MSGNet: A Multi-feature Lightweight Learning Network for Automatic Modulation Recognition

Automatic modulation recognition (AMR) has significant applications in communication system optimization, spectrum management, signal identification and classification, and security and protection. However, most of the existing AMR models are too large in terms of parameter number and computational complexity. Therefore, this letter proposes a multi-channel deepthwise separable convolution and gated recurrent unit network (MSGNet), which has a better recognition performance in the case of a lower number of parameters. MSGNet uses light weight modules: deepthwise separable convolution and gated recurrent unit, from the three input data channels to enter successively separated amplitude/phase samples, in-phase/quadrature-phase samples, and the real/imaginary samples obtained by Fourier transform of the signals, from which to obtain the signal space, time and frequency features. To speed up the training of the model and prevent overfitting, the batch normalization layer and dropout algorithm are added to the MSGNet model. Simulation results on the benchmark dataset indicate that the proposed model has a low number of parameters and computational complexity and a high level of recognition accuracy.