Time-Domain Adversarial Network for High-Fidelity Radar Signal Separation Under Interrupted Sampling Repeater Jamming

Abstract

To achieve high-fidelity single-channel radar signal separation in complex electromagnetic environments heavily corrupted by coherent interference, this article proposes a time-domain network based on adversarial learning for interrupted sampling repeater jamming (ISRJ) suppression. With the complex-valued radar baseband signals represented as dual-channel real tensors to fully preserve phase information, a 1-D encoder–decoder generator augmented with an attention mechanism is designed to adaptively extract temporal dependencies and filter jamming features within the high-dimensional latent space. To ensure precise waveform reconstruction, two independent multiresolution discriminators (MRDs) are established to provide hierarchical adversarial supervision across multiple temporal scales. Furthermore, a physics-informed hybrid loss framework is introduced, which optimizes waveform shapes via a scale-invariant signal-to-noise ratio (SI-SNR) metric while simultaneously enforcing physical boundaries through multidomain constraints. Experimental results on simulated ISRJ data indicate that the proposed network achieves a signal distortion ratio (SDR) of 8.44 dB and a signal-to-jamming-plus-noise ratio (SJNR) of 14.81 dB after pulse compression (PC), outperforming representative traditional and deep-learning-based baselines under the tested simulation settings. Under the tested implementation environment, the proposed network achieves an average single-sample inference time of 8 ms, demonstrating its potential for low-latency radar signal processing.