A Two-Stage DNN Model With Mask-Gated Convolution for Automotive Radar Interference Detection and Mitigation

Abstract

As the number of radar sensors on the road increases rapidly and many of these sensors share the same frequency spectrum, mutual interference cannot be avoided. This paper introduces a novel automotive radar interference mitigation approach using an autoencoder model which consists of separate neural networks for the detection and reconstruction steps. A mask-gated convolution is proposed to help the reconstruction neural network to learn the signal pattern from interference-free samples and to interpolate accordingly the signal segments at the disturbed positions. Through perturbation analysis it is shown that the reconstruction neural network can recover the distorted samples by utilizing their surrounding relevant samples. By exploiting the nature of interference in real-world scenarios, the proposed training approach does not need hand-labeled training data. Together with the proposed composite training loss, the neural network can recover the disturbed discrete beat signal with remarkable improvements in the signal-to-interference-plus-noise ratio (SINR) and the mean absolute percentage error (MAPE). Moreover, despite the use of a purely simulated training data set, the autoencoder can deal with real-world radar measurements which are more complex than the training data set.