基于时序深度残差收缩网络的混叠信号调制识别方法  

Mixed signal modulation recognition method based on temporal depth residual shrinkage network

作　　者：刘京华 魏祥麟 范建华 胡永扬 王晓波 于兵[1] LIU Jinghua;WEI Xianglin;FAN Jianhua;HU Yongyang;WANG Xiaobo;YU Bing(School of Electronic and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China;The 63rd Research Institute,National University of Defense Technology,Nanjing 210007,China)

机构地区：[1]南京信息工程大学电子与信息工程学院,江苏南京210044 [2]国防科技大学第六十三研究所,江苏南京210007

出　　处：《电信科学》2024年第10期27-38,共12页Telecommunications Science

摘　　要：基于深度学习进行信号自动调制识别在分类精度、可迁移性等方面普遍优于传统方法,引起广泛关注。但是,当前方法多数针对单信号样本进行识别,无法适用于混叠信号识别场景。针对该问题,对混叠信号调制识别方法进行了研究,结合长短期记忆(long short term memory,LSTM)网络和深度残差收缩网络(deep residual shrinkage network,DRSN),设计了时序深度残差收缩网络模型,其中包含残差模块、收缩模块和LSTM模块。残差模块和收缩模块负责提取混叠信号中的显著信息并自适应生成决策阈值,LSTM模块用于提取混叠信号中的时序隐含特征。三者结合可以有效提高混叠信号的识别精度。公开和实测数据集测试结果表明,所提方法识别精度优于5种典型方法,在高信噪比下的平均识别分类准确率可以达到92.7%;21种混叠信号中有12种识别准确率接近100%。Deep learning-based automatic signal modulation recognition has generally outperformed traditional methods in terms of classification accuracy and transferability,garnering widespread attention.However,most existing methods are designed to recognize single signal samples and are not applicable to recognize scenarios involving overlapping signals.To address this limitation,a modulation recognition method for aliased signals was investigated and a temporal deep residual shrinkage network model by integrating LSTM and DRSN was developed.There were three key modules in the model:a residual module,a shrinkage module,and a LSTM module.Salient information from overlapping signals was extracted by the residual module and the shrinkage module and decision thresholds were adaptively generated,while the LSTM module is tasked with extracting temporal hidden signals within the aliased data.The recognition accuracy of aliased signals was enhanced by the combination of these modules significantly.Testing on both public and private datasets demonstrates that the proposed method outperforms five state-of-the-art approaches,achieving an average recognition and classification accuracy of 92.7%under high signal-to-noise ratio conditions.Notably,the recognition accuracy for 12 out of 21 types of aliased signals approaches 100%.

关 键 词：调制识别 混叠信号 深度残差收缩网络 深度学习 

分 类 号：TN929.5[电子电信—通信与信息系统]