Research On Classification And Recognition Of Communication Signals Modulation Mode Based On CNN Architecture

With the rapid development of communication technology in modern society and the complicated communication environment,many different modulation methods have to be adopted for different communication signals in order to improve the utilization of the frequency band and ensure the reliability of transmission.Therefore,how to carry out efficient detection and identification of these signals has been explored and sought by innumerable people in both military and civilian fields.The purpose of the research on the classification of the communication signal modulation mode is to be able to correctly recognize the modulation mode of the received communication signal in the context of simultaneous transmission of multiple modulated signals.In this paper,the advantage of the characteristics of signal cyclic spectrum and the advantages of CNN architecture in classification are studied to identify digital modulated signals.First of all,the commonly used communication signal modulation methods are introduced.Then,the CNN-based modulation recognition architecture and its related theories are introducedwhich mainly introduce artificial neuron,CNN solution,and gradient calculation of convolutional layer and downsampled layer,and also introduce softmax regression.After introducing the theory of cyclic spectrum,the characteristics of the cyclic spectrum theory and its suppression of noise were analyzed.Secondly,according to the related theory of cyclic spectrum,the feature of the modulation mode of communication signal is extracted and analyzed.In this paper,the theory of cyclic spectrum of MPSK and MQAM is deduced and simulated.After that,the cyclic spectrum of the modulated signal is preprocessed,and the original feature set is formed,which is used as the input data of CNN.Finally,the construction of identification system of CNN architecture is completed,and the cyclic spectrum of the modulation mode is taken as the input data of CNN.The cyclic spectrum feature is extracted by the convolution layer and subsampling layer in the CNN architecture so that the modulation signal can be identified effectively in class and between classes.The results of simulation show that the modulation classification algorithm proposed in this paper are superior to other traditional modulation recognition algorithms at low signal-to-noise ratio(SNR),and When SNR is between-5dB and 5dB,the accuracy of modulation pattern recognition can be as high as 92%.