MST: A Multi-Scale Transformer Framework With Cross-Scale Token Fusion for Automatic Modulation Recognition

Automatic modulation recognition which is a core enabling technology for cognitive radio systems, faces considerable challenges in robust feature extraction under non-ideal environments. Under low signal-to-noise ratio (SNR) conditions and modulation parameter mismatches, conventional deep learning-based methods exhibit markedly degraded discriminative feature representation capability. This letter presents a multi-scale Transformer (MST) framework that significantly enhances feature extraction from raw signal sequences through parallel multi-resolution analysis. We decompose signals into three complementary scales using varying convolutional kernel sizes, enabling comprehensive analysis of global features, local characteristics, and transient details across amplitude, phase, and frequency domains. Subsequently, an effective cross-scale token fusion mechanism is developed to integrate information across different scales. Furthermore, we employ knowledge distillation to compress the model, nearly halving its size while maintaining performance. Experimental results show that under challenging conditions combining low SNR, large carrier frequency offsets, and significant sampling rate offsets, the proposed MST framework demonstrates superior performance compared to other state-of-the-art methods.