A Multi-level Complex Feature Mining Method Based on Deep Learning for Automatic Modulation Recognition

The previous modulation recognition models based on deep learning ignore the signal's complex characteristics and only consider the information carried by the signal in a single dimension, resulting in poor performance. Aiming at the complex characteristics of in-phase/quadrature (I/Q) data, this paper adopts a combination of complex convolution and one-dimensional real convolution, emphasizing the feature interaction between I and Q and enriching the feature representation of the signal. Besides, a multi-level complex attention block is introduced to enhance the informative representation of the entire feature space. Experimental results indicate that the proposed method's recognition accuracy of MQAM is significantly improved. Furthermore, the proposed method also alleviates the poor performance under a low signal-to-noise ratio, which is overall better than other deep learning-based modulation recognition models.