Feature Fusion Convolution-Aided Transformer for Automatic Modulation Recognition

Automatic Modulation Recognition (AMR) is becoming increasingly important due to its key role in wireless communications. In order to enrich the feature information and reduce computational complexity, a novel feature fusion convolution-aided transformer (FCAformer) approach is proposed in this paper. To enhance the differences between samples, we adopt Markov Transformation Field (MTF), which can extract potential correlation features between signal sequences and combine them with I/Q sequences to generate more information rich data as input for the transformer. We propose a novel patchify module that reduces invalid features and computational complexity by performing feature extraction. We also propose a convolution-aided encoder module consisting of an improved attention mechanism and a convolutional residual connection to improve performance through local-global feature fusion and further reduce the parameters. Experimental results on RadioRML2016.10a show that the proposed AMR method has both higher recognition accuracy and lower parameters than other contrastive methods.