Enhanced Automatic Modulation Recognition Using Polar Coordinate Features and Multi-Teacher Knowledge Distillation for Lightweight Networks

Automatic Modulation Recognition (AMR) is crucial for spectrum monitoring and cognitive radio applications. Deep learning methods have been widely adopted for AMR tasks due to their powerful feature learning capabilities. However, existing deep learning approaches face challenges in computational complexity, model deployment, and feature representation. This paper designs corresponding methods to address AMR challenges to a certain extent. First, we propose grid-based cumulative features in polar coordinate space that enhance inter-class separability and provide noise resistance performance. Second, we design a lightweight student network architecture utilizing depthwise separable convolutions and Convolutional Block Attention Module (CBAM) attention mechanisms to reduce computational parameters while maintaining performance. Third, we introduce an enhanced knowledge distillation framework employing multi-teacher feature-level distillation that replaces traditional soft labels with high-dimensional feature vectors for more granular knowledge transfer. Experimental validation on the RadioML2016.10A dataset demonstrates that our proposed method reduces the loss caused by lightweighting while decreasing model complexity, particularly under low Signal-to-Noise Ratio(SNR) conditions.