Signal Modulation Recognition based on Convolutional Autoencoder and Time-Frequency Analysis

With the development of communication technology, patterns of communication signals have become more complex and diverse. Therefore, the technology of identifying the modulation mode of the communication signal in transmission, especially the modulation recognition technology based on artificial intelligence, has become an extremely important technology in the communication field. For a variety of modulation methods, the traditional method is extremely complicated to implement, and cannot meet the requirements of accurate identification in a short time. In order to increase the speed and reduce the redundancy, this paper proposes a method based on the convolutional autoencoder and the residual network which can realize the denoising, identification and classification of different modulated signals. This method generates ten different modulation types of signals under each signal-to-noise ratio. After the model is trained, the data set is input to the convolutional autoencoder to denoise, and then the data set denoised by the autoencoder is input to the residual network to obtain the classification and recognition accuracy of each modulation type. And an average recognition rate of 92.86% was achieved at -2dB.