Modulation classification based on the collaboration of dual-channel CNN-LSTM and residual network

作     者：Li Hui Li Shanshan Zou Borong Chen Yannan Li Hui;Li Shanshan;Zou Borong;Chen Yannan

作者机构：School of Physics and Electronic Information EngineeringHenan Polytechnic UniversityJiaozuo 454003China 

出 版 物：《The Journal of China Universities of Posts and Telecommunications》 (中国邮电高校学报（英文版）)

年 卷 期：2022年第29卷第1期

页      面：113-124页

核心收录：

学科分类：0809[工学-电子科学与技术（可授工学、理学学位）] 080902[工学-电路与系统] 08[工学] 081104[工学-模式识别与智能系统] 0811[工学-控制科学与工程] 

基　　金：supported by the Project of Henan Science and Technology Research (222102210247)。 

主　　题：convolutional neural network deep neural network long short-term memory modulation classification residual network 

摘      要：Deep learning has recently been progressively introduced into the field of modulation classification due to its wide application in image, vision, and other areas. Modulation classification is not only the priority of cognitive radio and spectrum sensing, but also the link during signal demodulation. Combining the advantages of convolutional neural network(CNN), long short-term memory(LSTM), and residual network(ResNet), a modulation classification method based on dual-channel CNN-LSTM and ResNet is proposed to automatically classify the modulation signal more accurately. Specifically, CNN and LSTM are initially used to form a dual-channel structure to effectively explore the spatial and temporal features of the original complex signal. It solves the problem of only focusing on temporal or spatial aspects, and increases the diversity of features. Secondly, the features extracted from CNN and LSTM are fused, making the extracted features richer and conducive to signal classification. In addition, a convolutional layer is added within the residual unit to deepen the network depth. As a result, more representative features are extracted, improving the classification performance. Finally, simulation results on the radio machine learning(RadioML) 2018.01 A dataset signify that the network’s classification performance is superior to many classifiers in the literature.