Phy-SSMT: A Physical-Driven Semi-Supervised Framework for Phase-Preserving Separation of Smeared Spectrum Jamming

Abstract

A fundamental challenge in radar anti-jamming is the practical difficulty of acquiring extensive labeled datasets within complex electromagnetic environments. To address smeared spectrum jamming, which generates a dense swarm of false targets to mask legitimate radar returns, this letter proposes a physical-driven semi-supervised mean teacher (Phy-SSMT) framework for single-channel complex echo separation. To alleviate the reliance on labeled data, a semi-supervised learning paradigm is introduced to leverage unlabeled radar received signals through consistency regularization between a student model and an exponential moving average teacher model. Furthermore, physical constraints in the fractional Fourier transform domain and energy maintenance criteria are incorporated as complementary regularizers to guide the reconstruction of physically consistent waveforms. Experimental results demonstrate that under a severe average input signal-to-interference-plus-noise ratio of negative 20.10 dB, the proposed framework achieves an improvement factor of 48.33 dB, substantially outperforming supervised baselines. Ablation studies show that Mean Teacher consistency regularization mainly improves complex waveform fidelity, while the FrFT constraint provides a modest complementary contribution to coherent processing.