Robust DOA Estimation Based on Deep Neural Networks in Presence of Array Phase Errors

Deep learning (DL) framework is gradually applied to solve the problem of DOA estimation in array signal processing. DL-based DOA estimation methods are much more efficient than conventional model-based methods in the testing stage. However, the generalization of DL-based methods is limited in the presence of array phase errors, because array phase errors may change in different environments, leading to the difference between the phase errors in the training and the ones in testing. In this paper, we explore the magnitude property of array received signal to develop robust deep neural network (DNN)-based framework for DOA estimation, named as magnitude-based DNN method (shorten as MDNN). The proposed MDNN method performs independently of array phase errors and enjoys a simpler network than the original DNN method. Simulation results in different scenarios demonstrate that the MDNN method behaves much more robust to array phase errors than the original DNN-based method.