An optimised multi-channel neural network model based on CLDNN for automatic modulation recognition

Abstract

To achieve high accuracy blind modulation identification of wireless communication, a novel multi-channel deep learning framework based on the Convolutional Long Short-Term Memory Fully Connected Deep Neural Network (MC-CLDNN) is proposed. To make network training more efficient, we use the Gated Recurrent Unit (GRU) sequence model as the substructure. Furthermore, the skip connection is added to alleviate the problem of gradient disappearance in the network training and reduce the negative effect of pooling layer processes time series data on the subsequent sequence model. We test the feasibility of the model based on two open-source data sets RadioML2016.10a and RadioML2016.10b. The simulation results show that the proposed model can identify most modulation modes efficiently under the influence of various factors such as Additive White Gaussian Noise (AWGN), multipath fading, frequency offset. In the Signal-to-Noise Ratio (SNR) range of 0-18 dB, the overall recognition accuracy of the MC-CLDNN can reach 93% and the area under the Receiver Operating Characteristic (ROC) curve accounts for more than 99%. Therefore, the model has the characteristics of high recognition accuracy and strong generalisation ability. Its comprehensive performance is better than most of the existing deep learning models.