Research on Automatic Modulation Recognition Method Based on Deep Learning

In non-cooperative communication systems, Automatic Modulation Recognition (AMR) is a key technology for spectrum sensing, spectrum monitoring and spectrum utilization. Most traditional AMR methods ignore the diversity and intrinsic connections of features, and also face the challenges of low recognition rate and weak generalization ability. In this paper, we propose a parallel neural network combining skip-connected Convolutional Neural Network (CNN) with Gated Recurrent Unit (GRU) to extract spatiotemporal features in parallel from both the In-phase Quadrature (IQ) and Amplitude Phase (AP) components of the signal. The proposed network uses skip-connection structures to effectively reduce the problems of gradient vanishing and network degradation. GRU reduces the computational parameter count while retaining the advantages of Long-Short Term Memory (LSTM) networks. In addition, based on the interdependence between different feature channels, we introduce a lightweight and efficient channel attention network to re-weight all features, further improving the signal recognition rate. The results show that the proposed network achieves a recognition rate of 90% in environments where the Signal to Noise Ratio (SNR) is above 4 dB, and can effectively improve the recognition rate in low SNR conditions. Compared with two other parallel neural networks, the dual-stream CNN and dual-stream CNN-LSTM, our network can achieve higher recognition accuracy with lower complexity.