Robust DOA Estimation Method for MIMO Radar via Deep Neural Networks

Abstract

It is a serious problem that the performance loss is suffered by traditional Direction-of-Arrival (DOA) estimation methods in non-ideal environment, such as mutual coupling of array elements, coherent sources, colored noise and plethora targets. A data-driven robust DOA estimation framework is proposed for MIMO radar via deep neural networks (DNN), so as to overcome the problems mentioned before. The framework consists of an autoencoder, a feedforward network, a network parameters database and a series of parallel directed acyclic graph networks (DAGN). Assisted with feedforward network for target-number determination, matching parameters of networks will be loaded from database. The autoencoder acts like a noise filter, it reconstructs the noise-free covariance from the noisy signal and thus the generalization burden of the subsequent DOA estimation DAGN will be decreased. Each sub-network of the parallel DAGN consists of a convolutional neural network (CNN) and two bidirectional long short-term memory (BiLSTM) networks, from which the estimation of DOA will be obtained by regression. The simulation results show that the proposed method is superior to the traditional methods in a non-ideal environment, and can also perform well when the number of targets reaches the upper limitation of the degrees of freedom of MIMO radar.