RadarImputationNet: An Enhanced Long-Term Diffusion Model for Automotive Radar Interference Suppression and Signal Recovery

Automotive radar is crucial in advanced driver assistance systems (ADAS) and autonomous driving. However, radars encounter considerable challenges when other radars operate within a similar frequency band and generate mutual interference. The radar signal imputation network (RadarImputationNet) is proposed to suppress interference signals by accurately reconstructing missing or corrupted radar data. The long-term dependency enhancement module (LTDEM) is proposed to improve the precision of signal imputation. The proposed RadarImputationNet is used after interfering samples are removed from the pulse. The recent diffusion models have poor signal imputation accuracy and performance in recovering signal peaks. The proposed RadarImputation- Net model solves the dilemma by improving the long-term dependency of the diffusion model. A real-data experiment is conducted in the dynamic scene to verify the proposed RadarImputationNet. Metrics including root-mean-square error (RMSE), mean absolute error (MAE), and continuous ranked probability score (CRPS) are compared, and the proposed RadarImputationNet outperforms existing methods in RMSE, MAE, and CRPS. RMSE decreased by 19.30, representing a 50.7% reduction; MAE decreased by 13.91, representing a 48.6% reduction; and CRPS decreased by 0.17, representing a 48.3% reduction, indicating a significant improvement in the performance of the diffusion model. Furthermore, the proposed RadarImputationNet demonstrates exceptional proficiency in restoring the peak characteristics of the missing samples.