A Reference Signal-Aided Deep Learning Approach for Overlapped Signals Automatic Modulation Classification
Rui Zhang, Yanlong Zhao, Zhendong Yin, Dasen Li, Zhilu Wu
Abstract
Traditional likelihood-based and handcrafted feature-based methods for overlapped signals automatic modulation classification (OS-AMC) suffer from the uncertainty of the overlapped numbers in practical application scenarios, while existing deep learning methods still require a complex training process. In this letter, a deep learning approach with a hybrid network combining ConvNeXt and atrous self-attention transformer is proposed to solve this problem. Specifically, a reference signal-aided training is introduced to generate the decision threshold of the proposed network automatically, which omits the searching process of the decision threshold and makes the training process more efficient. The simulation results indicate that the proposed method can achieve superior classification accuracy with a simpler training process and lower computational complexity and memory cost.