Rapid radar work mode recognition based on knowledge distillation

Based on addressing the challenges of high computational complexity and poor real-time performance when deploying radar operating mode recognition models on embedded devices, this paper proposes a knowledge distillation-based rapid recognition framework. The proposed algorithm first trains a deep residual teacher network to extract deep features from radar reconnaissance signals. Subsequently, a lightweight fully convolutional student network is constructed. Through a knowledge distillation mechanism, the teacher model's critical modeling capabilities for complex modulation patterns are transferred to the student network. Experiments on a radar operating mode dataset demonstrate that the proposed algorithm compresses the student model's parameters to approximately 1/40th of the original teacher model. Compared to directly trained models of equivalent scale, it achieves a recognition accuracy improvement of 4 percentage points, reaching 94.3%. This significantly addresses the challenge of high-precision real-time recognition of radar operating modes in resource-constrained scenarios.