Modulation recognition method based on multiscale convolutional fusion coding networks

Abstract

To address the issue of insufficient feature extraction in existing modulation recognition methods that limited classification accuracy, a Transformer-based modulation recognition method was proposed. Convolutional kernels of varying sizes were employed to enhance multi-scale signal feature extraction, followed by feature fusion to strengthen learning capability while reducing computational demands. A multi-head self-attention mechanism was utilized to enable parallel processing for capturing diverse signal characteristics. A dual-branch multilayer perceptron structure was introduced to further improve adaptability and diversity learning while accelerating operational speed. Experimental results demonstrated the model’s robust stability and generalization capability, showing minimal performance variation under different test batch sizes with fixed training batches. On the RML2018.01A dataset, the proposed model achieves over 96% classification accuracy at 10 dB.

Keywords

• convolutional neural network
• modulation classification
• Transformer
• multi-scale fusion
• multilayer perceptron