基于轻量级残差网络的信号调制识别研究  

Research on signal modulation recognition based on lightweight residual networks

作　　者：张承畅[1] 王艺培 李吉利 罗元[1] ZHANG Chengchang;WANG Yipei;LI Jili;LUO Yuan(School of Optoelectronic Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学光电工程学院,重庆400065

出　　处：《实验技术与管理》2025年第3期114-122,共9页Experimental Technology and Management

基　　金：重庆市研究生教学改革项目(yjg222025,YKCSZ23112);重庆市教育教学改革研究项目(213153);重庆邮电大学教育教学改革重点项目(XJG20103)。

摘　　要：针对高复杂度的神经网络难以被部署在对低延迟和存储有严格要求的场景和接收设备中的问题,该文提出了一种基于轻量级残差网络的自动调制识别(AMR)框架。该框架将蓝图可分离卷积(BSConv)与CoordGate相结合以实现轻量化的设计。为了弥补轻量化设计造成的性能损失,该文提出了使用改进的基于软池化(SoftPool)的卷积注意力模块(CBAM)以提升模型的泛化能力和分类性能。实验结果表明,该文提出的轻量级AMR框架在性能提升的情况下参数量大幅减少,平均识别准确率为98.23%,参数量为87057。[Objective]Automatic modulation recognition(AMR)is a key technology in cognitive radio systems,effectively addressing the challenges of spectrum utilization efficiency by identifying signal modulation types.However,traditional AMR models typically rely on deep learning architectures that are computationally intensive,which limits their applicability in resource-constrained environments.This study aims to develop a lightweight residual neural network optimized for AMR tasks,striking a balance between performance and computational efficiency.[Methods]The proposed lightweight AMR network integrates the blueprint separable convolution(BSConv)and CoordGate modules to enhance feature extraction and reduce computational demands.BSConv reduces parameter counts by decomposing convolution operations into efficient pointwise and depthwise convolutions,making it particularly suitable for IQ modulation signal analyses.CoordGate,a spatial encoding module,enhances the network’s ability to identify spatially correlated features,thereby improving its robustness against noise and channel variations.Additionally,a modified SoftPool-based convolutional block attention module(CBAM)is introduced to retain critical signal information and mitigate the performance trade-offs of a lightweight design.The model is evaluated on the RadioML2016.10a dataset at a signal-to-noise ratio(SNR)of 2,focusing on modulation types such as BPSK,QPSK,8PSK,16QAM,and 64QAM.The dataset provides realistic conditions for testing,including noise,channel fading,and phase offset,making it an effective benchmark for evaluating AMR frameworks.[Results]Experimental results indicate that the proposed model achieves an average recognition accuracy of 98.23%,surpassing conventional models such as ResNet-50,SepConv,and LSTM by at least 6%in accuracy and 9%in F1-score.The framework reduces its parameter count to 87057,with a model size of only 0.33 MB,making it ideal for deployment in embedded systems and mobile devices.Ablation studies confirm the contributions of the

关 键 词：自动调制识别(AMR) 轻量级神经网络 深度学习 注意力机制 

分 类 号：TN971.1[电子电信—信号与信息处理]