objective With the rapid development of 6G networks, communication systems are moving toward high bandwidth, low latency, and massive connectivity. In this context, Integrated Sensing and Communication (ISAC) allows wireless signals to transmit information and sense the environment. It provides a basis for more intelligent and efficient 6G services. Environmental reconstruction is a core task in ISAC systems. It aims to recover the true spatial structure of targets and scenes from echo signals. However, practical environmental reconstruction still faces three main challenges. First, dense Base Station (BS) deployment in 6G communication systems places building targets in the near-field region of the imaging system. This causes strong coupling among the range, azimuth, and elevation dimensions in tomographic imaging, resulting in clear deviations between the reconstructed target geometry and its true shape. Second, the positioning error of User Equipment (UE) is much larger than the signal wavelength. Therefore, conventional Synthetic Aperture Radar (SAR) autofocus algorithms become ineffective, and new methods are needed to avoid the effects of positioning errors. Third, conventional Tomographic Synthetic Aperture Radar (TomoSAR) algorithms use a per-channel processing framework, in which Single-Look Complex (SLC) images are generated independently for each channel. When Inverse Synthetic Aperture Radar (ISAR) techniques are used to generate SLC images, data differences among channels cause inconsistent translational phase compensation. These inconsistencies introduce inter-channel phase errors during elevation focusing and produce false targets in the imaging results. Methods This paper first applies the nonparametric translational motion compensation method developed for ISAR imaging to SLC image generation. This strategy avoids the adverse effects caused by UE positioning errors. Existing ISAR studies usually assume that the target satisfies a turntable model. However, the actual SAR imaging geometry differs from this ideal assumption. Therefore, the mathematical mapping between the TomoSAR imaging results and the true spatial coordinates of the target is rederived under the SAR imaging geometry. Based on this mapping, the coordinate reconstruction problem is formulated as a system of nonlinear equations. A coordinate reconstruction method based on Particle Swarm Optimization (PSO) is then proposed to accurately recover the true geometric shape of the target. In addition, to address inconsistent translational phase compensation among channels in conventional per-channel processing frameworks, a joint phase correction TomoSAR framework is designed. A unified phase correction strategy is used to remove inter-channel phase errors, thereby improving elevation focusing and overall imaging quality. Results and Discussions The proposed methods are verified through simulation experiments on complex building targets under ideal and non-ideal trajectory conditions. Chamfer Distance (CD) is used to evaluate coordinate reconstruction accuracy. The CD values under ideal and non-ideal trajectories are 1.62 and 1.68, respectively, which indicates only slight performance degradation under non-ideal conditions. The imaging point cloud under the non-ideal trajectory shows clear point dropout. A comparison of the empirical Cumulative Distribution Function (CDF) curves of distance errors under the two trajectory conditions shows that the overall trends are similar. Clear differences appear only when the distance error exceeds 2 m. This finding indicates that, under the CD metric, the main differences between the imaging results for ideal and non-ideal trajectories are concentrated in point-cloud dropout regions and areas outside the main target. Thus, non-ideal trajectories mainly affect the scattering-intensity distribution. Comparative experiments between the joint phase correction framework and the conventional framework further show that conventional TomoSAR methods produce clear stacking effects at different elevations, with false targets appearing at incorrect elevation levels. This result suggests that independent translational phase compensation in each channel can induce inter-channel phase errors and seriously degrade elevation focusing. By contrast, joint phase correction improves imaging quality. Conclusions The experimental results verify the effectiveness of the proposed methods. By combining ISAR-based nonparametric translational motion compensation with PSO-based coordinate reconstruction, the true geometric shape of the target is successfully recovered. The joint phase correction strategy also eliminates false targets in elevation focusing caused by conventional per-channel processing. It therefore improves elevation focusing capability and overall image quality.
2026, 48(6): 2476-2492.
doi: 10.11999/JEIT250662
Abstract:
Objective Accurate characterization of ionospheric variability is essential for reliable Global Navigation Satellite System (GNSS) positioning, especially during geomagnetic storms, when rapid and highly structured disturbances occur. Existing empirical and physics-based ionospheric models often have limited ability to represent storm-time ionospheric dynamics and small-scale irregularities in real time. This study develops a unified data-driven ionospheric modeling framework that uses GNSS-derived Slant Total Electron Content (STEC) time series as input and learns spatiotemporal mappings to key ionospheric parameters, including Vertical Total Electron Content (VTEC) and the Rate Of TEC Index (ROTI). By using deep operator learning, the proposed framework improves short-term ionospheric modeling and forecasting under disturbed conditions and provides more reliable ionospheric corrections for single-frequency positioning. Methods A unified data-driven ionospheric modeling framework, named DeepION, is proposed based on the Deep Operator Network (DeepONet) architecture. The framework uses STEC time series as the primary input and learns nonlinear spatiotemporal mappings to key ionospheric parameters. DeepION models and predicts STEC and VTEC, whereas ROTI is derived from the predicted STEC series. In the network design, a Convolutional Neural Network (CNN) is used as the branch network to extract spatiotemporal features from historical STEC time series. The trunk network uses a multilayer fully connected structure with periodic time encoding. Its inputs include GNSS observation geometry and temporal information, which allows the model to capture the continuous temporal dynamics of ionospheric behavior. During data preprocessing, a VTEC-based modeling strategy is first used to estimate and remove receiver Differential Code Bias (DCB), thereby providing high-quality STEC observations. The model is then trained and validated using GNSS observations collected during the May 2024 geomagnetic storm. Its outputs include ray-path STEC values, gridded VTEC fields, and derived ROTI time series. The proposed framework is further evaluated by incorporating model-derived VTEC corrections into Single Point Positioning (SPP) experiments. Modeled and observed ionospheric parameters are compared under both geomagnetically quiet and disturbed conditions to assess the modeling accuracy and practical performance of DeepION. Results and Discussions The experimental results show that DeepION robustly characterizes ionospheric spatiotemporal variability under different space weather conditions. It captures both large-scale structures and small-scale disturbances during geomagnetic storms. For STEC forecasting, the model achieves a Root Mean Square Error (RMSE) of 12.82 TECU over a 3-day prediction horizon and maintains high consistency with GNSS observations (Fig. 4 ). The model also predicts ionospheric irregularities accurately, as indicated by the close agreement between predicted and observed ROTI time series at the mid-latitude NVSK station (Fig. 5 ). For VTEC modeling, DeepION-generated global VTEC maps reproduce equatorial anomalies and storm-enhanced density regions. These maps closely match the Center for Orbit Determination in Europe Spherical Harmonic (CODE-SH) model and outperform the Klobuchar and NeQuick empirical models in spatial resolution and structural fidelity (Fig. 6 ). Further ray-path-level analysis shows that STEC derived from DeepION-based VTEC mapping yields the lowest residual error at the mid-to-high-latitude NLIB station. It achieves an RMSE of 6.80 TECU, outperforming Klobuchar and NeQuick and slightly improving on CODE-SH (Fig. 7 ). In GNSS positioning applications, the SPP results show that DeepION-derived ionospheric corrections consistently reduce positioning errors at both the CUSV and NLIB stations. The improvement is especially clear in the vertical and geometric components during storm-time conditions, indicating stronger robustness under intensified geomagnetic disturbances (Fig. 8 , Fig. 9 ). Conclusions This study presents DeepION, a data-driven ionospheric modeling framework based on the DeepONet architecture. The framework learns spatiotemporal relationships between GNSS-derived STEC observations and key ionospheric parameters. With a CNN-based branch network and a periodically encoded trunk network, DeepION models and predicts STEC and VTEC, and then derives ROTI from the predicted STEC series. Experiments using global GNSS data during the May 2024 geomagnetic storm show that DeepION captures storm-time ionospheric variability and achieves stable performance in STEC forecasting and global VTEC reconstruction. Compared with conventional empirical and physics-based models, DeepION improves modeling accuracy and spatial representation. SPP experiments further show that ionospheric corrections derived from DeepION reduce positioning errors at both mid- and high-latitude stations, especially in the vertical and geometric components under disturbed geomagnetic conditions. These results indicate the practical value of DeepION for GNSS ionospheric correction during space weather events. Overall, DeepION provides a scalable framework for data-driven ionospheric modeling. Future work will extend it to multi-GNSS constellations, longer prediction lead times, and additional ionospheric observations.
2026, 48(6): 2493-2503.
doi: 10.11999/JEIT251085
Abstract:
Objective Geoacoustic parameters in shallow water are critical for characterizing underwater acoustic propagation. Traditional inversion methods, however, are limited by high computational complexity, high cost, and strong dependence on the accuracy of environmental models. To address these issues, an efficient and robust inversion method is proposed to improve the reliability and stability of shallow-water geoacoustic parameter estimation while preserving computational efficiency. Methods This method is developed from the Stokes parameters of the vector acoustic field. Signals received by a single vector hydrophone are processed with a warping transform to separate and extract the normal modes propagating in a shallow-water waveguide. The extracted signals are then used to calculate the Stokes parameters, which are normalized and used as input features for the inversion model. An attention-enhanced multi-task U-Net is constructed with a shared encoder and multiple prediction branches to estimate key geoacoustic parameters, including compressional wave velocity, shear wave velocity, density, compressional wave attenuation, and shear wave attenuation. In addition, channel attention and spatial attention, together with a multi-task loss function with uncertainty weighting, are used to improve feature extraction and adaptively balance the different parameter inversion tasks. Results and Discussions The attention mechanism is shown to suppress fluctuations in model predictions and to improve the accuracy and stability of geoacoustic parameter inversion. When 200 test samples are evaluated, the mean absolute percentage errors of both compressional wave velocity and seabed density remain below 5% (Table 3 ). After the attention mechanism is introduced, the errors in compressional wave velocity and seabed density are further reduced to below 3% (Table 5 ), which indicates improved prediction accuracy for these key parameters. The proposed method is also shown to be insensitive to parameter mismatch and to have strong robustness to environmental variation. Furthermore, the method is validated with measured data from a shallow-water region in the northern South China Sea, and its effectiveness and reliability in practical applications are confirmed (Table 6 and Fig. 9 ). These results show that the attention-enhanced multi-task U-Net effectively captures critical features from the Stokes parameters and yields more stable and accurate geoacoustic parameter estimation in shallow-water environments. Conclusions The inversion method based on the Stokes parameters and an attention-enhanced multi-task U-Net effectively improves the accuracy and stability of shallow-water geoacoustic parameter estimation and shows strong performance in the prediction of compressional wave velocity, shear wave velocity, and density. However, limitations remain in the inversion of seabed attenuation. Future work should focus on improving feature extraction methods and network architecture and on testing the applicability of the method under more complex marine conditions.
2026, 48(6): 2504-2515.
doi: 10.11999/JEIT251195
Abstract:
Objective The electromagnetic environment is becoming increasingly complex, and mainlobe suppression jamming degrades the detection performance of phased array radars. Adaptive Interference Cancellation (AIC) can suppress such jamming. However, it may distort the mainlobe pattern and introduce azimuth angle-measurement errors. Most existing studies focus on interference cancellation mechanisms, whereas the angle-measurement errors caused by cancellation have received limited attention. Receive-channel amplitude and phase errors can further reduce angle-measurement accuracy. This paper investigates the effect of receive-channel amplitude and phase errors on the angle-measurement errors of monopulse phased array radar without a difference-difference channel. Methods A monopulse phased array radar without a difference-difference channel is analyzed. Receive-channel amplitude and phase errors are modeled by normal distributions. The mean represents the systematic offset, and the standard deviation represents random fluctuation. The operating principles of phased array radar receivers, monopulse radar systems, sum-difference angle measurement, and mainlobe suppression jamming cancellation are first described. Two angle-measurement models are then derived: an ideal reference model and an amplitude and phase error model. Under ideal interference-free and error-free conditions, the effective angle-measurement range of the radar is ±2.5°. The jamming source is set at –1.2°, and the corresponding angle-measurement results are used as the reference for subsequent experiments. Monte Carlo simulations, with 100 independent tests for each parameter set, are performed to analyze the statistical characteristics of the angle-measurement errors. Heatmaps are used to present the absolute errors and their variation trends. Results and Discussions (1) Without receive-channel amplitude and phase errors, the jamming angle remains fixed at –1.2°. Before interference cancellation, the target indication angle is consistent with the true value. After cancellation, the absolute error between the target indication angle and the true value near the beam normal is no more than 0.1°. However, a cancellation null near the jamming angle causes abrupt changes in the azimuth indication, and the error increases as the target moves away from the beam normal. (2) Before cancellation, the azimuth angle-measurement error increases with the absolute amplitude-error mean and the incident angle. The error reaches more than 0.06° when the amplitude-error mean is ±0.9 dB and the incident angle is ±2.5°. Within an incident-angle range of ±2°, the error is generally below 0.02°. When the amplitude-error mean is fixed, the error increases with the amplitude-error standard deviation. When the phase-error standard deviation is fixed, the error increases with the absolute phase-error mean. The error exceeds 0.15° at a phase-error mean of ±0.9° and reaches approximately 0.6° at a phase-error standard deviation of 6° and an incident angle of ±2.5°. (3) After cancellation, the effect of phase error is strongest at an incident angle of 0.5°, where the azimuth angle-measurement error reaches approximately 0.4°. Outside this region, the error is generally controlled within 0.2° and decreases rapidly as the target moves away from the beam normal. Conclusions This paper quantifies the effect of receive-channel amplitude and phase errors on azimuth angle-measurement errors before and after interference cancellation. The main conclusions are as follows. First, amplitude and phase errors both cause random fluctuations in azimuth angle measurement, and phase errors have a stronger effect than amplitude errors. Second, in the absence of jamming, azimuth angle-measurement errors are smallest near the beam normal and increase as the target approaches the boundary of the effective angle-measurement range. Third, under jamming and cancellation conditions, the azimuth angle-measurement error reaches its peak near the beam normal and then decreases rapidly. This study provides guidance for azimuth angle-measurement error assessment, error budgeting, and mainlobe suppression jamming cancellation in engineering applications. Future work will focus on non-normal amplitude and phase errors, calibration dynamics, multiple-jamming-source scenarios, and experimental validation.
2026, 48(6): 2516-2528.
doi: 10.11999/JEIT250727
Abstract:
Objective Sub-bottom profiling is widely used in seabed geological and resource exploration, pipeline route inspection, and port and channel safety assessment. It is also a major topic in underwater acoustic detection. Accurate sub-bottom horizon picking is critical for interpreting sedimentary structures, analyzing seabed substrate characteristics, and identifying buried objects. However, existing horizon-picking methods often cannot balance picking quality, false-alarm suppression, and online real-time performance. To address this problem, this paper proposes a real-time sub-bottom horizon-picking method based on Maximum Correlated Kurtosis Deconvolution (MCKD) combined with continuity constraint. Methods The proposed method includes three stages: preprocessing, coarse horizon extraction, and fine horizon extraction. In preprocessing, raw echoes are enhanced by cascaded band-pass filtering and matched filtering. Fixed-delay correction is then performed to align the picked positions with pulse leading-edge arrivals. In coarse horizon extraction, synthesized periodic signals are constructed under multiple slicing step lengths. MCKD is then applied to enhance impulsive horizon responses and obtain potential horizon sequences. These candidates are screened and fused using a cross-step-length consistency criterion to suppress false alarms. In fine horizon extraction, a continuity constraint is applied within an online sliding window. This process removes isolated points, segments horizons, and performs curve fitting and correction, thereby further reducing residual false alarms and improving horizon continuity. Results and Discussions Simulation and field data experiments were conducted to evaluate detection probability, false-alarm probability, horizon positioning error, processing time, and extracted horizon profiles. Monte Carlo results show that fine horizon extraction further reduces false alarms and positioning errors while maintaining detection performance close to that of coarse horizon extraction (Fig. 5 , Fig. 6 ). When the echo Signal-to-Noise Ratio (SNR) is higher than –15 dB, the detection probability exceeds 70.000%, and the false-alarm probability remains below 0.200%. When the SNR is higher than –10 dB, the detection probability exceeds 99.000%, the false-alarm probability is below 0.100%, and the positioning error is close to one sample point (Fig. 6 ). In the sub-bottom survey simulation, the proposed method successfully extracts both the seabed surface and the buried sedimentary horizon under different noise conditions. Its results are more refined than those obtained using the Fractional Fourier Transform (FrFT)-based algorithm and are generally comparable to manual interpretation (Fig. 7 , Fig. 8 ). Field data results further confirm the effectiveness of the method. In comparison with signal-based algorithms, the proposed method achieves an average detection probability of 91.833%, an average false-alarm probability of 0.004%, and an average positioning error of 10.15 samples. By contrast, the FrFT-based algorithm produces a much higher false-alarm probability of 3.987% (Table 1 ). In comparison with image-based algorithms, although their detection probabilities exceed 95%, their false-alarm probabilities and processing times are much higher than those of the proposed method (Table 2 ). Qualitative results also show that the extracted horizons agree well with manual interpretation trends. The proposed method yields lower background noise, avoids obvious large-scale false horizons, and preserves local fluctuations and possible interruptions in real horizons (Figs. 9 –12 ). Overall, the method provides a better balance for online horizon picking by combining acceptable detection probability and positioning accuracy with an extremely low false-alarm probability and real-time processing capability (Table 1 , Table 2 ). Conclusions This study presents a real-time sub-bottom horizon-picking method based on MCKD combined with continuity constraint. The method is organized into preprocessing, coarse horizon extraction, and fine horizon extraction. It effectively extracts the seabed surface and sedimentary horizons while meeting real-time processing requirements. Simulation results show that, when the SNR exceeds –10 dB, the method achieves a detection probability greater than 99.000%, a false-alarm probability below 0.100%, and a positioning error close to one sample point. Field data processing results show an average detection probability of 91.833%, an average false-alarm probability of 0.004%, and an average positioning error of 10.15 samples. These results confirm the effectiveness and practical value of the proposed method for real-time sub-bottom horizon picking. The method maintains high detection performance, suppresses false alarms, and achieves millisecond-level processing, making it suitable for online sub-bottom horizon-picking tasks in practical applications.
2026, 48(6): 2529-2540.
doi: 10.11999/JEIT251340
Abstract:
Objective Monaural speech separation aims to extract individual speaker signals from a single-channel mixture. It is a core technology for addressing the “cocktail party problem” and has substantial application value in low-resource, low-latency scenarios such as mobile voice assistants, teleconferencing, and hearing aids. However, the lack of spatial cues in single-channel signals, together with the substantial overlap of multiple speakers in both time-domain waveforms and frequency-domain spectra, makes accurate separation highly challenging, especially when the integrity and clarity of the target speech must be preserved. Current deep learning-based models often show limitations in three closely related aspects: effective coordination of multi-scale dependencies, efficient fusion of time-frequency information, and control of computational complexity. To address these challenges, a novel Multi-Scale Attention model integrating Time-Frequency domain information (MSA-TF) is proposed to improve separation performance, computational efficiency, and generalization capability. Methods The MSA-TF model contains three key components. First, a lightweight Time-Frequency fusion module is designed. The module first divides the frequency band into four subbands on the basis of speech priors, such as low-frequency energy concentration and high-frequency detail sensitivity, to extract spectral features efficiently. A dynamic gating mechanism with decomposed convolutions and SiLU activation is then applied to adaptively enhance speaker-discriminative features and suppress redundant channels associated with noise. Finally, a cross-attention mechanism is used to promote deep interaction between time-domain and frequency-domain features during the encoding stage. Global semantic information from the time domain guides the selection and weighting of useful frequency-domain features, allowing mutual correction and complementarity. This module adds only 0.8 M parameters. Second, a Multi-scale Interaction Separator is proposed to address the limitations of sequential or loosely coupled multi-scale processing in models such as SepFormer. Multi-granularity features, ranging from frame-level F 1 to syllable-level semantic F 4, are extracted through cascaded dilated convolutions. Its core is the “GF-LF Iterative Feedback” mechanism. The Global Flash module, based on efficient FLASH attention, captures long-range dependencies and syllable-level context. This global information is upsampled and injected into local features ( F k) through residual connections. Local Flash modules, also based on FLASH attention, then process the enhanced local features (\begin{document}$ {\boldsymbol{F}}_k^{\prime} $\end{document} Results and Discussions Extensive experiments are conducted on the standard WSJ0-2mix and Libri-2mix datasets, with Scale-Invariant Signal-to-Noise Ratio (SI-SNR) and Signal-to-Distortion Ratio (SDR) used as evaluation metrics. Ablation studies (Table 1 ) confirm the individual and joint contributions of the proposed modules. When only the time-frequency module is added to the TDAnet baseline, SI-SNR increases by 0.3 dB and SDR by 0.4 dB with only a small increase in parameters, confirming its contribution to signal structure modeling, particularly for high-frequency details. When only the multi-scale interaction module is incorporated, SI-SNR increases by 2.5 dB and SDR by 2.7 dB, highlighting its central role in modeling long-term dependencies. When the time-frequency and multi-scale modules are combined in the complete MSA-TF core, a synergistic effect is obtained, reaching 17.6 dB SI-SNR, which exceeds the sum of the individual gains. This result indicates that the dual-dimensional features provided by time-frequency fusion and the deep dependency modeling enabled by multi-scale interaction strengthen each other. Spectrogram analysis (Fig. 3 ) further shows that the time-frequency module effectively suppresses residual high-frequency noise and produces clearer spectral contours for the target speech. On the WSJ0-2mix test set (Table 2 ), MSA-TF achieves state-of-the-art performance, with 17.6 dB SI-SNR and 17.8 dB SDR. It matches the performance of SuperFormer and substantially outperforms strong baselines such as Conv-TasNet by 2.3 dB SI-SNR, while maintaining a reasonable parameter count of 15.6 M. For generalization evaluation on the completely unseen Libri-2mix dataset (Table 4 ), MSA-TF, trained only on WSJ0-2mix, achieves 14.2 dB SI-SNR and 14.7 dB SDR. Its performance is comparable to that of Conv-TasNet models trained specifically on Libri-2mix, which achieve 14.4 dB SI-SNR, and it outperforms BLSTM-TasNet trained on Libri-2mix. This strong cross-dataset adaptability indicates that the model captures universal time-frequency characteristics and multi-scale dependency structures in speech signals rather than overfitting to a specific dataset distribution. Conclusions An MSA-TF model is presented to address key challenges in monaural speech separation through deep integration of multi-scale time-frequency information interaction. The proposed lightweight Time-Frequency fusion module efficiently supplements time-domain features with discriminative frequency-domain information. The Multi-scale Interaction Separator, with its iterative feedback mechanism, enables dynamic bidirectional information flow across scales and substantially improves the joint modeling of short-term details and long-term dependencies. Combined with an efficient attention design, the model achieves superior performance without excessive computational cost. Experimental results show that MSA-TF achieves leading separation performance on standard benchmarks and shows strong generalization ability on unseen data distributions, confirming the effectiveness of this comprehensive design. The model provides an efficient, robust, and generalizable solution for practical low-resource application scenarios. Future work may examine advanced cross-modal fusion techniques and dynamic scale adjustment strategies to further improve robustness and performance in more complex and variable acoustic environments.
2026, 48(6): 2541-2550.
doi: 10.11999/JEIT251028
Abstract:
Objective This paper presents a comprehensive channel model for broadband Earth-Moon-Earth (EME) communication. The model addresses the limitations of traditional simplified models, which cannot accurately describe the Moon’s complex scattering behavior, terrain-induced scattering, or shadowing effects. Existing approaches usually treat the Moon as a point reflector or rely on empirical scattering laws. These methods are insufficient for broadband, high-capacity EME systems. To address these limitations, a unified large-scale fading model is proposed to statistically characterize lunar terrain-induced reflection and shadowing. A small-scale fading model is also developed to analyze multipath propagation and Doppler effects, decompose the channel, and quantify dynamic impairments. Link-level simulations are conducted to validate the model. This study helps close a gap in broadband EME channel modeling and provides a basis for designing and optimizing future deep-space communication systems. Methods A dual-scale modeling framework is proposed for broadband EME channels. At the large scale, a unified integral path-loss model is developed for both wide-beam and narrow-beam scenarios. Lunar terrain is statistically represented by a Gaussian height distribution to describe shadowing and roughness effects. A distributed integral formulation is used to calculate the effective Radar Cross Section (RCS) under narrow-beam conditions. At the small scale, the channel is decomposed into quasi-specular reflection and diffuse scattering components. The corresponding delay-power profiles are derived from surface roughness and lunar scattering mechanisms. Doppler shift and Doppler spread are analytically modeled according to Earth-Moon orbital motion. Monte Carlo simulations and numerical integration are used to verify the models. System-level performance is evaluated by the Bit Error Rate (BER) under different channel conditions, equalization schemes, and frequency-offset correction methods. Results and Discussions The proposed channel model captures both large-scale fading and small-scale fading in broadband EME communication. Simulation results show that the large-scale fading model represents the non-uniform power distribution across the lunar disk through an RCS-based integral model. In large-aperture antenna scenarios, the received-power correction approaches 20 dB compared with the traditional average model. At the small scale, quasi-specular reflection and diffuse scattering characterize the multipath delay spread caused by lunar surface reflection. The Doppler model quantifies the effects of Earth’s rotation and lunar orbital motion. At 1.296 GHz, the maximum two-way Doppler shift is approximately 4.5 kHz, and the two-way Doppler spread ranges from –39.88 Hz to 39.88 Hz. Simulations under low Signal-to-Noise Ratio (SNR) conditions show that conventional equalization schemes, including Least Mean Squares (LMS), Recursive Least Squares (RLS), and RAKE receiver with Maximum-Ratio Combining (RAKE-MRC), stagnate near BER = 0.1. Frequency-offset correction methods, including Fast Fourier Transform (FFT)-based estimation and Maximum Likelihood Estimation (MLE), also degrade under large frequency offsets. These results show that severe multipath delay spread is a major bottleneck in broadband EME communication. Conclusions This paper develops and validates a comprehensive channel model for broadband EME communication. Compared with conventional point-target and empirical methods, the proposed model describes path loss, shadowing, multipath delay spread, and Doppler effects more accurately. The results show that lunar terrain and surface scattering produce severe multipath delay spread, which greatly limits the performance of traditional equalization methods. By contrast, Doppler variation in the EME link is relatively slow, and Doppler spread is limited. Existing MLE-based frequency-offset correction can effectively compensate for this effect. Future work should explore adaptive receiver strategies, including machine learning and joint compensation, to address the severe delay spread caused by complex lunar reflection channels. This model provides a foundation for reliable EME links and future deep-space communication networks.
2026, 48(6): 2551-2559.
doi: 10.11999/JEIT250853
Abstract:
Objective As one of the coding schemes used in Fifth-generation (5G) wireless communication systems, Low-Density Parity-Check (LDPC) codes can achieve performance close to the Shannon limit through iterative decoding. However, in practical wireless transmission environments, the decoding performance of LDPC codes is vulnerable to burst interference in wireless channels. The Normalized Min-Sum (NMS) decoding algorithm is highly sensitive to the distribution characteristics of the input log-likelihood ratios (LLRs). Burst interference causes the LLRs to deviate from a Gaussian distribution, which degrades decoding performance. Meanwhile, 5G LDPC decoders are often configured with a fixed number of Processing Elements (PEs) based on the maximum lifting size to cover the full code-length range. In Ultra-reliable Low-Latency Communications (URLLC) short-code transmission scenarios, the lifting size is much smaller than the maximum lifting size. This condition leaves many PEs idle for long periods and results in low hardware resource utilization. To address these issues, a Clipped Normalized Min-Sum List (CNMSL) decoding algorithm is proposed. By co-designing burst-interference smoothing and idle-resource reuse, the proposed algorithm improves hardware resource utilization and decoding performance. Methods The statistical characteristics of LLRs over Additive White Gaussian Noise (AWGN) and interference channels are first analyzed, and the negative effect of burst interference on decoding performance is qualitatively attributed to the increased proportion of saturated LLRs induced by such interference. The correlation between the optimal clipping threshold and the channel noise variance, burst interference variance, and burst probability is then examined. It is shown that, under limited variations in channel parameters, the optimal clipping threshold converges to a finite interval, referred to as the optimal threshold interval. On this basis, the CNMSL decoding algorithm is proposed. A list-decoding architecture is constructed by reusing idle PEs in 5G LDPC decoders. In this architecture, each decoding path performs independent and synchronous decoding to generate candidate codewords, and the optimal decoding result is selected through a Cyclic Redundancy Check (CRC). Meanwhile, an independent clipper is configured for each path, with parameters determined from the optimal threshold interval, thereby effectively suppressing the adverse effects of burst interference. Results and Discussions Experimental results show that the layered NMS algorithm nearly fails over interference channels when no clipping mechanism is used. With a single clipping threshold, the algorithm operates normally, and its Block Error Rate (BLER) shows a convex downward trend, first decreasing and then increasing as the clipping threshold is reduced. Under various channel conditions for both short and long codes, the single-clipping layered NMS algorithm with a clipping threshold of 3.5 achieves a gain of about 1 dB at \begin{document}$ \mathrm{BLER}=10^{-2} $\end{document} Conclusions A CNMSL decoding algorithm is proposed to improve the error-correction performance of the conventional layered NMS decoding algorithm over interference channels. By reusing idle PEs for list decoding to generate multiple candidate paths, the proposed algorithm introduces no additional hardware overhead. In addition, an optimal threshold interval is defined to configure the clipper for each decoding path. This strategy limits the proportion of saturated LLRs and makes the input LLRs follow a Gaussian or near-Gaussian distribution. Experimental results show that, compared with the layered NMS decoding algorithm with a single clipper, the proposed CNMSL algorithm achieves a gain of approximately 0.5 dB for both short and long codes. Meanwhile, PE utilization is increased by an average of 69%.
2026, 48(6): 2560-2570.
doi: 10.11999/JEIT251377
Abstract:
Objective As wireless communication evolves toward high efficiency, low latency, and ubiquitous connectivity, higher requirements are placed on Automatic Modulation Recognition (AMR) to ensure link reliability in complex electromagnetic environments. Deep learning has improved recognition performance compared with traditional methods, which often rely on subjective feature design and have limited robustness. However, existing YOLO-based AMR models are not fully optimized for specific signal characteristics or practical deployment. These models often have excessive parameters and high computational complexity, which makes them unsuitable for resource-constrained hardware, such as edge nodes and Field-Programmable Gate Arrays (FPGAs), and limits their ability to meet real-time communication requirements. To address these bottlenecks, this paper proposes Precision and Lightweight Structure-YOLO (PLS-YOLO), a lightweight AMR model based on YOLOv10n. By optimizing network channels, replacing core modules, and improving the downsampling mechanism, the proposed model enables efficient integration of modulation signal classification and localization. It also reduces the parameter count and computational complexity, thereby supporting AMR deployment in resource-constrained scenarios. Methods The method includes two main stages: dataset preprocessing and PLS-YOLO model construction. In the preprocessing stage, the public RadioML2016.10a and RadioML2016.10b benchmark datasets for signal modulation recognition are used. For In-phase and Quadrature (IQ) signals in these datasets, the Short-Time Fourier Transform (STFT) is used to map one-dimensional temporal signals into two-dimensional time-frequency spectrograms containing phase and amplitude information. This process provides richer feature representations for the model. A random sampling strategy without replacement is then used to stitch individual time-frequency samples into 3×3 composite images (Fig. 4 ). Target labels matching the input format of YOLO-series models are generated at the same time. The dataset is divided into training, validation, and test sets at a ratio of 7:1.5:1.5 by stratified sampling to ensure consistent signal-type distributions across all subsets. The model is built on YOLOv10n, with targeted improvements designed to balance the parameter count and recognition performance. The C2f module in the original backbone network is replaced with the CSPPC module, which is based on the CSP architecture and consists of feature splitting, Partial Convolution (PConv) processing, and feature fusion. This design reduces parameters while improving recognition performance. The feature dimensionality reduction process in the backbone network is also reconstructed to reduce the increase in computational complexity caused by parameter redundancy. The traditional downsampling module is replaced with CGBlock, which improves the capture of complex modulation signal features by fusing context-aware information. Finally, standard convolutions in the PSA and v10Detect modules are replaced with PConv to further reduce computational complexity and jointly optimize lightweight design and recognition performance. Results and Discussions Experimental results on RadioML2016.10a show that PLS-YOLO achieves a mean Average Precision (mAP) of 68.4% within the Signal-to-Noise Ratio (SNR) range of –20 to 18 dB. The mAP increases to 94.3% when SNR ≥ 0 dB. Compared with the baseline YOLOv10n model, PLS-YOLO improves mAP by 0.6%, reduces the parameter count by 47.33%, and decreases computational complexity by 34.15%. Its inference speed also increases by 5 frames per second (fps) (Table 2 ). These results show that the model effectively balances recognition performance and lightweight deployment by reducing computational cost while improving precision. To verify robustness, additional experiments are conducted on RadioML2016.10b. As shown in Table 4, PLS-YOLO achieves an mAP of 73.30% over the –20 to 18 dB range and 95.4% at SNR ≥ 0 dB. It outperforms mainstream models such as MCNet and LSTM2, confirming its strong recognition performance. Furthermore, Fig. 5 shows that converting IQ data into spectrograms is more suitable for PLS-YOLO recognition of digital modulation signals. By contrast, the recognition performance for analog modulation signals remains limited. Future work should therefore improve feature modeling and recognition capability for analog signals. Conclusions This study proposes PLS-YOLO, a lightweight AMR model based on YOLOv10n. To jointly improve modulation recognition performance and model compactness, the network structure is optimized through channel dimensionality reduction, core module replacement, downsampling mechanism improvement, and PConv substitution. These strategies reduce key limitations of existing YOLO-based AMR models, including parameter redundancy, high computational complexity, and limited adaptability to resource-constrained scenarios such as edge nodes and FPGAs. Experiments on the RadioML2016.10a and RadioML2016.10b benchmark datasets show that PLS-YOLO achieves strong overall performance. While integrated signal classification and localization are maintained, both parameter count and computational complexity are substantially reduced compared with the baseline YOLOv10n model, with a clear improvement in recognition performance. The results verify the effectiveness and feasibility of the proposed optimization strategies and provide a practical technical path for AMR implementation. The remaining limitations in analog modulation signal recognition also indicate a clear direction for future research.
2026, 48(6): 2571-2580.
doi: 10.11999/JEIT251251
Abstract:
Objective Entanglement-Assisted Quantum Error-Correcting Codes (EAQECCs) provide an effective way to protect quantum information by using pre-shared entanglement between the sender and receiver. Existing constructions of EAQECCs mainly rely on classical cyclic or constacyclic codes and often require strong algebraic constraints, which limit the range of achievable parameters. This paper develops a general and systematic framework for constructing new families of EAQECCs from Twisted Reed-Solomon (TRS) codes over finite fields. The study has two aims. The first is to extend classical Reed-Solomon-based code design to the twisted setting so that richer algebraic structures can be used. The second is to determine the exact number of maximally entangled pairs required to attain the quantum Singleton bound. The final objective is to construct Maximum-Distance Separable (MDS) EAQECCs with greater flexibility and broader parameter ranges than existing methods. Methods The proposed method starts from the definition of TRS codes over finite fields. A twist parameter is introduced into the generator matrix, which changes the structure of the corresponding parity-check matrices. By systematically analyzing the associated coset-sum matrices in the twisted and untwisted cases, the rank of the relevant matrix product is determined. This rank equals the number of required entangled pairs and therefore provides the theoretical basis for the construction of EAQECCs. A detailed algebraic analysis shows that the matrix contains a submatrix with entries \begin{document}$ {M}_{l,j}=\displaystyle\sum\nolimits_{y\in W}{\left({\xi }^{j}y\right)}^{tl} $\end{document} \begin{document}$ t\zeta^{jl} $\end{document} Results and Discussions Comprehensive parameter analysis and explicit examples verify the theoretical results. Comparative analysis further shows the flexibility of the proposed framework. Unlike previous constructions that require divisibility conditions such as \begin{document}$ a\mid (q+1) $\end{document} \begin{document}$ a\mid (q-1) $\end{document} Table 4 ) reveals several new parameter regimes that are not accessible with classical or cyclic-code-based constructions. In particular, the proposed method yields larger code lengths and more flexible entanglement consumption rates \begin{document}$ {c}/{n} $\end{document} Conclusions This study establishes an algebraic framework for constructing MDS EAQECCs from TRS codes. By rigorously analyzing the rank properties of coset-sum matrices, the required entanglement is determined precisely, and the conditions under which the constructed codes attain the quantum Singleton bound are identified. Two broad classes of MDS EAQECCs are obtained, corresponding to \begin{document}$ a\mid \left(q+1\right) $\end{document} \begin{document}$ a\mid \left(q-1\right) $\end{document}
2026, 48(6): 2581-2590.
doi: 10.11999/JEIT251219
Abstract:
Objective Vehicular applications have diverse Quality of Service (QoS) requirements that traditional spectrum-focused networks cannot adequately meet. Although network slicing based on Mobile Edge Computing (MEC) provides customized service provisioning, existing methods often fail to jointly consider slice generation and adaptive access control. To address these limitations, this paper proposes a two-stage vehicular network slicing framework that integrates resource-aware slice generation with dynamic pricing and access control. The framework supports efficient resource allocation and slice access management. It also improves service quality, resource utilization, and system adaptability for both the MEC Network Service Provider (MEC-NSP) and vehicles through a Stackelberg game-based interaction mechanism. Methods The proposed solution uses a two-layer coupled mechanism consisting of resource pre-allocation and Stackelberg game-based pricing and access control. In the first stage, a three-dimensional resource pre-allocation mechanism jointly optimizes communication, computation, and caching resources to satisfy vehicular latency and bandwidth requirements. The resource allocation problem is formulated as a Mixed-Integer Nonlinear Programming (MINLP) problem. It is then decoupled into uplink and downlink subproblems, which are solved using branch-and-bound and interior-point methods, respectively. In the second stage, a Stackelberg game is developed to balance MEC-NSP profit and vehicle QoS. The MEC-NSP acts as the leader and sets dynamic slice prices. The network controller acts as the follower and determines the optimal slice selection probabilities. This interaction is solved using the Iterative Slices Pricing Algorithm (ISPA), which is proven to converge to a Nash equilibrium. Results and Discussions Simulation results show that the proposed framework consistently outperforms baseline algorithms, including Fixed Slice Pricing, Average Resource Allocation, Random Selection, and Dynamic Combinatorial Double Auction (DCDA), under different network conditions. In bandwidth-constrained scenarios, the proposed framework increases MEC-NSP profit by up to 20.77% compared with the Random Selection approach. When resources are abundant, with 150% capacity, it maintains profit gains of 3～9% over other baselines. The ISPA converges to equilibrium after approximately 175 iterations. The flexible pricing mechanism balances network loads, improves cache hit rates, and reduces resource bottlenecks, thereby supporting high QoS satisfaction. Conclusions The proposed dual-layer framework integrates slice generation and pricing for resource-aware network slicing in vehicular MEC environments. By coupling three-dimensional resource pre-allocation with a Stackelberg game-based pricing strategy, the framework improves MEC-NSP profit, resource utilization, and vehicle QoS. Future work will study blockchain-based mechanisms for trusted negotiation and decentralized resource orchestration in cross-domain cooperation under multi-operator and multi-vendor environments.
2026, 48(6): 2591-2601.
doi: 10.11999/JEIT250749
Abstract:
Objective Connected autonomous vehicles (CAVs) need to perceive complex driving environments accurately and in real time, including traffic object detection, drivable-area segmentation, and lane-line detection. Existing single-task perception models are difficult to meet these diverse requirements simultaneously, while running multiple models in sequence increases latency and is unsuitable for safety-critical autonomous driving. In addition, the performance of deep learning models depends heavily on large-scale and diverse training data, whereas the data collected by an individual CAV are usually limited and scenario-specific. Federated learning (FL) provides a feasible way to train a more generalized global model without uploading raw vehicle data. However, conventional FL still faces several challenges in Internet of Vehicles scenarios, including malicious CAVs that upload low-quality or poisoned local models, privacy leakage from uploaded model parameters, and the single point of failure caused by centralized aggregation. To address these problems, this paper proposes a Secure Multi-task Federated Panoptic Perception (SMFPP) algorithm for connected autonomous vehicles. The goal is to achieve efficient multi-task driving perception, reliable collaborative model training, malicious-node resistance, privacy-preserving aggregation, and fault tolerance in a unified framework. Methods The proposed SMFPP framework consists of a cloud layer, an edge layer, and a terminal layer. The cloud layer publishes FL tasks and provides global coordination, the edge layer contains multiple road side units (RSUs) responsible for model aggregation, and the terminal layer consists of CAVs that train local models with their own driving-scene data. First, a YOLO-based multi-task panoptic perception model is designed as the local model. Its backbone network extracts visual features from vehicle images, while the spatial pyramid pooling fast module and feature pyramid network perform multi-scale feature fusion. Three task-specific detection heads are constructed to perform traffic object detection, drivable-area segmentation, and lane-line detection in parallel, thereby reducing perception latency and improving real-time applicability. The overall loss function combines the losses of the three tasks, including object detection losses and segmentation-related losses. Second, to defend against poisoning attacks and low-quality updates in FL, a hybrid-scoring-based CAV selection algorithm is proposed. For direct scoring, each CAV evaluates the cosine similarity between its local model and the previous global model, as well as the update age of its local model. For indirect scoring, RSUs select high-reputation CAVs to conduct encrypted cross-validation of other CAVs’ local models, and the model quality is evaluated according to local validation performance. The direct and indirect scores are normalized and used as the input of K-means clustering to select CAVs with high model quality, good timeliness, and strong consistency. The reputation value of each CAV is then updated for subsequent training rounds. Third, to avoid centralized aggregation risks, a Shamir-secret-sharing-based global model aggregation algorithm is developed. Each selected CAV splits its local model parameters into secret shares and sends them to different RSUs. Each RSU aggregates only the received shares and returns an aggregation share to CAVs. When a CAV receives at least the threshold number of aggregation shares, it reconstructs the global model by Lagrange interpolation. In this way, no single RSU can obtain a complete local model, and the system can still work when some RSUs fail. Results and Discussions Experiments are conducted on the BDD100K dataset, which contains diverse driving scenes from urban roads, suburban areas, and highways. In the basic simulation setting, five CAVs participate in FL training, each CAV uses 20% of the dataset to train its local model, and five RSUs serve as edge aggregation nodes. The effectiveness of the hybrid-scoring-based CAV selection algorithm is first verified under different numbers of malicious CAVs. As shown in Fig. 2 , the mAP50 values of normal CAVs increase with training rounds and gradually converge, whereas malicious CAVs maintain low and unstable mAP50 values. Fig.3 further shows that a malicious CAV is excluded from global aggregation throughout training, and the global model achieves better convergence when the proposed selection algorithm is used. Compared with FedAvg, FedCS, and FedAsync, the proposed SMFPP framework obtains the best overall performance, achieving 86.4% Recall, 77.7% mAP50, 89.6% mIoU, 82.4% Accuracy, and 27.2% IoU, as reported in Table 1 . For traffic object detection, SMFPP achieves 86.4% Recall and 77.7% mAP50, and its mAP50 is higher than those of MultiNet, DLT-Net, Fast R-CNN, YOLOv8n, YOLOP, A-YOLOM, and YOLOMH, as shown in Table 2 . For drivable-area segmentation, SMFPP reaches 89.6% mIoU, outperforming MultiNet, DLT-Net, PSPNet, YOLOv8n(seg), YOLOP, A-YOLOM, and YOLOMH by 20.2%, 18.8%, 2.0%, 12.2%, 0.3%, 0.9%, and 0.4%, respectively, as shown in Table 3 . For lane-line detection, SMFPP achieves the best Accuracy of 82.4% and IoU of 27.2%, as shown in Table 4 . In both daytime and nighttime scenarios, SMFPP performs better than YOLOP on all evaluation metrics, demonstrating its adaptability to different illumination conditions (Table 5 ). In terms of efficiency, SMFPP contains only 3.45 million parameters and reaches 52.4 fps, indicating that it is lightweight and suitable for edge deployment (Table 6 ). The scalability experiment shows that, when the number of CAVs increases from 5 to 20, the global model performance improves consistently, with Recall, mAP50, mIoU, Accuracy, and IoU increasing to 88.6%, 79.5%, 90.4%, 85.2%, and 28.4%, respectively (Table 7 ). Moreover, the Shamir-secret-sharing-based aggregation scheme remains effective as long as at least three of the five RSUs survive. When the number of available RSUs is below the threshold, the global model cannot be correctly recovered; once enough RSUs recover, model training continues normally (Fig. 5 ). These results demonstrate that SMFPP improves perception accuracy, robustness against malicious CAVs, privacy protection, fault tolerance, and scalability. Conclusions This paper presents SMFPP, a secure multi-task federated panoptic perception algorithm for connected autonomous vehicles. The proposed method integrates a lightweight YOLO-based multi-task perception model with FL to solve the data-island problem among CAVs and to support collaborative training without exposing raw driving data. The hybrid-scoring-based CAV selection algorithm effectively filters malicious or low-quality local models by jointly considering model quality, update timeliness, and model similarity. The Shamir-secret-sharing-based aggregation scheme further protects local model parameters and removes the single point of failure in centralized FL aggregation. Simulation results on the BDD100K dataset verify that SMFPP achieves competitive or superior performance in object detection, drivable-area segmentation, and lane-line detection, while maintaining high inference efficiency and strong robustness. Future work may consider larger-scale heterogeneous vehicular networks, dynamic communication conditions, adaptive threshold settings, and real-world deployment in edge-assisted intelligent transportation systems.
2026, 48(6): 2602-2610.
doi: 10.11999/JEIT251269
Abstract:
Objective The physical-layer security game is used to reveal the interaction between User Equipment (UE) and attackers, and to provide performance bounds for anti-jamming transmission and physical-layer authentication schemes based on the equilibria. However, existing game models overlook intelligent attackers that transmit jamming or spoofing signals, do not account for maritime wireless channels affected by evaporation ducts and sea wave fluctuations, and do not readily support performance evaluation of Large Language Model (LLM)-based inference tasks such as vessel traffic monitoring. Methods An anti-jamming maritime communication game for LLM inference is formulated. In this game, the jammer first selects the jamming power and channel to reduce the signal-to-interference-plus-noise ratio at the server at lower jamming cost. The UEs then select the transmit power, channel, LLM sparsity ratio, and control center to send sensing data, such as images, temperature, and humidity, so that inference accuracy is improved with lower latency. A physical-layer authentication game for maritime wireless networks with LLM inference is further formulated. The spoofing attacker first selects the number of spoofing packets to reduce authentication accuracy at lower cost. The control center then selects either the fast authentication mode based on channel state or the safe authentication mode based on the received signal strength and packet arrival interval from multiple ambient transmitters, as well as the test threshold, to improve accuracy at lower cost. Results and Discussions Based on the Stackelberg Equilibrium (SE) under an LLM with 7 billion parameters, the performance bounds of the Reinforcement Learning (RL)-based anti-jamming inference scheme are derived to show the effects of evaporation duct height, sea wave height, maximum LLM sparsity ratio, and quantization level on inference accuracy and latency. In addition, the performance bounds of the RL-based maritime spoofing detection scheme are derived from the SE of the physical-layer authentication game to show the effect of the maximum number of spoofing packets on authentication accuracy. Simulations are conducted for five UEs with antenna heights of 3 m, which offload images, temperature, and humidity data using transmit power of up to 200 mW at 5.8 GHz with a bandwidth of 20 MHz, to five control centers with antenna heights of 6 m. The jammer uses a Deep Q-Network to select the jamming power, with a maximum transmit power of 200 mW for each 5.8 GHz channel. The spoofing attacker uses a Deep Q-Network to select the number of spoofing packets, up to 100. The results show that the inference accuracy and latency of the RL-based anti-jamming maritime communication scheme for LLM inference converge to the performance bounds, with gaps of less than 0.6%, after 2,500 time slots. In addition, the RL-based authentication scheme converges after 1,000 time slots, with a gap of less than 1.6%. Conclusions In this paper, a maritime physical-layer security game for LLM inference is formulated to address scenarios including anti-jamming sensing data transmission and spoofing detection. The aim is to investigate how UEs determine the transmit power and channel, and how the control center selects authentication modes and test thresholds to improve physical-layer security. The attacker selects attack modes and parameters to reduce inference accuracy, increase latency, and even cause denial of service. Based on the SE and the related conditions, the performance bounds show that inference accuracy increases with the maximum transmit power and decreases linearly with the sparsity ratio. Furthermore, the effect of the maximum number of spoofing packets on inference accuracy is analyzed. Simulation results show that the RL-based maritime physical-layer security schemes converge to the performance bounds, which validates the accuracy and effectiveness of the game model.
2026, 48(6): 2611-2621.
doi: 10.11999/JEIT251145
Abstract:
Objective The proliferation of Deep Neural Networks (DNNs) in safety-critical domains such as autonomous driving and biomedical diagnostics has raised serious concerns about their vulnerability to adversarial threats, particularly backdoor attacks. In these attacks, hidden triggers are embedded during training, causing models to behave normally on clean inputs while producing malicious outputs when specific triggers are present. Existing backdoor methods mainly operate in either the spatial domain or the frequency domain, but they face a fundamental tradeoff between Attack Success Rate (ASR) and stealth. Spatial triggers often introduce visible artifacts, whereas frequency-domain amplitude perturbations disrupt spectral energy distributions and can therefore be detected by advanced defenses such as spectral anomaly detection. This study addresses the need for a backdoor paradigm that simultaneously achieves high attack performance, minimal perceptual distortion, and robustness against state-of-the-art defense methods. The objective is to develop a frequency-domain backdoor attack based on phase manipulation, which is better aligned with human visual perception and structural consistency, thereby overcoming the limitations of existing methods. Methods FDPS integrates frequency-domain phase manipulation, perceptual similarity screening, and standard data poisoning. The method first converts input images from RGB to Y'CbCr color space. This conversion isolates the chrominance channels while preserving the luminance component. Discrete Fourier Transform (DFT) is then applied to the chrominance components to obtain complex frequency spectra. Phase information is computed with the atan2 function, and selected high-frequency components are shifted to embed the trigger. Image reconstruction is performed through Inverse Discrete Fourier Transform (IDFT). The framework further incorporates Learned Perceptual Image Patch Similarity (LPIPS) filtering. This filter removes generated samples that do not satisfy the similarity threshold. The screening process ensures that all retained triggers remain visually imperceptible. The accepted poisoned samples are assigned the target class labels and then combined with the clean training data according to standard protocols. Results and Discussions FDPS achieves near-perfect ASR, reaching 99%, while maintaining Benign Accuracy (BA) across three datasets and two network architectures (Table 1 ). The method embeds triggers by manipulating phase information in the Cb and Cr chrominance channels through Fourier transforms, and LPIPS filtering helps preserve visual stealth. Experimental results show that poisoned images retain semantic focus, as confirmed by Grad-CAM visualizations that remain aligned with the clean-image patterns (Fig. 4 ). The method also shows strong resistance to defense mechanisms. Under Neural Cleanse, FDPS yields an anomaly index of 1.73, which is below the detection threshold of 2 (Figs. 3 ～5 ). Under STRIP, the entropy distribution of poisoned samples substantially overlaps with that of clean samples. Additional analysis shows that high-frequency phase perturbation achieves strong attack performance with limited poisoning. In particular, on the GTSRB dataset, FDPS achieves 99% ASR with only 2% poisoned training samples, while minimizing the effect on model utility (Fig. 8 ; Table 3 ). Conclusions An end-to-end frequency-domain strategy is proposed to embed covert triggers into image classification models while preserving fidelity on clean samples. By shifting selected high-frequency phase components in the chrominance channels and applying LPIPS-based filtering, FDPS achieves 99% ASR with negligible BA loss and minimal visible artifacts. It also evades representative detection methods, including Grad-CAM, Neural Cleanse, Adversarial Neuron Pruning (ANP), and STRIP. These findings indicate that high-frequency phase perturbation constitutes an effective and stealthy backdoor mechanism. Future work should extend this strategy to broader modalities and develop dedicated frequency-domain anomaly detectors as principled countermeasures.
2026, 48(6): 2622-2630.
doi: 10.11999/JEIT251360
Abstract:
Objective Sequences with good correlation properties are widely used in wireless communications, cryptography, and radar systems. However, a sequence set cannot simultaneously achieve ideal autocorrelation and ideal cross-correlation. This limitation has led to the study of two signal classes with ideal correlation properties: Zero Correlation Zone(ZCZ) sequences and Golay Complementary Sets(GCS). A Golay-ZCZ sequence set combines the advantages of both. Its constituent sequences exhibit ideal periodic autocorrelation and cross-correlation within the ZCZ, and the sums of their aperiodic autocorrelations are zero at all nonzero shifts. Therefore, a Golay-ZCZ set is both a ZCZ set and a GCS. It can thus be used in the applications of both sequence classes. An array set is a two-dimensional extension of a sequence set. Although Golay-ZCZ sequence sets have been widely studied and constructed, research on Two-Dimensional (2D) Golay-ZCZ array sets remains limited. This study proposes three constructions of 2D Golay-ZCZ array sets based on 2D multivariable functions and the concatenation operator. These array sets can be used as precoding matrices for massive Multiple Input Multiple Output(MIMO) omnidirectional transmission. Methods Three construction methods for 2D Golay-ZCZ array sets are proposed, including one direct construction and two indirect constructions. The resulting parameters have not been reported in existing studies. In the first construction, a 2D Golay-ZCZ array set is generated using 2D multivariable functions, with parameters expressed as prime powers. This direct function-based approach enables efficient synthesis of the target arrays. The second and third constructions generate 2D Golay-ZCZ array sets through horizontal and vertical concatenation of Two-Dimensional Complete Complementary Codes(2D CCC), respectively. In these indirect constructions, the parameters are not restricted to prime powers. This property broadens the applicability of the methods and increases parameter flexibility. Results and Discussions The first construction generates a 2D Golay-ZCZ array set with array size \begin{document}$ p_{1}^{{m}_{1}}\times p_{2}^{{m}_{2}} $\end{document} \begin{document}$ ({p}_{1}-1)p_{1}^{{\pi }_{1}(2)-1}\times ({p}_{2}-1)p_{2}^{{\sigma }_{1}(2)-1} $\end{document} \begin{document}$ {p}_{1} $\end{document} \begin{document}$ {p}_{2} $\end{document} Fig. 1 ). The second construction generates a ZCZ array set with array size \begin{document}$ {L}_{1}\times {N}^{2}{L}_{2} $\end{document} \begin{document}$ ({L}_{1}-1)\times (N-1){L}_{2} $\end{document} \begin{document}$ (N,N,{L}_{1},{L}_{2}) $\end{document} \begin{document}$ {N}^{2}{L}_{1}\times {L}_{2} $\end{document} \begin{document}$ (N-1){L}_{1}\times ({L}_{2}-1) $\end{document} \begin{document}$ (N,N,{L}_{1},{L}_{2}) $\end{document} Table 1 ), the construction methods proposed in this paper generate parameter sets that have not been reported in the existing literature. The constructed array sets provide considerable flexibility in array dimensions and ZCZ sizes. This flexibility is valuable for the design of precoding matrices in MIMO omnidirectional transmission systems. In practical implementations, the dimension of a precoding matrix is typically determined by the number of transmit antennas, whereas the ZCZ size must match the maximum multipath delay spread of the channel. Owing to this parameter flexibility, the proposed 2D Golay-ZCZ array sets support adaptive selection under different antenna configurations and channel conditions. Conclusions Three construction methods for 2D Golay-ZCZ array sets are proposed. These methods generate array sets with flexible array sizes and large ZCZ widths. The first construction is based on a 2D multivariable function and can include previous results as special cases without using kernels. The second and third constructions rely on the concatenation operator and provide greater parameter flexibility. The proposed 2D Golay-ZCZ arrays have potential applications in MIMO omnidirectional transmission. The parameter-flexible array sets can be selected according to different antenna configurations and channel conditions. This property suppresses multi-antenna interference within the zero-correlation zone and maintains uniform transmitted energy.
2026, 48(6): 2631-2640.
doi: 10.11999/JEIT250850
Abstract:
Objective With the increasing penetration of distributed renewable energy and the growing dependence on cyber-physical infrastructure, Networked MicroGrids (NMGs) are increasingly vulnerable to False Data Injection Attacks (FDIAs). These attacks threaten frequency stability and system security. Traditional secondary control methods are limited by constrained communication resources and fixed sampling mechanisms. They often fail to maintain resilient operation under stealthy FDIAs and dynamic disturbances. To address these challenges, this study develops a multi-layer resilient control strategy that integrates event-triggered communication/control, data-driven attack observation, and double-replay Q-learning. The objective is to improve communication efficiency, attack detection, and stability recovery in NMGs under complex cyber threats. Methods The proposed Event-Triggered Control-Radial Basis Function-Double-Replay Q-Learning (ETC-RBF-DRQL) framework integrates an Event-Triggered Control (ETC) mechanism, a Radial Basis Function Unknown Input Observer (RBF-UIO), and a Double-Replay Q-Learning (DRQL) compensator to achieve resilient frequency control in NMGs under FDIAs. The ETC mechanism reduces redundant data transmission while maintaining system stability. The RBF-UIO estimates system states and detects anomalous deviations. After an attack is detected, the DRQL module adaptively generates compensation signals to suppress the attack effect and restore system stability. The framework is formulated using a modular dynamic model of NMGs, which supports stability analysis under communication and attack constraints. Simulation experiments are conducted on a 4-node distributed microgrid testbed in MATLAB/Simulink. The testbed includes different renewable energy sources and realistic communication links to verify the effectiveness and scalability of the proposed approach. Results and Discussions The proposed ETC-RBF-DRQL framework is validated on a 4-node NMG under FDIA scenarios. Simulation results show that the method achieves better overall performance in frequency regulation, communication efficiency, and attack resilience. Specifically, the frequency deviation peak is reduced from 0.021 8 Hz under periodic Proportional-Integral (PI) control to 0.012 1 Hz. The steady-state average deviation and fluctuation standard deviation are reduced to 0.009 7 Hz and 0.007 4 Hz, respectively (Fig. 4 , Table 2 ). The average communication event rate decreases to 11.9 pkt/s, corresponding to a 76.2% reduction compared with periodic sampling (Table 2 ). The proposed framework also maintains reliable attack detection performance, with a detection rate of 91.5%, a false alarm rate of 4.8%, and an area under the curve (AUC) of 0.968 (Table 2 ). These results indicate that the proposed method can coordinate frequency recovery, communication overhead reduction, and FDIA mitigation in NMGs. Conclusions This paper investigates a multi-layer resilient control framework for NMGs under FDIAs and communication constraints. The proposed ETC-RBF-DRQL method integrates event-triggered communication/control, RBF-UIO-based attack detection, and DRQL-based adaptive compensation. It therefore enables closed-loop coordination among anomaly detection, attack suppression, and frequency stability recovery. Simulation results on a 4-node NMG show that, compared with conventional PI-based schemes, the proposed approach reduces frequency deviation peaks and shortens recovery time while lowering communication overhead. Theoretical analysis further confirms its feasibility and stability under bounded estimation errors. This study focuses on sensor-side FDIAs and simplified communication conditions. Future work will consider more complex multi-type attacks and hardware-in-the-loop validation to support engineering applications.
Power Side-channel Leakage Assessment and Chosen-ciphertext Attack on the Decoding Function of Kyber
2026, 48(6): 2641-2654.
doi: 10.11999/JEIT251243
Abstract:
Objective The standardization of Post-Quantum Cryptography (PQC) has made the implementation security of Kyber a practical concern. Kyber, standardized as Module-Lattice-based Key-Encapsulation Mechanism (ML-KEM), is a lattice-based scheme with favorable efficiency and security based on the hardness of the Module Learning With Errors (MLWE) problem. However, its deployment on embedded devices can still produce measurable physical leakage. Existing studies have shown that side-channel attacks can target several Kyber modules, but two issues remain insufficiently studied. First, the leakage strengths of different auxiliary functions on the decapsulation and re-encryption path have not been compared under a unified assessment framework. This limits the identification of the most vulnerable implementation-level weak point. Second, although chosen-ciphertext attacks and power analysis have been studied, the decoding function poly_frommsg() has not been fully examined from the perspective of periodic leakage modeling and low-query key recovery. To address these issues, this work evaluates function-level leakage in the key operations of Kyber decapsulation and develops a chosen-ciphertext Simple Power Analysis (SPA) attack against the most vulnerable decoding function. The study provides a practical attack method and implementation-oriented security insights for protecting post-quantum cryptographic software on embedded platforms. Methods A function-oriented evaluation-and-attack framework is established for the execution path of Kyber.CCAKEM.Dec(). Four representative target functions are selected: the Barrett reduction function poly_reduce(), the encoding function poly_tomsg(), the decoding function poly_frommsg(), and the hash function G(). For each function, the intermediate variable with the largest data-dependent bit transition under crafted ciphertext inputs is first analyzed from the perspective of Hamming-distance leakage. Two ciphertext sets are then constructed so that the selected intermediate variable takes two maximally distinguishable values. For each set, 50 power traces are collected. The experiments are performed on an STM32F407IG embedded platform, and power signals are captured using a PicoScope 6406E oscilloscope at a sampling rate of 5 GS/s. Welch’s t-test-based Test Vector Leakage Assessment (TVLA) is used to quantify leakage significance, with ±4.5 used as the decision threshold for leakage detection. After poly_frommsg() is identified as the most vulnerable point, a chosen-ciphertext SPA attack is designed. The attack first constructs ciphertexts according to the coefficient range of the secret polynomial. It then extracts 256 Points of Interest (PoIs) from reference traces through local-maximum search. Finally, a grouped threshold model is built according to the periodic energy structure of the PoIs. The recovered message bits are mapped back to the coefficients of the secret polynomial, enabling full private-key reconstruction for Kyber512 and Kyber768. Results and Discussions The leakage assessment shows clear differences among the four target functions. For poly_reduce(), the intermediate variable t directly depends on the coefficients of the intermediate polynomial mp, and the maximum Hamming distance reaches 13 bit. The measured TVLA peaks are therefore concentrated around 50 for both Kyber512 and Kyber768 (Fig. 5 ). For poly_tomsg(), the relevant binary transition corresponds to a Hamming distance of only 1 bit, and the observed TVLA values are much smaller, at approximately 6 (Fig. 6 ). For poly_frommsg(), the message-dependent mask flips between 0 and 0xffff, producing a Hamming distance of 16 bit and the strongest leakage among all tested functions. The TVLA peaks reach about 60, identifying this module as the primary attack target (Fig. 7 ). For the hash function G(), the leakage is weaker and less regular, but several sampling points still exceed the TVLA threshold. This result indicates that theoretical indistinguishability under chosen-ciphertext attack (IND-CCA) reinforcement through the Fujisaki-Okamoto (FO) transform does not automatically remove physical leakage (Fig. 8 ). These results show that implementation-level vulnerability is strongly associated with data-dependent bit transitions. They also show that linear message-expansion functions may expose more stable power signatures than some arithmetic modules. Based on this observation, the proposed attack focuses on poly_frommsg(). Local-extrema analysis shows that the 256 message-bit operations generate 256 stable PoIs. Their energy values show a periodic pattern with an approximate period length of 8 (Fig. 10 , Fig. 11 ). Instead of applying a single global threshold to all PoIs, the proposed grouped threshold model divides the PoIs according to their positions within the period and computes location-aware thresholds. This design suppresses position-dependent drift and improves the consistency of bit decisions. The resulting message-recovery procedure reliably reconstructs the bit sequence from one attack trace under each chosen ciphertext. Combined with the precomputed ciphertext table, only 6 chosen ciphertexts are required to recover the private key of Kyber512, and only 9 chosen ciphertexts are required for Kyber768. Compared with the prior poly_frommsg()-based method, which requires 8 and 12 ciphertexts, respectively, the proposed method reduces the ciphertext requirement by 25.0% while maintaining a 100% success rate (Table 4 ). Compared with the attack on poly_tomsg(), the proposed method exploits a function with stronger leakage observability and therefore achieves higher decision stability and equal or better overall efficiency. The periodic PoI model is thus not only an empirical observation, but also a direct basis for the attack design and a key reason for the practical gain in low-query key recovery. Conclusions This work shows that Kyber contains different implementation-level vulnerabilities along its decapsulation path and that poly_frommsg() is the most critical leakage point in the tested software implementation. By combining function-level TVLA assessment with a chosen-ciphertext SPA attack, the study identifies leakage sources in poly_reduce(), poly_tomsg(), poly_frommsg(), and G(). It also converts the observed periodic leakage structure of poly_frommsg() into an effective grouped threshold model for key recovery. The resulting attack reduces the number of required ciphertexts for Kyber512 and Kyber768 to 6 and 9, respectively, while preserving a 100% success rate. These findings indicate that practical protection of post-quantum software should go beyond algorithm-level security claims. Masking, execution randomization, balanced implementations, and function-level leakage testing should be considered explicitly during deployment and validation.
A Spatiotemporal Coupling Traffic Flow Prediction Model with Dynamic Graph Recursion and State Space
2026, 48(6): 2655-2665.
doi: 10.11999/JEIT251198
Abstract:
Objective Accurate traffic flow prediction is a key task in intelligent transportation systems. However, it remains challenging to capture dynamically evolving spatial structures and complex spatiotemporal dependencies in urban road networks. To address these issues, this paper proposes DGGRU-Mamba, a spatiotemporal traffic flow prediction framework that integrates dynamic graph recurrent modeling with a structured state space mechanism. The model jointly captures dynamic spatial dependencies and long-range temporal dependencies. Methods DGGRU-Mamba contains two core modules: Dynamic Graph Recurrent Modeling (DGRM) and Spatiotemporal Mamba (ST-Mamba). A spatiotemporal embedding generator is first designed to jointly encode periodic temporal information and node-specific spatial features, thereby supporting adaptive graph construction. DGRM dynamically updates time-varying adjacency structures through Dynamic Graph Gated Recurrent Units (DGGRUs), which enables adaptive modeling of evolving spatial dependencies. ST-Mamba uses structured state transitions to efficiently capture long-range temporal dependencies. In addition, a dual-branch prediction scheme with Forecast and Backcast branches is used to improve multi-step prediction accuracy and reduce cumulative errors. Results and Discussions DGGRU-Mamba is evaluated on four benchmark datasets, namely PEMS03, PEMS04, PEMS07, and PEMS08. Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) are used as evaluation metrics. Experimental results show that DGGRU-Mamba achieves strong performance on all datasets. On PEMS04, compared with the mainstream attention-based model STAEformer, DGGRU-Mamba reduces MAE, RMSE, and MAPE by approximately 4.2%, 3.8%, and 2.9%, respectively. Its inference time is also shortened by 4.82 s. These results indicate that the proposed framework improves prediction accuracy while maintaining high computational efficiency. The performance gains mainly arise from the complementary effects of DGRM and ST-Mamba, which strengthen dynamic spatial dependency modeling and long-range temporal dependency learning with lower computational cost. Conclusions This paper proposes DGGRU-Mamba, a spatiotemporal traffic flow prediction framework for modeling dynamic spatial structures and long-range temporal dependencies in complex traffic networks. By integrating dynamic graph recurrent modeling with a structured state space mechanism, the framework achieves a favorable balance between prediction accuracy and computational efficiency. Experiments on multiple benchmark datasets verify its effectiveness and scalability in multi-step traffic flow prediction. Future work will consider external factors, such as weather and traffic events, to further improve its applicability in real traffic scenarios.
2026, 48(6): 2666-2677.
doi: 10.11999/JEIT260044
Abstract:
Objective Agricultural pest images are often affected by complex background interference, large appearance differences across morphological stages, diverse shooting angles, and substantial scale variation. These factors limit feature extraction and morphological adaptability in existing fine-grained classification models. To address these challenges, an Agricultural Pest Multi-Dimensional dataset (APMD) is constructed to cover multiple morphological stages, viewing angles, and object scales. In addition, a Dynamic Focus and Semantic Prompt Network for fine-grained pest classification (DFS-PestNet) is proposed. The network adopts a decoupled parallel architecture that combines a main feature stream and a prompt enhancement stream. A Spatial Dependency Perception (SDP) module is designed to dynamically focus on key discriminative regions, such as pest spots and wing veins, thereby improving local subtle feature extraction under complex backgrounds. An Advanced Haptic-Visual Prompting (AHVP) module is introduced to integrate category semantics and spatial position information into shallow and middle-level features, which improves adaptability to morphological variations across developmental stages. Dual-branch Saliency Sampling (DSS) is further adopted to adaptively aggregate key features from essential pest body parts through learnable prototype components and dual-branch saliency fusion. This strategy improves the recognition of small targets, including tiny pests and early-stage larvae. Experimental results show that the proposed model achieves better classification performance than baseline and mainstream methods on both public and self-constructed datasets. These results verify the effectiveness and application potential of the model in complex agricultural scenarios and provide a technical reference for intelligent pest monitoring and precise control in smart agriculture. Methods To improve classification accuracy under complex background interference and multi-morphological conditions, APMD is first constructed. This dataset contains image data covering different pest morphological stages, viewing angles, and scales. Specifically, it includes 15,680 images from 58 species, which are divided into training, validation, and testing sets at a standard ratio of 7:2:1 (Fig. 1 ) (Table 1 ). The dataset provides high-quality data support for research on fine-grained pest classification. DFS-PestNet is then proposed. In this network, the SDP module is designed to adaptively locate and enhance key discriminative pest regions. By reducing the effects of pose variation and complex background interference, this module enables more accurate fine-grained feature extraction. The AHVP module is also incorporated into the network to embed category semantics and spatial position information. This module guides the network to focus on key discriminative features across different morphological periods, thereby improving recognition robustness under large morphological changes during the pest life cycle. Furthermore, DSS is proposed to adaptively aggregate features from essential pest body parts. This strategy strengthens the recognition of challenging small targets and reduces the difficulty of small-target recognition in fine-grained pest classification. Results and Discussions The performance of DFS-PestNet in fine-grained pest classification is evaluated through multidimensional experiments. First, qualitative visualization is conducted. Grad-CAM heatmaps show that, compared with the baseline model, which is easily affected by complex farmland backgrounds and plant stems, DFS-PestNet effectively suppresses background noise and focuses on fine-grained discriminative parts, such as pest heads and antennae (Fig. 6 ). The model also shows clear advantages in capturing features of tiny targets, such as leafhopper nymphs, and pests at different life stages, such as Chilo suppressalis hidden within stems. The t-SNE feature reduction results further confirm that the proposed model reduces feature confusion in multi-morphological scenarios. High-dimensional features show clearer inter-class separation and tighter intra-class clustering in a two-dimensional visual space (Fig. 7 ). Second, quantitative ablation and parameter optimization experiments are performed. The ablation studies validate the synergistic effect of the three improved modules, namely SDP, AHVP, and DSS (Table 2 ). Their combination increases the classification accuracy of the baseline model by 2.21%, reaching 77.24%, with all core evaluation metrics achieving the best values. Hyperparameter optimization further identifies 6 as the optimal number of prompt position tokens and 0.2 as the optimal feature dropout rate (Fig. 8 ). This configuration ensures sufficient semantic representation while achieving a good balance between simulating natural occlusion and improving model robustness. Finally, comparative experiments with mainstream state-of-the-art models are conducted. Compared with existing advanced Convolutional Neural Network (CNN) and Transformer architectures, such as Gate-ViT and EST, DFS-PestNet achieves the highest accuracies of 77.24% and 98.01% on the large-scale public dataset IP102 and the challenging self-constructed APMD dataset, respectively (Table 3 ) (Table 4 ). These results show consistent improvements across fine-grained classification metrics. Moreover, while maintaining high classification accuracy, the proposed model achieves inference speeds of 158 frame/s and 164 frame/s on the two datasets, respectively. In summary, DFS-PestNet achieves strong classification accuracy and high inference efficiency for complex pest feature extraction across large scale variation and multiple morphological stages. This provides a practical basis for efficient deployment in smart agriculture. Conclusions To address multi-morphological variation and small-target recognition in fine-grained pest classification, the APMD dataset is constructed, and DFS-PestNet is proposed based on the MPSA baseline. Specifically, the SDP module is introduced to adaptively focus on pose- and morphology-invariant discriminative features. The AHVP module embeds category semantics and spatial position information into shallow and middle-level networks. The DSS module adaptively aggregates key body-part features to improve small-target recognition. Experimental results show that DFS-PestNet outperforms mainstream models on both the IP102 and APMD datasets across different developmental stages, angles, and scales. Future work will focus on lightweight model design for efficient edge deployment and open-set recognition for early warning of unknown pest categories in complex real-world environments.
2026, 48(6): 2678-2689.
doi: 10.11999/JEIT251303
Abstract:
Objective The rapid development of remote sensing imaging has generated large-scale and diverse data for remote sensing land-cover classification. In recent years, Mamba-based models have been successfully applied to image processing because of their distinctive architectures and strong global modeling capability. Among them, multi-scale vision Mamba models are well suited to complex spatial distributions. This property matches remote sensing scenes, in which ground objects often have large scale variations and complex orientations. To fully use the advantages of Mamba in feature extraction and fusion for remote sensing data, a Mamba-based Multi-modal and Multi-scale fusion model for Remote Sensing land-cover classification (M3RS) is proposed. Methods M3RS mainly contains three stages for feature extraction and fusion. First, a Multi-Scale Spatial Encoder based on Spatial Mamba is used to extract features from Light Detection And Ranging (LiDAR) images and Synthetic Aperture Radar (SAR) images. Considering the unique data structure of HyperSpectral Image (HSI), a Multi-Scale Spatio-Spectral Encoder is proposed to extract complex spatio-spectral features by using Spatial Mamba and Spectral Mamba. Next, a Multi-Modal Feature Fusion Module, consisting of the proposed Cross-Mamba and Channel-Concatenated Mamba, is designed to fuse multi-modal features. Cross-Mamba efficiently fuses multi-modal spatial features through the interaction of State Space Model (SSM) parameters from different modalities. Channel-Concatenated Mamba further fuses multi-modal features by constructing four channel scanning strategies. Finally, an improved Multi-Scale Feature Fusion Module is adopted to fuse multi-scale features layer by layer. This design provides highly discriminative features for classification and improves the accuracy of remote sensing land-cover classification. Results and Discussions Comparative experiments are conducted on three publicly available multi-modal remote sensing land-cover classification datasets. The proposed model is compared with seven mainstream models. The results show that M3RS achieves the best Overall Accuracy (OA), Average Accuracy (AA), and Kappa coefficient among all compared methods. On the Muufl dataset, the OA of M3RS is 3.49%, 3.80%, and 4.02% higher than those of representative Convolutional Neural Network (CNN)-, Transformer-, and Mamba-based models, respectively (Table 1 , Fig. 8 ). On the Houston2013 and Augsburg datasets, the OA of M3RS exceeds those of all compared algorithms by an average of 3.37% and 3.11%, respectively (Tables 2 and 3 ). These results indicate that integrating a multi-modal and multi-scale architecture with Mamba improves the accuracy of remote sensing land-cover classification. In addition, the ablation experiment verifies the contribution of each proposed module to classification performance (Table 4 ). Spectral Mamba provides a clear accuracy gain, and the fusion modules further improve the overall performance to different degrees. The hyperparameter experiment also provides a useful configuration for multi-scale remote sensing image fusion (Table 5 ). Compared with a Transformer model using the same multi-scale architecture, M3RS achieves higher classification accuracy, reduces the parameter count by 37.4%, and shortens the training time by 10.7%. These results show that Mamba improves both accuracy and efficiency in this framework (Fig. 9 ). Conclusions M3RS uses Mamba to fuse multi-modal and multi-scale features, thereby improving remote sensing land-cover classification. The heterogeneous encoders in M3RS address differences among multi-modal data and provide richer complementary information for fusion and classification. Cross-Mamba and Channel-Concatenated Mamba account for both the similarities and differences between Mamba and Transformer. They achieve efficient multi-modal spatial feature interaction and comprehensive multi-modal feature fusion, respectively, forming a hierarchical fusion strategy. The multi-scale architecture also alleviates the difficulty caused by complex spatial distributions of remote sensing land covers. The proposed Multi-Scale Feature Fusion Module, composed of Spatial Mamba and channel attention, integrates multi-scale features and provides a reliable basis for subsequent classification. Future work will further optimize the model by exploring the principles of Mamba and refining feature alignment in cross-attention-based multi-modal interaction, thereby improving the reliability of feature fusion.
2026, 48(6): 2690-2700.
doi: 10.11999/JEIT251271
Abstract:
Objective With the rapid development and wide use of image editing tools, such as Adobe Photoshop and Meitu, realistic forged images can now be created and disseminated with increasing ease. This trend poses challenges to visual content authentication in journalism, forensic analysis, and social security. Existing image forgery detection methods usually define the task as pixel-wise binary classification. This formulation may cause label conflicts, especially when the same object has different labels in different images. In addition, most methods mainly focus on spatial-domain features and make limited use of complementary information from other views, such as noise-domain clues. Methods To address these limitations, this paper proposes an image forgery detection algorithm based on multi-view feature extraction and dual-edge contrastive learning. The detection task is reformulated as intra-image inconsistency detection, which avoids label conflicts caused by conventional pixel-wise classification. To reduce semantic ambiguity near tampered boundaries, a dual-edge contrastive learning strategy is designed. Inner-edge and outer-edge features are extracted and contrasted separately, and non-edge tampered and non-tampered features are also contrasted. This strategy guides the model to focus on difficult edge samples and improves boundary detection accuracy. A dual-branch multi-view feature encoder is further developed to extract complementary forgery clues. The spatial-domain branch uses a High-Resolution Network (HRNet) backbone to extract multi-scale spatial features. A mixture-of-experts gating mechanism dynamically weights features across scales and fuses residuals between adjacent scales, which helps capture subtle forgery traces. The noise-domain branch extracts multiple noise-related features, including noise fingerprint features, Spatial Rich Model (SRM) filter responses, Bayar convolution features, max-pooling features, average-pooling residuals, and learnable Fourier-domain features with adaptive masking. A mixture-of-experts strategy is also used to dynamically assign weights to these heterogeneous features according to the characteristics of each input image. During training, the fused multi-view features are optimized using the dual-edge contrastive learning framework, which strengthens discrimination between tampered and non-tampered regions, particularly near their boundaries. During inference, K-means clustering is applied to the learned feature representations to locate tampered regions without explicit pixel labels. Results and Discussions Extensive experiments are conducted on widely used benchmark datasets, including NIST, Columbia, COVERAGE, DSO, and CASIA-v1. These datasets cover different forgery types, including splicing, copy-move, object removal, and post-processing. The proposed method consistently outperforms state-of-the-art methods. Compared with the best existing methods, it improves the average permuted F1 (pF1) and permuted Intersection over Union (pIoU) by 26.0% and 10.1%, respectively (Table 3 ). Visualization results show more accurate localization of tampered regions, especially along tampered boundaries, with fewer false positives and clearer edge delineation (Fig. 5 ). Ablation studies further verify the effectiveness of each key component, including multi-view feature extraction, the mixture-of-experts fusion mechanism for noise features, and the dual-edge contrastive learning strategy (Tables 4 ～6 ). Conclusions This paper presents an image forgery detection framework that addresses the limitations of conventional classification-based methods by modeling the task as intra-image inconsistency detection. Dual-edge contrastive learning reduces semantic ambiguity at tampered boundaries, and the multi-view feature encoder extracts complementary spatial-domain and noise-domain clues. Experimental results on different datasets show improved detection accuracy and boundary precision. Future work will explore the extension of the inconsistency detection paradigm to additional modalities, such as text, for multimodal forgery detection.
2026, 48(6): 2701-2711.
doi: 10.11999/JEIT251252
Abstract:
Objective With the rapid development of satellite imaging technology, demand has increased for high-resolution multispectral remote sensing images in a wide range of applications. However, satellite platforms differ in sensor parameters and imaging conditions, which leads to clear domain shifts among datasets acquired by different satellites. Most existing Deep Learning (DL)-based pansharpening methods are therefore trained separately on individual satellite datasets and have limited cross-satellite generalization. To address this limitation, this study proposes a Semantic-guided Unified Multi-scale Deep Unrolling Network (SUM-DUN). SUM-DUN is designed based on classical optimization theory and adopts a three-dimensional (3D) multi-scale deep unrolling architecture for unified feature extraction and fusion. Multimodal Large Language Models (MLLMs) are used to generate semantic text prompts from the input images. These prompts guide the model to adaptively adjust feature representations and improve fusion quality. The proposed method aims to support unified remote sensing image fusion through a tailored network architecture and a prompt-guided mechanism, thereby providing reliable data for high-level image interpretation tasks. Methods Following the Maximum A Posteriori (MAP) estimation principle, the optimization process for High-Resolution Multispectral (HRMS) image recovery is unfolded into the proposed SUM-DUN (Fig. 1). Each iterative stage of SUM-DUN contains two main modules: a Gradient Descent Module (GDM) and a Semantic-guided Proximal Mapping Network (SPMN). These modules approximate the operations in Eq. (5) and Eq. (6), respectively. GDM performs gradient descent updating based on the current feature estimate and the degradation model. SPMN is implemented using a Transformer-based architecture, as shown in Fig. 2(b), and incorporates semantic text prompts generated from each input image pair by MLLMs. These prompts guide the network to select suitable feature propagation strategies for the current image pair. This process helps suppress noise and reduce discrepancies among different satellite sensors. Through upsampling and downsampling operations, the network also transmits multispectral (MS) and panchromatic (PAN) features across iterative stages. Thus, multi-scale spatial and spectral information is progressively preserved and enhanced during the deep unrolling process. Results and Discussions To verify the effectiveness of the proposed method, it is compared with seven representative baselines, including two traditional methods, BDSD and PRACS, and five DL-based methods, AWFLN, FusionMamba, PanMamba, WFANet, and TMDiff. In the reduced-resolution evaluation, ground-truth HRMS images are available. Several widely used reference-based metrics are adopted, including Spectral Angle Mapper (SAM), Spatial Correlation Coefficient (SCC), Peak Signal-to-Noise Ratio (PSNR), Erreur Relative Globale Adimensionnelle de Synthèse (ERGAS), Averaged Universal Image Quality Index (QAVE), and the Universal Image Quality Index for 4-band and 8-band images (Q4/Q8). These metrics jointly assess spectral fidelity, spatial consistency, and overall image quality. In the full-resolution evaluation, ground-truth HRMS images are unavailable. Therefore, no-reference quality indices are used. Specifically, Hybrid Quality with No Reference (HQNR), its spectral distortion component and spatial distortion component are used to assess fusion quality in real-world scenarios. Quantitative results on the GF-1 and WV-2 test datasets show that the proposed method consistently achieves the best or second-best performance across all metrics under both reduced-resolution and full-resolution settings (Tables 2 and 3 ). These results indicate that the proposed method can preserve spectral fidelity and spatial consistency while maintaining robust performance across different satellites and challenging imaging conditions. Ablation studies further validate the effectiveness of the 3D architecture, the multi-scale network design, and the spatial-channel prompt guidance mechanism. Removing or modifying any of these components causes performance degradation to different degrees (Tables 4 and 5 ). Conclusions This study proposes a semantic-guided unified multi-scale deep unrolling method for pansharpening. The method uses semantic prompts generated by an MLLM to support efficient and unified fusion of images from different satellites. The proposed approach is built on a deep unrolling framework and uses a 3D convolutional architecture to process satellite datasets with different numbers of spectral bands. A multi-scale network design is further used to extract spatial and spectral features at different levels, thereby improving fusion performance. In addition, a Semantic Prompt Integration Module (SPIM) is designed to adaptively route spatial and channel features based on semantic information. SPIM enables more effective feature propagation and improves both spatial detail reconstruction and spectral consistency. Extensive experiments show that the proposed method achieves state-of-the-art performance in visual quality and quantitative evaluation.
2026, 48(6): 2712-2721.
doi: 10.11999/JEIT250868
Abstract:
In the era of big data, image quality varies greatly. Reconstructing high-resolution images from low-quality inputs is therefore an important task in computer vision. Existing super-resolution methods based on window self-attention, such as SwinIR, have limited receptive fields and insufficient ability to capture high-frequency details. These limitations weaken the reconstruction of fine image structures. To address these issues, this study proposes the Heterogeneous Windows Transformer Network for Image Super-Resolution (HWT-SRNet), a new architecture built on SwinIR. Through targeted module design, HWT-SRNet improves high-frequency detail extraction while expanding the receptive field, providing an effective solution for image super-resolution. Methods Based on SwinIR, this study designs two key modules to improve super-resolution reconstruction. First, the Residual Heterogeneous Windows Transformer Block (RHWTB) alternates square windows and pale-shaped windows. This design preserves local feature extraction while expanding the receptive field, enabling the network to model both fine-grained details and global structural information. The window size and alternation frequency are optimized to balance computational efficiency and feature extraction. Second, the High-Frequency Prior Feature Extraction Network (HFPFEN) is used to compensate for the limited high-frequency modeling ability of Transformer-based super-resolution models. HFPFEN extracts high-frequency prior information from images using a Difference of Gaussians (DoG) filter. The DoG filter emphasizes edges and textures by computing the difference between lightly and heavily blurred images. The extracted high-frequency information is then fused with the heterogeneous window attention mechanism. This design allows HWT-SRNet to enhance fine details while maintaining structural coherence. Because the DoG filter is applied in the spatial domain, the model can capture and reconstruct sharp edges and textures without frequency-domain transformation. Results and Discussions Experiments are conducted on five widely used benchmark datasets: Set5, Set14, BSD100, Urban100, and Manga109. HWT-SRNet is compared with representative advanced methods, including ACT, ART, and CAT. The results show superior performance across key evaluation metrics (Table 1 ). Compared with baseline models, HWT-SRNet improves the Peak Signal-to-Noise Ratio (PSNR) by 0.10 dB to 0.37 dB, confirming its effect in improving image quality. The Structural Similarity Index Measure (SSIM) also shows consistent improvement, indicating better perceptual quality and more accurate reconstruction. Qualitative results further show that HWT-SRNet restores sharper edges, preserves textures more effectively, and reduces blurring artifacts. Ablation studies are conducted to evaluate the contributions of RHWTB and HFPFEN (Tables. 4 , 5 ). The results confirm that heterogeneous window attention and high-frequency prior extraction jointly improve local feature refinement and global context modeling. Therefore, HWT-SRNet provides an efficient solution for receptive field expansion and high-frequency detail reconstruction. Conclusion This paper proposes HWT-SRNet to address the limited receptive fields and insufficient high-frequency detail capture of existing super-resolution algorithms. By integrating heterogeneous window attention with high-frequency prior feature extraction, the model achieves more effective fusion of local and global features. Experimental results confirm that HWT-SRNet improves both PSNR and SSIM and outperforms representative advanced methods. However, this study does not specifically examine the model’s adaptability to noise interference in real-world scenarios. Future research can further improve the robustness of HWT-SRNet to noisy and degraded inputs and evaluate its generalization on specialized datasets, such as medical and satellite images.
2026, 48(6): 2722-2732.
doi: 10.11999/JEIT251379
Abstract:
To address the disconnect between conventional fixed-threshold early warning methods for power MOSFETs and their physical failure mechanisms, this paper proposes a lifetime prediction framework that integrates Explainable Artificial Intelligence (XAI). First, an adaptive dual-threshold partitioning strategy is designed by combining K-means clustering with the Proximal Policy Optimization (PPO) algorithm. The initial solution obtained by K-means is used as the search starting point. A multi-objective reward function is then constructed to balance interval proportion, state-transition sensitivity, and threshold-spacing penalties. This function guides the agent in threshold optimization and enables accurate partitioning of degradation stages. Second, SHAPley additive explanations (SHAP) analysis is introduced to improve the interpretability of the black-box decision-making process. It verifies the rationality of threshold decisions from the perspective of feature-mechanism correlations. The results show that the low threshold is mainly governed by steady-state features in the healthy stage and meets the safety baseline requirement. The high threshold is dominated by dynamic features of late-stage accelerated degradation and accurately identifies the critical point. These findings confirm the reliability and transparency of the threshold decisions. Based on this framework, an early warning mechanism is triggered when degradation data exceed the reliable low threshold. A Residual-connected Stacked Gated Recurrent Unit (R-SGRU) is then used for Remaining Useful Life (RUL) prediction. Experiments on the NASA dataset show that the proposed model outperforms several baseline models, including Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN). The test-set Mean Squared Error (MSE) is below 0.001 5, and R2 is above 0.98. This study provides accurate and reliable decision support for early warning in MOSFETs. It also links data features with physical mechanisms through explainable techniques, supporting the development of trustworthy artificial intelligence for device prognostics. Objective This study addresses two key issues in power MOSFET lifetime prognostics: the disconnect between conventional fixed-threshold early warning methods and physical mechanisms, and the limited interpretability of existing approaches. A framework integrating adaptive dual-threshold partitioning with XAI is proposed to support predictive maintenance with both physical credibility and high prediction accuracy. Methods An adaptive dual-threshold partitioning strategy is proposed by integrating K-means clustering with PPO reinforcement learning. Threshold positions are optimized using a multi-objective reward function to accurately identify degradation stages. SHAP analysis is used to quantify the contributions of 13-dimensional morphological features based on Shapley values. This validates the physical rationality of threshold decisions from a mechanistic perspective. When degradation data exceed the low threshold, an early warning is triggered. The R-SGRU network is then used for RUL prediction by capturing long-term dependencies through its gating mechanism. The proposed method is validated using the NASA dataset, forming a complete technical route from intelligent early warning to accurate prediction. Results and Discussions The thresholds optimized by PPO achieve the best performance across all metrics (Table 1 ). SHAP analysis reveals the physical rationale for the threshold decisions. In the healthy stage, the low threshold is mainly governed by steady-state features. By contrast, the high threshold is determined by accelerated degradation dynamics. This result establishes a quantitative correlation between data-driven results and physical failure mechanisms. SHAP interaction heatmaps (Figs. 6 and 7 ) further show the synergistic effects among features. Device failure is a complex process driven by the coordinated evolution of multiple features. The R-SGRU prediction model based on the optimized thresholds shows excellent performance on the NASA dataset (Table 5 ). Across the four device groups, the model achieves an MSE below 0.001 5 and an R2 above 0.98, outperforming the baseline models. Conclusions This study proposes an XAI-based framework for predicting the RUL of power MOSFETs. For threshold partitioning, an adaptive dual-threshold strategy combining K-means clustering and PPO reinforcement learning is adopted. A multi-objective reward function enables accurate identification of nonlinear degradation stages, and its performance is validated across four test devices. For interpretability, SHAP analysis provides mechanistic support for threshold decisions. The results show that low thresholds depend on steady-state features in the healthy period, whereas high thresholds are dominated by late-stage accelerated degradation features. This pattern is consistent with actual failure mechanisms. Feature interaction heatmaps reveal complex cooperative effects among multiple features and improve the understanding of the decision-making process. The R-SGRU prediction model shows strong time-series modeling capability and ensures high stability and accuracy. This work establishes a complete technical route from intelligent early warning to accurate prediction. It achieves adaptive threshold optimization and links data-driven results with physical mechanisms through interpretability analysis. The findings provide reliable support for the intelligent operation and maintenance of power MOSFETs.
2026, 48(6): 2733-2742.
doi: 10.11999/JEIT251084
Abstract:
Objective The increasing demand for immersive media propels point cloud video into the spotlight for applications such as virtual and augmented reality. However, the massive data volume of point cloud streams poses a significant challenge to current network infrastructures, jeopardizing the user’s Quality of Experience (QoE) under limited bandwidth. Existing Adaptive BitRate (ABR) streaming solutions are hindered by two primary limitations. Viewport prediction models often focus solely on temporal features, leading to insufficient accuracy for long-term predictions in complex Six-Degrees-of-Freedom (6DoF) movement. Concurrently, dynamic quality allocation strategies struggle to make optimal online decisions under the uncertainties of prediction errors and network fluctuations, failing to effectively balance conflicting QoE metrics. This research addresses these challenges by proposing an integrated framework that combines high-precision viewport prediction with intelligent, context-aware quality allocation to enhance QoE for point cloud video streaming. Methods The proposed method integrates a graph-based viewport prediction scheme with a context-aware quality allocation mechanism. For viewport prediction, an “anchor point graph” is constructed to explicitly model the user’s spatial movement patterns. This graph is processed using representation learning to generate low-dimensional embeddings for each anchor point, which encapsulate rich spatial context. These learned spatial features are concatenated with real-time 6DoF viewport data to form a fused feature sequence. A stacked Long Short-Term Memory (LSTM) network processes this sequence to accurately predict the user’s future viewport trajectory. For quality allocation, the sequential decision-making process is modeled as a contextual bandit problem, adopting the LinUCB algorithm as the decision engine. At each decision epoch, a context vector is constructed for each spatial tile, incorporating critical information such as its predicted utility, historical quality level, and location relative to the predicted viewport. The LinUCB algorithm utilizes this context to select an optimal action for each tile, thereby maximizing cumulative QoE under the bandwidth budget, as detailed in Algorithm 1. Results and Discussions Extensive simulations validate the framework’s performance using the public 8i Voxelized Full Bodies dataset, real-world user viewport traces, and 5G network bandwidth profiles. In the viewport prediction task, the proposed model significantly outperforms baselines, achieving a stable average F1-score of 0.984 (Fig. 4 ) and maintaining a consistently low Root-Mean-Square Error (RMSE) over long prediction horizons (Fig. 3 ). In the end-to-end streaming evaluation, the integrated framework demonstrates remarkable improvements in overall QoE. Cumulative Distribution Function (CDF) plots reveal that the proposed scheme consistently delivers higher QoE, user-perceived utility, and video quality, while incurring the lowest quality fluctuation (Fig. 5 ). Notably, under fluctuating network conditions, the solution improves the mean QoE by 54.82% compared to the next-best baseline at an average bandwidth of 100 Mbps (Fig. 6 ), highlighting its efficiency in resource-constrained environments. Conclusions This paper presents a complete adaptive streaming framework to address the QoE optimization challenge for point cloud video. By developing a novel 6DoF viewport prediction model that leverages graph representation learning, long-term prediction accuracy is significantly enhanced. Furthermore, by framing dynamic quality allocation as a contextual bandit problem, the system makes intelligent, online decisions that adapt to both prediction outcomes and dynamic network conditions. Comprehensive experimental results validate the effectiveness of this integrated approach, which consistently outperforms existing solutions in both prediction accuracy and overall user QoE.
2026, 48(6): 2743-2753.
doi: 10.11999/JEIT251290
Abstract:
Objective The rapid development of Generative Adversarial Network (GAN) and Diffusion Model (DM) techniques has sharply increased the number of fake images. The wide dissemination of such images poses potential and unpredictable risks to individuals, society, and national security. Efficient and highly generalizable deepfake detection methods are therefore needed. Cross-domain detection has become a central task in deepfake detection. However, existing methods often rely on specific artifacts or fixed parameters for feature extraction. They also learn spatial and frequency modalities separately, lack dynamic interaction mechanisms, and provide insufficient global feature association. To address these limitations, a Pyramid Interactive Dual-Stream Network (PIDSNet) is proposed. This network integrates dynamic artifact tracking with spatio-frequency interaction analysis. Methods PIDSNet consists of spatial and frequency branches (Fig. 1 ) and four cooperative modules: the Multi-Branch Feature Extraction (MBFE) module, the Frequency Domain Feature Extraction (FDFE) module, the Pyramid Spatio-Frequency Interaction (PSFI) module, and the Multi-Head Pyramid Squeeze Attention (MHPSA) module. MBFE (Fig. 2 ), which serves as the basic unit of the spatial branch, uses multilevel and multibranch dilated convolutions to reduce information loss as the receptive field expands. It extracts global and local features jointly. FDFE, which is central to the frequency branch, combines MBFE with spectral convolution to dynamically identify frequency-domain artifact features. This design reduces the reliance of traditional frequency-domain methods on fixed parameters and frequency bands. It also improves the adaptive capture of artifacts from different generative models. PSFI drives interaction between the two branches (Fig. 3 ). It constructs a Gaussian pyramid in the spatial domain and a Laplacian pyramid in the frequency domain to capture low-frequency global information and high-frequency details, respectively. Dynamic weighting at each pyramid level supports adaptive spatio-frequency feature fusion and builds a dynamic interaction mechanism. MHPSA integrates Multi-Head Self-Attention (MHSA) with dilated convolution (Fig. 4 ). It retains the local detail capture ability of the Pyramid Squeeze Attention (PSA) module and strengthens global feature modeling, thereby improving model adaptability and robustness. Results and Discussions To evaluate cross-domain detection across different generative paradigms, PIDSNet is trained on the ProGAN subset and tested on multiple GAN and DM datasets. For GAN detection, the mean Accuracy (Acc.) reaches 95.2% on the ForenSynths test set containing four GANs (Table 3 ). This value is 5.3 and 5.2 percentage points higher than those of LGrad and FreqNet, respectively. On the GANGen dataset containing nine GANs (Tables 4 and 5 ), the mean Acc. reaches 95.5%. This result represents a 20.1 percentage-point improvement over F3Net. Compared with FreqNet, PIDSNet improves mean Acc. and mean Average Precision (A.P.) by 4.1 and 1.3 percentage points, respectively. For DM detection, tests are conducted on the DiffusionForensics and Ojha datasets. On DiffusionForensics (Table 6 ), the mean Acc. reaches 95.4%, which is 4.8 and 13.2 percentage points higher than those of LGrad and FreqNet, respectively. On Ojha (Table 7 ), the mean Acc. and mean A.P. reach 96.1% and 99.4%, respectively. More importantly, PIDSNet has only 2.4 M parameters (Table 8 ) and achieves mean Acc. and mean A.P. values of 95.7% and 98.7% across 25 datasets, outperforming competing methods. These experiments indicate that PIDSNet, although trained only on the ProGAN subset, adapts to multiple GAN types and effectively detects DM-generated images with different spatial and frequency artifact characteristics. This confirms its strong cross-model and cross-paradigm generalization ability. In addition, Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations indicate that PIDSNet identifies detection-relevant regions in face images, although face images are absent from the training data (Fig. 5 ). Conclusions This study addresses the weak domain adaptability and poor generalization of current GAN and DM detection methods, which often rely on domain-specific artifacts or fixed parameters and have limited modality interaction. A spatio-frequency collaborative learning framework and a dynamic artifact tracking mechanism are constructed to reduce reliance on specific artifacts and fixed parameters. This design improves the extraction of general forgery features. The effectiveness of PIDSNet is validated on image datasets generated by 25 different GAN and DM models. Compared with current advanced models, the mean Acc. and A.P. are improved, confirming strong performance in cross-domain deepfake detection. However, PIDSNet still has limitations. For specific models such as S3GAN, whose high-frequency energy distribution is close to that of real images, performance can still be improved. Future work should further optimize frequency-domain feature extraction, improve detection under compression distortion and noise interference, and study artifact separation and detection for images generated by multiple models. These directions may further improve model adaptability in complex real-world settings.
2026, 48(6): 2754-2764.
doi: 10.11999/JEIT251375
Abstract:
Objective As device dimensions in conventional two-dimensional integrated circuits approach fundamental physical limits, further improvements in performance and integration density face major challenges. Monolithic Three-Dimensional Integrated Circuits (M3D ICs), in which multiple active device layers are sequentially stacked on a single wafer, provide an effective way to address these limitations. In M3D ICs, Monolithic Interlayer Vias (MIVs) are used to realize vertical interconnections between device tiers. Compared with Through-Silicon Vias (TSVs), MIVs have much smaller dimensions, lower parasitic capacitance, and shorter interconnect delay. However, because of their small electrical variations and large numbers, defects are mainly manifested as subtle delay shifts. This places stringent demands on test accuracy, efficiency, and robustness to Process, Voltage, and Temperature (PVT) variations. Existing MIV test methods have limited scalability, show strong sensitivity to PVT variations, and have difficulty achieving both small-delay defect detection and fault localization in large-scale arrays. To address these issues, a parallel MIV test method based on a Time-to-Digital Converter (TDC) is presented to enable efficient and reliable testing of large MIV arrays with low area and time overhead. Methods Large-scale MIVs are logically organized into a two-dimensional array. Each basic test cell consists of a device-under-test MIV, a tri-state buffer, and a D flip-flop, and multiple cells are cascaded to form row and column test chains. By systematically exploiting the inherent input capacitance mismatch between the data and clock terminals of the D flip-flop, an embedded TDC structure that incorporates the MIV under test is constructed. Test stimuli are generated by a Digitally Controlled Delay Line (DCDL), which produces START and STOP pulse signals with multiplicatively adjustable phase differences. These signals are injected into different propagation paths of the test chains, so that time quantization is achieved through a signal-chasing mechanism. Structural symmetry between the test chains is used to reduce the effect of PVT variations. As the phase difference between START and STOP is progressively amplified, multiple TDC readings are collected to characterize defect-induced small delay variations and to distinguish them from measurement noise and PVT-induced fluctuations. After fault information is obtained for individual test chains, cross-analysis of the row and column test results is used to localize faults within the two-dimensional MIV array. Results and Discussions Simulation results based on the Nangate 45 nm standard cell library show that, under fault-free conditions, TDC readings obtained at different phase-difference settings exhibit a stable linear proportional relationship (Fig. 7 ). Extensive Monte Carlo simulations are conducted to determine a robust deviation tolerance threshold of 2, which effectively separates normal variations caused by PVT fluctuations from abnormal shifts induced by defects. Fault-injection experiments confirm that small-delay defects occurring in both the START chain and the STOP chain can be effectively detected and distinguished (Fig. 8 ). In terms of quantitative detection capability, the minimum detectable resistive open defect is approximately 8.4 kΩ, whereas the maximum detectable leakage defect and resistive short defect are about 67 kΩ and 32 kΩ, respectively, outperforming existing methods (Fig. 9 ). Moreover, the row-column decomposition architecture effectively alleviates the increase in test time as the MIV array size grows, which substantially reduces the overall test overhead. Area evaluation indicates that the average area overhead of the embedded built-in self-test structure is only 5.594 μm2 per MIV, which makes the method suitable for high-density M3D integration. Conclusions A parallel testing approach based on a TDC is presented for large-scale MIV arrays. By combining row-column decomposition, phase-difference multiplication, and a proportional deviation-based decision mechanism, efficient detection and accurate localization of both hard faults and small-delay defects are achieved. Structural symmetry within the test chains effectively improves robustness to PVT variations. Simulation results confirm that the proposed method can reliably detect resistive open, leakage, and short defects while maintaining low area and time overhead. Compared with existing techniques, the proposed method achieves a favorable balance among test accuracy, robustness to PVT variations, test efficiency, and hardware cost. Owing to its scalability and practical feasibility, it provides an effective and reliable solution for MIV testing in advanced M3D ICs.
News
more >
- JEIT has been included by DOAJ
- Call for papers of Radar Signal Processing Issue for JEIT
- WeChat public platform has been open for JEIT
- Call for papaers of Online Social Network for JEIT
- JEIT attended the 973 Project " social network "
- 2016 Annual NNFS Application Code and Research Direction
- JEIT Held the First Meeting of the Seventh Editorial Board
- JEIT won the title of "2014 Chinese TOP 100 Outstanding Academic Journals"
Conference
more >
- Call for papers of the 9th CWSN2015 Conference
- Notice to the 26th Annual Conference of Circle and System Branch, Chinese Institute of Electronics
- Notice of MAN2015 & ICMAN2015 Conference
- Call for papers 27th Annual Conference of Circle and System Branch, Chinese Institute of Electronics
- Notice on hosting the annual meeting of the 2016 Software Radar Technology Application Conference
- Notice on hosting the annual meeting of the 2016 Software Radar Technology Application Conference
- Call for paper of the 27th academic annual conference of circuits and systems branch
- Call for Papers of the Maritime Informantion Processing and Fusion
Author Center
Wechat Community