基于时频Grad-CAM的调制识别网络可解释分析  

Interpretability of Modulation Recognition Network Based on Time-Frequency Gradient-Weighted Class Activation Mapping

作　　者：梁先明[1] 倪帆 陈文洁 张家树[3] LIANG Xianming;NI Fan;CHEN Wenjie;ZHANG Jiashu(The 10th Research Institute of China Electronics Technology Group Corporation,Chengdu 610036,China;School of Information Science and Technology,Southwest Jiaotong University,Chengdu 610031,China;School of Computer and Artificial Intelligence,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]中国电子科技集团公司第十研究所,四川成都610036 [2]西南交通大学信息科学与技术学院,四川成都610031 [3]西南交通大学计算机与人工智能学院,四川成都610031

出　　处：《西南交通大学学报》2024年第5期1215-1224,共10页Journal of Southwest Jiaotong University

基　　金：国家自然科学基金项目(62071396);四川省自然科学基金项目(2022NSFSC0531)。

摘　　要：针对时频深度学习调制识别方法存在可解释性差的问题,提出一种基于时频梯度加权类激活映射(GradCAM)的调制识别网络可解释框架.该框架通过时频Grad-CAM可视化深度模型中隐含层的关键特征,从视觉上解释网络隐含层提取的时频深度特征对于正确与错误识别中的作用,揭示低信噪比环境下网络性能下降的内在机理,并通过量化和排序网络中每层不同卷积核的贡献值来判断网络的冗余程度.仿真实验结果验证了基于时频Grad-CAM的调制识别网络可解释性框架的有效性;可解释分析结果表明,在低信噪比环境下,网络特征提取区域有大量噪声存在,且本文所测试的调制识别网络冗余程度较为严重.Aiming at the poor interpretability of modulation recognition methods based on time-frequency deep learning,an interpretable framework of a modulation recognition network is proposed,utilizing time-frequency gradient-weighted class activation mapping(Grad-CAM).Through the key features of the hidden layer in the Grad-CAM visual deep model,the significance of the deep features extracted from the network hidden layer are illustrated in terms of correct and error recognition,revealing the decline of network performance in the environment of low signal-to-noise ratio(SNR).The contribution values of different convolution cores at each network layer are quantified and sorted to determine the network redundancy.The simulation results verify the interpretable framework of the time-frequency deep learning network for modulation recognition.The interpretable analysis results reflect that there is a large amount of noise present in the feature extraction region of the network in a low signal-to-noise ratio environment,and the tested modulation recognition network exhibits a high degree of redundancy.

关 键 词：可解释深度学习 梯度类加权激活映射 调制识别 时频分析 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]