Direction of arrival estimation of unknown emitter by deep neural networks with array imperfections

In array signal processing, high-resolution direction-of-arrival (DOA) estimation by eigendecomposition method requires knowledge of the array covariance matrix and an exact characterisation of the array. The dependence of most such methods on the array manifold substantially degrades their array performance when the actual sensors deviate from their assumed nominal values. This article aims at formulating a deep neural network framework for DOA estimation with feasible computational complexity. The two-stage algorithm is composed of a detection network and a series of parallel DOA estimation networks: the multilayer perceptron is trained to detect the presence of one or more sources in an angular sector at the first stage, the DOA estimation networks then estimate their respective DOAs at the second stage. When a single type of array imperfection is considered, numerical results reveal that the DNN-based method can obtain DOA estimates with much higher precision than the existing techniques for significant location errors and gain/phase perturbations, they also illustrate the potential usefulness of the proposed algorithm towards better DOA estimation accuracy for mutual coupling as the mutual coupling matrix is not simplified. Moreover, the proposed method is finally extended to deal with the DOA estimation problem when multiple types of array perturbations coexist.