Toward Interference-Tolerant Automatic Modulation Recognition via Multi-Stage Feature Extraction Network

Automatic modulation recognition (AMR) is critical for enhancing communication system performance and spectrum utilization in complex environments. However, existing deep learning-based AMR methods face significant challenges under the low signal-to-noise ratio (LSNR) condition, where noise interference severely distorts modulation characteristics, reducing recognition accuracy. To address this issue, multi-stage feature extraction network (MFENet) is proposed, an interference-tolerant AMR model designed to improve LSNR performance. MFENet integrates Fast Fourier Transform and convolutional filters in the Modulation Signal Preprocessing (MSP) module to enhance spectral features. The Multi-stage Feature Extraction (MFE) module, incorporating multi-scale sandglass, feature extraction stem, and Res2-Attention Block, enables deep cross-scale feature fusion and extraction. Finally, the AMR classifier refines these features using channel attention and a multilayer perceptron. Experimental results on the RML2016.10a and RML2016.10b datasets demonstrate that MFENet significantly outperforms state-of-the-art models in recognition accuracy, especially under the LSNR condition.