Automatic modulation recognition is an important aspect of wireless communication. In recent years, the rapid development of deep learning technology has provided new solutions for modulation recognition. Deep learning-based modulation recognition has strong feature extraction and classification capabilities, resulting in higher recognition accuracy compared to traditional detection methods. However, the commonly used neural networks currently all face the problem of low recognition accuracy under low signal-to-noise ratio (SNR). To address this issue, this paper proposes a hybrid neural network model based on multi-channel input, which utilizes three channels of input: I/Q signals, time-frequency (T-F) distribution matrix, and signal-to-noise ratio. By incorporating SNR to introduce an environmental perception mechanism and a gated fusion module, the model's ability to understand the features of complex sequence information is enhanced. In addition, a phased block based on the Transformer architecture is employed to learn local and global features by combining tokens of different scales. Experimental results on the open-source dataset RML2016.10A indicate that the proposed method outperforms the current state-of-the-art modulation recognition methods, with an average accuracy of 47.60% at signal-to-noise ratios from -20dB to 0dB, and an overall average recognition accuracy of 68.52%.
Xiaoting Wang (Mon,) studied this question.