An Encoder-Decoder Retnet Model for Fmcw Radar Interference Mitigation

Interference mitigation is one of the primary challenges in radar technology, particularly for automotive radars that utilize frequency-modulated continuous wave radars, where reliability is crucial. In this study, the RetNet-ED model, an encoder-decoder architecture that adapts the retentive network for the signal-to-signal task of interference mitigation, is proposed. The core innovation is the design of a stable, deep architecture that successfully applies the RetNet mechanism to the complex domain of radar signals. Specifically, a segmented processing pipeline to handle long sequences of signal and gated residual connections are added, a key stabilization technique that enables the deep network to converge effectively. Extensive experiments using the open-source ARIM-v2 dataset demonstrate that RetNet-ED delivers state-of-the-art performance. In particular, it exceeds a robust Transformer baseline in Signal-to-Distortion Ratio (SDR) and sets a new benchmark for both Amplitude and Phase Estimation Error, demonstrating superior reconstruction accuracy across all tested Signal-to-Noise Ratios (SNR). This work demonstrates that the RetNet model, with proper architectural adaptation, is a successor to the transformer model, providing a high-performance alternative for complex signal processing tasks and offering a new, validated approach to developing high-quality interference cancellation models.