Multi-Channel Convolutional Distilled Transformer for Automatic Modulation Classification
Zhenhua Chen, Xinze Zhang, Kun He · 2024
Automatic modulation classification (AMC) is of great importance in the field of radio, and deep neural networks have achieved promising progress on the AMC task. However, existing research has not fully exploited the diverse features present in radio signals, indicating untapped potential for further enhancement in model performance. To this end, we propose a novel model architecture called multi-channel convolutional distilled Transformer (MCDformer). In MCDformer, we first propose a concise frequency domain denoising module (FDDM) to effectively leverage the frequency-domain information and enhance the denoising process of input signals. Following extracting waveform features using a convolutional neural network backbone, we propose a distilled Transformer module (DTM) to capture time-domain features and learn the correlations among different features. Within the DTM, a distillation layer is introduced to further condense attention maps while reducing computational complexity. By integrating modules that learn features from different aspects, MCDformer can effectively leverage the complementary nature of diverse features, thereby acquiring richer information from the signals and achieving improved classification performance. Experimental results demonstrate that our proposed MCDformer model significantly outperforms existing state-of-the-art AMC models. Code is available at https://github.com/JHL-HUST/MCDformer/.