Modulation Recognition with Enhanced Constellation Based on Convolutional Neural Network

Modulation recognition is essential for cognitive radio and non-cooperative communications. Recently, convolutional neural network (CNN) has been applied in modulation recognition and achieved great progresses. However, the current methods using constellation as signal features usually ignore the time correlation information between adjacent samples. Motivated by this deficiency, in this paper, we propose a CNN-based modulation recognition method with enhanced constellation. To eliminate the redundancy in the original constellation, only the information representing whether there exist any constellation points is extracted from the original constellation. Then, the phases carrying sample correlation information are superimposed to synthesize the enhanced constellation as the representative features. The enhanced constellation is further improved by incorporating the amplitude information as pixel values. In addition, to adapt to the varying levels of correlation between and within rows in enhanced constellation, we optimize the existing CNN network with the proper convolution kernel. Simulation results of the proposed method verify its effectiveness and robustness in modulation recognition.