Automotive Radar Interference Mitigation Technology Based On Complex-Valued Network And Prior Information

Due to the advantages of being cheap and able to penetrate dust,millimeter wave(mmWave)radar has become an indispensable sensor in cars.With the increasing number of vehicles and radio equipment equipped with mm Wave radars and the limited radar spectrum bandwidth,the probability of interference between the mm Wave radars is increasing,which would affect the target detection of mm Wave radars.Therefore,based on the background of automotive mm Wave radar,this thesis carries out in-depth research on the automotive radar interference mitigation technology based on complex-valued networks and prior information.The main research content and technological innovation of this thesis include:1.Radar interference mitigation method in the time-frequency(t-f)domain with multiple interference sourcesIn view of the lack of open-source automotive Frequency Modulated Continuous Wave(FMCW)radar interfered signal dataset,and the previous interference mitigation research is limited to the scenario with only one interference source,an FMCW radar interfered signal dataset in the multi-interference sources scenario is proposed.Moreover,to solve the problem that the existing interference mitigation networks have too large number of parameters and poor real-time processing capability,considering the target echo signal and interference signal of FMCW radar have significantly different distributions in the t-f domain,this thesis proposes a deep-learning-based radar interference mitigation method in the t-f domain.The results show that a simple convolutional neural network can achieve good interference mitigation performance in the t-f domain.Besides,the network trained with only simulated radar signals has good generalization performance when processing measured radar signals.2.Radar interference mitigation algorithm based on complex-valued network and sparse prior information constraintThis part aims at the problem that the deep-learning-based radar interference mitigation approaches rely on large-scale signal datasets and it is hard for them to suppress strong interferences.Considering using realvalued neural networks to process complex-valued radar signals would ignore the phase relationship between the real and imaginary parts of radar signals,the complex-valued modules are used to build a complex-valued fully convolutional network.Compared with the real-valued ones,with the number of network parameters reduced by half,better interference mitigation performance is achieved in the low signal-to-interferencenoise-ratio scenarios.Besides,using the mean square error as the loss function would make the trained network easily lead to overfitting.Considering the sparse distribution characteristic of the target echo signal,a new regularization term is constructed,which is combined with the mean square error to accelerate the training convergence of the network.The results show that a better interference mitigation result is achieved when trained using smaller training datasets,and better generalization performance is realized.3.Radar interference mitigation algorithm based on contrastive learningTo solve the problem of residual interference components in the recovered signal after interference mitigation,the dilated convolution is used to build the interference mitigation network to enlarge the receptive field,and the optimal setting of dilated rate of convolutional layers is proposed to keep all the elements of the t-f spectrum are processed and maximize the receptive field.Moreover,a new contrastive learning strategy and contrastive loss function are proposed.The interference signals are generated separately and taken as negative samples,and the clean reference signals are taken as positive samples.By using the contrastive learning,the residual interference in the recovered signal is further suppressed.Besides,the proposed method does not require an extra feature extraction network and hyper-parameter,and does not add computational complexity.In a word,to solve the problem that mutual interference between automotive mmWave radars would affect its normal operation,this thesis proposes a deep-learning-based method to process the received signals for interference mitigation.By suppressing the interferences in the t-f domain,the network parameters can be reduced,enabling faster signal processing.Then,by introducing the complex-valued network,sparse target prior information,and interference feature,the dependence of the large-scale datasets is reduced,and better interference mitigation performance is achieved for both strong interferences and residual interference components.Besides,better generalization performance can be seen when processing the measured radar signal.