DOA Estimation Using an Unfolded Deep Network in the Presence of Array Imperfections

Deep learning has gained remarkable popularity due to its widespread success in a number of practical applications such as image classification and speech recognition. In this paper, we consider the application of deep learning to DOA estimation problem in the presence of array imperfections. Firstly, through spatially overcomplete formulation, the DOA estimation problem is converted to a sparse linear inverse problem, where one seeks to recover a sparse signal from a few noisy linear measurements. Then an iterative sparse signal recovery algorithm, iterative shrinkage/thresholding algorithm(ISTA), is “unfolded” to form an interpretable and learnable deep network-learned ISTA(LISTA). LISTA is applied to recover the DOA spectrum after trained with vast amounts of training data. We find that LISTA is of good robustness against noise and array imperfections due to the training procedure. Comprehensive simulations and experiments have been carried out, and the superiority of the proposed method can be clearly seen.