Enhancing Robustness in Automatic Modulation Classification via Energy-Guided Multi-Scale

Abstract Automatic Modulation Classification (AMC) is pivotal in non-cooperative com munication systems. This study addresses the challenges of feature coupling and noise-induced degradation, particularly under low signal-to-noise ratio (SNR) conditions. We introduce the Energy-Guided Multi-Scale TemporalNetwork (EMST-Net), which employs a Multi-Stream Fusion Convolution module to decouple in-phase and quadrature components, enhancing feature expressive ness. An Energy-Guided Temporal Saliency Attention mechanism adaptively emphasizes informative signal segments while suppressing noise. A gateddilated temporal convolutional network captures both short-term and long-range depen dencies. Extensive experiments on three public datasets (RML22, RML2016.10a, and RML2018.01a) demonstrate that EMST-Net achieves overall accuracies of 71.59%, 62.79%, and 61.74%, respectively, outperformingcomparative models. Notably, it exhibits strong robustness under low SNR conditions (e.g., 86.27% accuracy at 0 dB on RML22). The model is computationally efficient, with only 260K parameters, 24.2M FLOPs, and a storage overhead of 0.811MB. The source code is available at: https://github.com/jiahuazhou123-cpu/EMST-Net