RTDNet: Modulation-Conditioned Attention Network for Robust Denoising of LPI Radar Signals

Accurate processing of low-probability-of-intercept (LPI) radar signals poses a critical challenge in electronic warfare support (ES). These signals are often transmitted at very low signal-to-noise ratios (SNRs), making reliable analysis difficult. Noise interference can lead to misinterpretation, potentially resulting in strategic errors and jeopardizing the safety of friendly forces. Accordingly, effective noise suppression techniques that preserve the original waveform shape are crucial for reliable analysis and accurate parameter estimation. In this study, we propose the recognize-then-denoise network (RTDNet), which effectively removes noise while minimizing signal distortion. The proposed approach first employs a modulation recognition network to infer the modulation scheme and then feeds the inferred label to an attention-based denoiser to guide feature extraction. By leveraging prior information, the attention mechanism preserves key features and reconstructs challenging patterns such as polytime and polyphase codes. Simulation results indicate that RTDNet more effectively removes noise while maintaining the waveform shape and salient signal structures compared with existing techniques. Furthermore, RTDNet improves modulation classification accuracy and parameter estimation performance. Finally, its compact model size and fast inference meet the performance and efficiency requirements of ES missions.