A Cross-Scale Embedding Based Fusion Transformer for Automatic Modulation Recognition

A novel Cross-Scale Embedding Based Fusion Transformer, called CE-FuFormer for automatic modulation recognition was proposed in this letter. In CE-FuFormer, a cross-scale embedding layer is introduced to convert the input signal into an embedding with multiple scale features. To reduce the computational complexity of the model, we allocate fewer channel dimensions for large convolution kernels. A fusion transformer encoder (FTE) module is proposed for simultaneous local and global information extraction of the input modulated signal features. In addition, we rationally allocate the channel dimensions of the CNN part in the FTE module to reduce the generation of parameters while maintaining network performance. The proposed CE-FuFormer outperforms other state-of-the-art (SOTA) models on the datasets RadioML2016.10a and RadioML2018.01a.