Automatic Modulation Recognition Based on Spatio-temporal Features Fusion

Automatic Modulation Recognition (AMR) has substantial implications in civilian and military communications. Lately, the neural network model has been increasingly used for AMR. However, the previous models have ignored the spatio-temporal characteristics of the signals, which leads to insufficient classification accuracy under low signal-to-noise ratio, especially for certain specific modulation types, like MQAM. Combining the advantages of the complex-valued convolutional neural networks and the long short-term memory networks (LSTM), this paper proposes a novel AMR method based on spatio-temporal features fusion with complex convolution modules and LSTMs, denoted as bi-channel complex-convolutional neural networks with LSTM (BCLNN). BCLNN efficiently explores the spatial-temporal properties of raw complex signals and achieves higher classification accuracy for AMR, especially under a low signal-to-noise ratio and for MQAM modulation. Experimental results dedicate that when the SNR is between −20dB and −10dB, the average accuracy of BCLNN is at least 7.9% higher than that of other networks. BCLNN has a more than 12% improvement of average recognition rate for 16QAM and 64QAM when the SNR is between −20dB and 18dB.