PLS-YOLO：一种轻量化的信号调制识别模型

周晓波; 张凡; 佘超; 周郭飞; 孟建平

doi:10.11999/JEIT251377

PLS-YOLO：一种轻量化的信号调制识别模型

doi: 10.11999/JEIT251377 cstr: 32379.14.JEIT251377

1.
北京交通大学电子信息工程学院北京 100044
2.
公安部第一研究所北京 100048

基金项目: 国家自然科学基金面上项目(52472174)，国家重点研发计划(2023YFB3208101)

详细信息

作者简介:
周晓波：男，副教授，研究方向为信号处理与专用集成电路设计

张凡：女，硕士生，研究方向为信号处理与数字集成电路设计

佘超：男，工程师，研究方向为信号处理与电路设计

周郭飞：男，研究员，研究方向为微弱信号处理技术与集成电路设计

孟建平：男，教授，研究方向为忆阻器与神经形态计算

通讯作者:
张凡　24125011@bjtu.edu.cn

中图分类号: TN929.5
计量
- 文章访问数: 12
- HTML全文浏览量: 5
- PDF下载量: 4
- 被引次数: 0
出版历程
- 收稿日期: 2025-12-30
- 修回日期: 2026-04-17
- 录用日期: 2026-04-17
- 网络出版日期: 2026-05-06

PLS-YOLO: A Lightweight Model for Signal Modulation Recognition

1.
School of Electronics and Information Engineering, Beijing Jiaotong University, Beijing 100044, China
2.
The First Research Institute of the Ministry of Public Security, Beijing 100048, China

摘要

摘要: 自动调制识别是无线通信频谱监测与安全保障的关键技术。针对当前基于深度学习的自动调制识别模型难以兼顾高识别准确率与低模型复杂度的问题，该文提出一种基于视觉目标检测的轻量化识别方法。首先利用短时傅里叶变换将IQ信号转化为时频图，并采用拼图形式进行预处理，将调制识别任务转化为视觉目标检测问题。随后以YOLOv10n为基座，构建了 Precision and Lightweight Structure-YOLO (PLS-YOLO)模型。该模型通过重构网络核心模块、优化主干网络通道降维策略、设计新型下采样结构以及改进注意力模块的前馈网络，有效实现了识别性能与轻量化结构的平衡。在RadioML2016.10a和RadioML2016.10b数据集上的实验结果表明，PLS-YOLO的平均精度分别达到68.4%和72.6%；与YOLOv10n相比，其参数量减少了47.33%，浮点运算次数降低了34.15%，且帧率提升了5帧/秒。研究结果证实，PLS-YOLO模型在显著降低计算成本的同时，依然保持了优异的识别精度。
- 深度学习 /
- 信号调制方式识别 /
- 轻量化 /
- YOLOv10
Abstract: Objective As wireless communication evolves toward high communication efficiency, low latency, and ubiquitous connectivity, stringent demands are deployed on Automatic Modulation Recognition (AMR) technology to ensure link reliability within complex electromagnetic environments. While deep learning has significantly enhanced recognition accuracy compared to traditional methods, which are often characterized by high subjectivity and poor robustness, existing YOLO-based AMR models remain unoptimized for specific signal characteristics and deployment scenarios. These models typically suffer from excessive parameters and high computational complexity, rendering them unsuitable for resource-constrained hardware such as edge nodes and FPGAs, and unable to meet real-time communication demands. A lightweight AMR model based on YOLOv10n, denoted as PLS-YOLO, is proposed to resolve the critical bottlenecks that restrict practical deployment of AMR techniques. By employing target strategies such as optimizing network channel, replacing core modules, and enhancing down-sampling mechanism, the integration of modulation signal classification and localization are realized swiftly. Furthermore, significant reductions in model parameters and computational complexity are achieved, thereby facilitating the adaptation of AMR models to resource-limited scenarios and providing technical support. Methods The experimental methodology centers on two core stages: dataset preprocessing and the construction of the PLS-YOLO model. In the preprocessing phase, the public benchmark datasets RadioML2016.10a and RadioML2016.10b from the field of signal modulation recognition are utilized as the foundation. For the In-phase and Quadrature (IQ) signals within these datasets, the Short-Time Fourier Transform (STFT) is employed to map one-dimensional temporal signals into two-dimensional time-frequency spectrograms containing critical information such as phase and amplitude, thereby providing richer feature representations for the model. Subsequently, a random sampling strategy without replacement is adopted to stitch single time-frequency samples into 3×3 aggregated images (Fig. 4), while target labels matching the input format of YOLO series models are synchronously generated. The dataset is ultimately partitioned into training, validation, and test sets at a ratio of 7:1.5:1.5 via stratified sampling to ensure the consistency of signal type distribution across all subsets. The model construction is based on the YOLOv10n architecture, with specific improvements implemented to achieve balance between parameter quantity and recognition performance in modulation recognition tasks. The C2f module in the original backbone network is replaced by the CSPPC module, based on the CSP architecture and comprising feature splitting, partial convolution processing, and feature fusion, to achieve the dual objectives of parameter reduction and recognition rate enhancement. Furthermore, the feature dimensionality reduction process of the backbone network is reconstructed to effectively mitigate the surge in computational load caused by parameter redundancy. The traditional down-sampling module is replaced by the innovative CGBlock, enhancing the capability to capture features of complex modulation signals by fusing context-aware information, thereby elevating recognition performance. Finally, standard convolutions in the PSA module and the v10Detect module are replaced with Partial Convolutions to further reduce computational complexity, realizing a synergistic optimization of lightweight design and recognition performance. Results and Discussions Experimental results on the RadioML2016.10a dataset indicate that the PLS-YOLO model achieves a mean Average Precision (mAP) of 68.4% within the signal-to-noise ratio (SNR) range of -20 to 18 dB, which further increases to 94.3% when the SNR is no less than 0 dB. Compared with the basic YOLOv10n model, PLS-YOLO attains a slight mAP improvement of 0.6% while reducing the parameter count by 47.33% and computational complexity by 34.15%, alongside an increase in inference speed by 5 FPS (Table 2). These findings verify that the model effectively balances performance with lightweight requirements by significantly decreasing computational costs while enhancing precision. To validate robustness, supplementary experiments are conducted on the RadioML2016.10b dataset. As shown in Table 4, the model achieves an mAP of 72.6% across the -20 to 18 dB range and 95.4% for SNR ≥ 0 dB, outperforming mainstream models such as MCNET and LSTM2, thereby demonstrating the superior performance of PLS-YOLO. Furthermore, as illustrated in Fig.5, it is observed that converting IQ data into spectrograms for PLS-YOLO recognition is more adaptive to digital modulation signals, whereas performance on analog modulation signals remains suboptimal; consequently, future research should focus on enhancing the recognition capabilities for analog signals. Conclusions This study proposes PLS-YOLO, a lightweight Automatic Modulation Recognition model based on YOLOv10n. To achieve synergistic optimization of modulation recognition performance and model lightweighting, the model structure is systematically improved through targeted strategies, including network channel dimension reduction, core functional module iteration, down-sampling mechanism innovation, and partial convolution replacement. Consequently, core bottlenecks prevalent in existing YOLO-based AMR models—such as parameter redundancy, high computational complexity, and limited adaptability to resource-constrained scenarios like edge nodes and FPGAs—are significantly reduced. Experimental results on the RadioML2016.10a and RadioML2016.10b benchmark datasets demonstrate that PLS-YOLO exhibits superior comprehensive performance. While the integrity of integrated signal classification and localization functions is maintained, both parameter and computational complexity are significantly reduced compared to the baseline YOLOv10n, accompanied by a notable enhancement in recognition accuracy, thereby significantly outperforming mainstream comparative models. In conclusion, the effectiveness and feasibility of the proposed optimization strategies are verified, providing a reliable technical path for the engineering implementation of AMR technology, while the identified potential for improvement in analog modulation signal recognition clarifies specific directions for future research.
- Deep Learning /
- Automatic Modulation Recognition /
- Lightweight /
- YOLOv10

HTML全文

图 1 PLS-YOLO 网络结构图

下载: 全尺寸图片幻灯片

图 2 部分卷积架构

下载: 全尺寸图片幻灯片

图 3 CSPPC架构

下载: 全尺寸图片幻灯片

图 4 拼图后数据

下载: 全尺寸图片幻灯片

图 5 各个调制方式下不同信噪比识别率

下载: 全尺寸图片幻灯片

图 6 混淆矩阵

下载: 全尺寸图片幻灯片

表 1 RadioML2016.10a数据集

类型	数据
样本数量	总样本数：22万单调制数：2万
调制方式	PSK、BPSK、CPFSK、GFSK、4-PAM、16-QAM、64-QAM、QPSK、AM-DSB、AM-SSB、WBFM
信噪比范围	–20 dB～18 dB 间隔为2 dB
样本大小	(2，128)

下载: 导出CSV

表 2 主流YOLO模型对比

模型	P	R	mAP	FPS	FLOPs/G	参数量	权重文件
YOLOv8n	0.721	0.625	0.671	121	8.1	3007793	5.4 MB
YOLOv10n	0.746	0.612	0.678	119	8.2	2635602	5.8 MB
YOLOv10s	0.750	0.620	0.681	109	24.5	8043474	15.7 MB
PLS-YOLO	0.735	0.632	0.684	124	5.4	1388277	3.4 MB

下载: 导出CSV

表 3 消融实验对比

CSPPC	PCF	CGBlock	GPSA+ v10PDetect	mAP	FLOPs/G	参数量
×	×	×	×	0.678	8.2	2635602
√	×	×	×	0.681	6.9	2245394
×	√	×	×	0.681	6.9	1920754
×	×	√	×	0.687	8.1	2699033
×	×	×	√	0.673	8.2	2625682
√	√	×	×	0.681	5.6	1404338
√	√	√	×	0.685	5.4	1457333
√	√	√	√	0.684	5.4	1388277

下载: 导出CSV

表 4 PLS-YOLO与当前主流AMR算法的性能对比

数据集	模型	识别精度 (%)		数据集	模型	识别精度 (%)
数据集	模型	0～18 dB	–20～18 dB	数据集	模型	0～18 dB	–20～18 dB
10a	MCNET^[17]	82.33	56.63	10b	MCNET^[17]	85.69	60.95
	LSTM2^[18]	91.41	60.56		LSTM2^[18]	88.65	63.28
	PET-CGDNN^[19]	89.45	60.38		PET-CGDNN^[19]	92.64	63.91
	MCLDNN^[20]	91.68	61.93		MCLDNN^[20]	93.02	64.46
	AMC-NET^[21]	91.32	62.40		AMC-NET^[21]	93.34	65.14
	FEA-T^[22]	88.54	60.55		FEA-T^[22]	93.12	64.44
	多模态联合两阶段网络^[23]	93.36	64.52		多模态联合两阶段网络^[23]	93.13	65.59
	文献[24]	92.06	62.44		文献[24]	93.24	64.61
	PLS-YOLO	94.30	68.40		PLS-YOLO	95.40	73.30
注：RadioML2016.10a数据集简称为：10a, RadioML2016.10b数据集简称为：10b。特别说明的是，由于PLS-YOLO属于目标检测模型，其评价指标为平均精度，涵盖了分类准确度与定位精度的综合考量。而MCNET、MCLDNN等均为分类模型，使用准确率作为指标。尽管两者在评价维度上存在差异，但为了直观展示本文模型在特征提取能力上的先进性，表4将两类指标进行并列参考性对比。

下载: 导出CSV

参考文献(24)

[1]	郑庆河, 李秉霖, 于治国, 等. 深度学习使能的自动调制分类技术研究进展[J]. 电子与信息学报, 2025, 47(11): 4096–4111. doi: 10.11999/JEIT250674. ZHENG Qinghe, LI Binglin, YU Zhiguo, et al. Research progress of deep learning enabled automatic modulation classification technology[J]. Journal of Electronics & Information Technology, 2025, 47(11): 4096–4111. doi: 10.11999/JEIT250674.
[2]	ZHANG Bojun. Electromagnetic signal modulation recognition based on subgraph embedding learning[EB/OL]. https://arxiv.org/abs/2508.13474, 2025.
[3]	YIN Ziqi, FENG Shanshan, LIU Shang, et al. LIST: Learning to index spatio-textual data for embedding based spatial keyword queries[J]. The VLDB Journal, 2025, 34(3): 23. doi: 10.1007/s00778-024-00886-5.
[4]	郑庆河, 刘方霖, 余礼苏, 等. 一种结合小波去噪卷积与稀疏Transformer的调制识别方法[J]. 电子与信息学报, 2025, 47(7): 2361–2374. doi: 10.11999/JEIT241159. ZHENG Qinghe, LIU Fanglin, YU Lisu, et al. A modulation recognition method combining wavelet denoising convolution and sparse Transformer[J]. Journal of Electronics & Information Technology, 2025, 47(7): 2361–2374. doi: 10.11999/JEIT241159.
[5]	O’SHEA T J, CORGAN J, and CLANCY T C. Convolutional radio modulation recognition networks[C]. Proceedings of the 17th International Conference on Engineering Applications of Neural Networks, Aberdeen, UK, 2016: 213–226. doi: 10.1007/978-3-319-44188-7_16.
[6]	MURAT A A and KIRAN M S. A comprehensive review on YOLO versions for object detection[J]. Engineering Science and Technology, an International Journal, 2025, 70: 102161. doi: 10.1016/j.jestch.2025.102161.
[7]	周鑫, 何晓新, 郑昌文. 基于图像深度学习的无线电信号识别[J]. 通信学报, 2019, 40(7): 114–125. doi: 10.11959/j.issn.1000-436x.2019167. ZHOU Xin, HE Xiaoxin, and ZHENG Changwen. Radio signal recognition based on image deep learning[J]. Journal on Communications, 2019, 40(7): 114–125. doi: 10.11959/j.issn.1000-436x.2019167.
[8]	李松璋. 基于深度学习的宽带信号识别技术研究[D]. [硕士论文], 北京交通大学, 2023. doi: 10.26944/d.cnki.gbfju.2023.001608. LI Songzhang. Research on broadband radio signal recognition based on deep learning[D]. [Master dissertation], Beijing Jiaotong University, 2023. doi: 10.26944/d.cnki.gbfju.2023.001608.
[9]	张红. 基于改进YOLOv5的宽带接收信号检测与调制识别研究[D]. [硕士论文], 西京学院, 2024. doi: 10.27831/d.cnki.gxjxy.2024.000125. ZHANG Hong. Research on detection and modulation recognition of wideband received signal based on improved YOLOv5[D]. [Master dissertation], Xijing University, 2024. doi: 10.27831/d.cnki.gxjxy.2024.000125.
[10]	CHEN Hui, CHEN Kai, DING Guiguang, et al. Yolov10: Real-time end-to-end object detection[C]. Proceedings of the 38th Conference on Neural Information Processing Systems, Vancouver, Canada, 2024: 107984–108011. doi: 10.52202/079017-3429.
[11]	ZOU Yukang and LIU S. Small object detection algorithm based on improved YOLOv10 for traffic sign[J]. Transportation Research Interdisciplinary Perspectives, 2025, 32: 101501. doi: 10.1016/j.trip.2025.101501.
[12]	HOU Sizu and WANG Wenyao. Fault-line selection method in active distribution networks based on improved multivariate variational mode decomposition and lightweight YOLOv10 network[J]. Energies, 2024, 17(19): 4958. doi: 10.3390/en17194958.
[13]	MEI Junhui and ZHU Wenqiu. BGF-YOLOv10: Small object detection algorithm from unmanned aerial vehicle perspective based on improved YOLOv10[J]. Sensors, 2024, 24(21): 6911. doi: 10.3390/s24216911.
[14]	CHEN Jierun, KAO S H, HE Hao, et al. Run, don't walk: Chasing higher FLOPS for faster neural networks[C]. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, Canada, 2023: 12021–12031. doi: 10.1109/CVPR52729.2023.01157.
[15]	ZHAO Yi’an, LV Wenyu, XU Shangliang, et al. DETRs beat YOLOs on real-time object detection[C]. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2024: 16965–16974.
[16]	WANG C Y, LIAO H Y M, WU Y H, et al. CSPNet: a new backbone that can enhance learning capability of CNN[C]. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, Seattle, USA, 2020: 390–391. doi: 10.1109/CVPRW50498.2020.00203.
[17]	HUYNH-THE T, HUA C H, PHAM Q V, et al. MCNet: An efficient CNN architecture for robust automatic modulation classification[J]. IEEE Communications Letters, 2020, 24(4): 811–815. doi: 10.1109/LCOMM.2020.2968030.
[18]	RAJENDRAN S, MEERT W, GIUSTINIANO D, et al. Deep learning models for wireless signal classification with distributed low-cost spectrum sensors[J]. IEEE Transactions on Cognitive Communications and Networking, 2018, 4(3): 433–445. doi: 10.1109/TCCN.2018.2835460.
[19]	ZHANG Fuxin, LUO Chunbo, XU Jialang, et al. An efficient deep learning model for automatic modulation recognition based on parameter estimation and transformation[J]. IEEE Communications Letters, 2021, 25(10): 3287–3290. doi: 10.1109/LCOMM.2021.3102656.
[20]	XU Jialang, LUO Chunbo, PARR G, et al. A spatiotemporal multi-channel learning framework for automatic modulation recognition[J]. IEEE Wireless Communications Letters, 2020, 9(10): 1629–1632. doi: 10.1109/LWC.2020.2999453.
[21]	ZHANG Jiawei, WANG Tiantian, FENG Zhixi, et al. AMC-Net: An effective network for automatic modulation classification[C]. 2023 IEEE International Conference on Acoustics, Speech and Signal Processing, Rhodes Island, Greece, 2023: 1–5. doi: 10.1109/ICASSP49357.2023.10097070.
[22]	CHEN Yantao, DONG Binhong, LIU Cuiting, et al. Abandon locality: Frame-wise embedding aided transformer for automatic modulation recognition[J]. IEEE Communications Letters, 2023, 27(1): 327–331. doi: 10.1109/LCOMM.2022.3213523.
[23]	王祯, 刘伟, 卢万杰, 等. 面向低信噪比序列的多模态联合自动调制方式识别方法[J]. 电子与信息学报, 2025, 47(12): 5082–5093. doi: 10.11999/JEIT250594. WANG Zhen, LIU Wei, LU Wanjie, et al. Multi-modal joint automatic modulation recognition method towards low SNR sequences[J]. Journal of Electronics & Information Technology, 2025, 47(12): 5082–5093. doi: 10.11999/JEIT250594.
[24]	曹洁, 刘锦辉, 宋蓓蓓. 一种轻量化的IQ路信号调制方式识别方法[J]. 无线通信技术, 2024, 33(3): 18–23. doi: 10.3969/j.issn.1003-8329.2024.03.004. CAO Jie, LIU Jinhui, and SONG Beibei. A lightweight IQ signal modulation recognition method[J]. Wireless Communication Technology, 2024, 33(3): 18–23. doi: 10.3969/j.issn.1003-8329.2024.03.004.