基于时序卷积网络与双向长短期记忆网络融合的电磁信号调制识别算法

黄敏; 王雅琪; 马立云; 王玉明

doi:10.11884/HPLPB202537.250126

基于时序卷积网络与双向长短期记忆网络融合的电磁信号调制识别算法

doi: 10.11884/HPLPB202537.250126

1.
河北科技大学信息科学与工程学院，石家庄 050018
2.
陆军工程大学石家庄校区，石家庄 050003

基金项目: 电磁环境效应重点实验室基础科研计划项目(JCKYS2023DC02); 电磁环境效应重点实验室基金项目（6142205240201）

详细信息

作者简介:
黄　敏，huangmin@hebust.edu.cn

通讯作者:
王玉明，wangyuming@aeu.edu.cn。

中图分类号: TP181
计量
- 文章访问数: 60
- HTML全文浏览量: 11
- PDF下载量: 2
- 被引次数: 0
出版历程
- 收稿日期: 2025-05-12
- 修回日期: 2025-08-26
- 录用日期: 2025-08-21
- 网络出版日期: 2025-09-06

Electromagnetic signal modulation recognition algorithm based on fusion of temporal convolutional network and bidirectional long-short-term memory network

Huang Min^1
,,
Wang Yaqi¹,
Ma Liyun²,
Wang Yuming^{2
, ,}

1.
School of Information Science and Engineering, Hebei University of Science and Technology, Shijiazhuang 050018, China
2.
Shijiazhuang Campus, Army Engineering University, Shijiazhuang 050003, China

摘要

摘要: 针对电磁频谱环境复杂多域这一问题，提出了基于时序卷积网络与双向长短期记忆网络融合的电磁信号调制识别算法。首先，设计了双向长短期记忆网络（Bi-LSTM）来捕捉时序数据的双向依赖关系，提升对复杂调制模式的判别能力；其次，将时序卷积网络（TCN）与Bi-LSTM通过级联架构进行融合，实现了分层时序特征提取与双向动态建模；最后，加入改进的局部敏感哈希注意力机制（LSH Attention），降低注意力矩阵复杂度的同时提高识别的精准度。数据预处理方面提出了一种KNN-BH处理方法，能够提高频谱特征的提取精度。在RML2016.10a数据集上的实验结果表明，相较7个对比算法，TCN-LSTM-LSH Attention算法的识别效果最佳，其对11类信号调制的整体识别准确率达到64.71%。证实了该算法在电磁频谱应用中的潜能与价值。
- 电磁频谱 /
- 调制识别 /
- 注意力机制 /
- TCN /
- Bi-LSTM
Abstract: Background
With the increasing complexity of electromagnetic spectrum, efficient signal modulation identification algorithm is beneficial to electromagnetic spectrum management, which is very important for modern communication system. However, traditional algorithms have limitations in feature extraction and lack of accuracy.
Purpose
This paper presents an improved algorithm for electromagnetic signal modulation recognition. This algorithm integrates the temporal convolution network (TCN), the bidirectional long-short-term memory (Bi-LSTM) network, and the improved locality-sensitive hashing attention mechanism (LSH Attention) to enhance the accuracy of recognition.
Methods
Firstly, Bi-LSTM is designed to capture the bidirectional dependency of time series data and enhance the discrimination ability for complex modulation modes. Secondly, TCN and Bi-LSTM are fused through a cascaded architecture to achieve hierarchical time series feature extraction and bidirectional dynamic modeling. Finally, LSH Attention is added to reduce the complexity of the attention matrix while improving the recognition accuracy. In terms of data preprocessing, a KNN-BH processing method is proposed, which can improve the extraction accuracy of spectral features.
Results
Experimental results on the RML2016.10a dataset show that compared with seven baseline algorithms, the TCN-LSTM-LSH Attention algorithm has the best recognition performance, with an overall recognition accuracy of 64.71% for 11 types of signal modulations.
Conclusions
This algorithm demonstrates great potential in electromagnetic spectrum applications and is highly suitable for use in high-precision modulation recognition tasks in communication systems.
- electromagnetic spectrum /
- modulation recognition /
- attention mechanism /
- temporal convolutional network /
- bidirectional long-short-term memory

HTML全文

图 1 因果卷积结构

Figure 1. Structure of causal convolution

下载: 全尺寸图片幻灯片

图 2 感受野扩大流程

Figure 2. Receptive field dilation process

下载: 全尺寸图片幻灯片

图 3 Bi-LSTM算法流程

Figure 3. Bi-LSTM algorithm flow

下载: 全尺寸图片幻灯片

图 4 LSH Attention原理流程

Figure 4. Principle process of LSH Attention

下载: 全尺寸图片幻灯片

图 5 哈希函数原理

Figure 5. Hash function principle

下载: 全尺寸图片幻灯片

图 6 RevNet主要结构

Figure 6. Main structure of RevNet

下载: 全尺寸图片幻灯片

图 7 TCN-LSTM-LSH Attention算法原理图

Figure 7. The schematic diagram of TCN-LSTM-LSH Attention

下载: 全尺寸图片幻灯片

图 8 加窗前后频谱泄漏对比图

Figure 8. Comparison chart of spectrum leakage before and after windowing

下载: 全尺寸图片幻灯片

图 9 实验流程图

Figure 9. Experimental flow chart

下载: 全尺寸图片幻灯片

图 10 各个算法对电磁信号调制识别的概率以及召回率

Figure 10. Accuracy of electromagnetic signal modulation recognition and recall by each algorithm

下载: 全尺寸图片幻灯片

表 1 传统注意力机制与LSH Attention对比

Table 1. Comparison of traditional Attention and LSH Attention

characteristic	traditional attention	LSH Attention
time complexity	O(L²)	O(LlogL)
space complexity	O(L²)	O(L)

下载: 导出CSV

表 2 对比实验数据

Table 2. Comparison of experimental data

algorithm	overall accuracy/%	highest accuracy/%	inference time/ms	recall/%
CNN	51.74	81.89	251	51.66
LSTM	58.1	87.59	368	58.28
TCN	59.2	90.23	279	69.43
LSTM-LSH attention	59.6	90.6	506	60.08
CNN-LSTM	60.48	91.08	439	60.42
CNN-LSTM-LSH attention	61	91.54	597	61.15
TCN-LSTM	62.25	93.16	572	62.65
TCN-LSTM-LSH attention	64.71	96.46	603	64.81

下载: 导出CSV

参考文献(16)

[1]	阮天宸, 吴启晖, 赵世瑾, 等. 认知学习: 电磁频谱空间机器学习新范式[J]. 电子学报, 2023, 51(6): 1430-1442 Ruan Tianchen, Wu Qihui, Zhao Shijin, et al. Cognitive learning: a new paradigm for machine learning in electromagnetic spectrum environment[J]. Acta Electronica Sinica, 2023, 51(6): 1430-1442
[2]	卢河旭, 陈丹伟, 罗圣美. 基于YOLOv5架构改进的电磁信号目标识别方法[J]. 计算机测量与控制, 2024, 32(11): 243-250,294 Lu Hexu, Chen Danwei, Luo Shengmei. An improved recognition method for electromagnetic signal targets based on YOLOv5 architecture[J]. Computer Measurement & Control, 2024, 32(11): 243-250,294
[3]	陆欢, 周顺勇, 彭梓洋, 等. 无线电信号的调制识别研究进展[J]. 四川轻化工大学学报(自然科学版), 2024, 37(6): 31-42 Lu Huan, Zhou Shunyong, Peng Ziyang, et al. Research progress on modulation identification of radio signals[J]. Journal of Sichuan University of Science & Engineering (Natural Science Edition), 2024, 37(6): 31-42
[4]	Weaver C S, Cole C A, Krumland R B, et al. The automatic classification of modulation types by pattern recognition[R]. Palo Alto: Defense Technical Information Center, 1969.
[5]	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
[6]	Yang Cheng, He Zhimin, Peng Yang, et al. Deep learning aided method for automatic modulation recognition[J]. IEEE Access, 2019, 7: 109063-109068. doi: 10.1109/ACCESS.2019.2933448
[7]	秦博伟, 蒋磊, 许华, 等. 基于残差生成对抗网络的调制识别算法[J]. 系统工程与电子技术, 2022, 44(6): 2019-2026 doi: 10.12305/j.issn.1001-506X.2022.06.30 Qin Bowei, Jiang Lei, Xu Hua, et al. Modulation recognition algorithm based on residual generation adversarial network[J]. Systems Engineering and Electronics, 2022, 44(6): 2019-2026 doi: 10.12305/j.issn.1001-506X.2022.06.30
[8]	张航, 吴泓霖, 余勤, 等. 基于多特征信息的深度学习网络调制识别算法[J]. 计算机工程与设计, 2022, 43(10): 2762-2768 Zhang Hang, Wu Honglin, Yu Qin, et al. Modulation recognition algorithm for multiple feature information based on deep learning[J]. Computer Engineering and Design, 2022, 43(10): 2762-2768
[9]	Ding Yixin. Automatic modulation recognition of communication signal based on wavelet transform combined with singular value and NCA-CNN[C]//Wireless and Optical Communications Conference (WOCC). 2023: 1-6.
[10]	Zhang Bojun. Multimodal machine learning algorithm for enhanced signal modulation recognition in wireless communication systems[C]//IEEE 49th Conference on Local Computer Networks (LCN). 2024: 1-6.
[11]	Lea C, Flynn M D, Vidal R, et al. Temporal convolutional networks for action segmentation and detection[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition. 2017: 1003-1012.
[12]	郝建华. 基于深度学习的时间序列预测算法研究[D]. 济南: 山东师范大学, 2024: 89-91 Hao Jianhua. Research on time series forecasting algorithms based on deep learning[D]. Ji’nan: Shandong Normal University, 2024: 89-91
[13]	Qiao Junbo, Han Suai, Sun Zeyang. Modulated recognition and function realization based on deep learning[C]//IEEE International Symposium on Broadband Multimedia Systems and Broadcasting (BMSB). 2023: 1-6.
[14]	刘佳旭, 白再冉, 张艳菊. 基于BiLSTM-Attention的议论文篇章要素识别[J]. 计算机系统应用, 2025, 34(5): 202-211 Liu Jiaxu, Bai Zairan, Zhang Yanju. Discourse elements identification in argumentative essays based on BiLSTM-attention[J]. Computer Systems & Applications, 2025, 34(5): 202-211
[15]	胡国乐, 李鹏, 林事力, 等. 基于相位变换和CNN-BiLSTM的自动调制识别算法[J]. 电讯技术, 2024, 64(11): 1780-1787 Hu Guole, Li Peng, Lin Shili, et al. An automatic modulation recognition algorithm based on phase transformation and CNN-BiLSTM[J]. Telecommunication Engineering, 2024, 64(11): 1780-1787
[16]	房虹辰. 基于LSTM和CNN的频谱感知方法研究[D]. 银川: 宁夏大学, 2022: 23-25 Fang Hongchen. Research on spectrum sensing method based on LSTM and CNN[D]. Yinchuan: Ningxia University, 2022: 23-25