Radio Signal Modulation Pattern Recognition Based on Time-Frequency Adaptive Decomposition and Hybrid Neural Network

Automatic modulation classification of electromagnetic signals has an important role in the field of signal processing. The current methods for identifying the modulation pattern of radio signals mainly used end-to-end neural network models, which suffered from the problems such as network redundancy and non-interpretability. In order to solve the above problems, this paper proposes a radio signal modulation pattern recognition method based on time-frequency adaptive decomposition and hybrid neural network. The discrete wavelet decomposition is used to decompose the original radio signals into adaptive high and low frequencies. Then feed the resulting high and low frequency portions into a multi-channel hybrid neural network (MHNN) for training to obtain the classification results. In the discrete wavelet decomposition process, the number of decomposition layers and the decomposition threshold are determined adaptively according to the difference of SNRs. In the design of multi-path hybrid neural network structure, the signal characteristics of high and low frequency parts are considered comprehensively. The network channels with different convolutional kernel scales and different layers are constructed to reduce the number of model parameters and training time while ensuring the classification accuracy. Using RadioML2016.10a dataset for experiments, the recognition accuracy of MHNN is prior to the end-to-end deep learning methods, which can achieve the accuracy of 92.30% and the F1 of 91.84% when SNR=2dB. It shows that the proposed radio signal modulation pattern recognition method based on the time-frequency adaptive decomposition and hybrid neural network can balance the accuracy and lightweight. Moreover, it has certain interpretability for the original input data, distinguishing the main body and details of the signal by dividing the signal into high and low frequencies, which provides a new idea for the interpretability of deep convolutional neural networks.