Effective High-Resolution Off-Grid DOA Estimation With Mutual Coupling via CNN Framework

Achieving high-resolution direction-of-arrival (DOA) estimation under the condition of mutual coupling (MC) is of great practical importance. Traditional parametric approaches are limited by either the computational complexity or the resolution. While, existing neural network (NN)-based end-to-end methods achieve DOA estimation by spatial spectrum regression task or multilabel classification task, which results in the performance being limited by off-grid errors. Besides, the MC between the sensors may seriously affect the precision and resolution of these approaches. In this article, an NN framework consisting of convolutional NN (CNN) is designed for effective high-resolution end-to-end off-grid DOA estimation with MC. The designed CNN framework consists of source enumeration CNN (SECNN) and multiple parallel DOA estimation CNNs (DOACNN) to enhance its generalization for different numbers of sources. The SECNN is designed to provide knowledge of source numbers for DOACNN, and the DOACNN is to achieve DOA estimation based on the provided knowledge. Both SECNN and DOACNN are trained by the output covariance of certain arrays thus robust to MC. Meanwhile, the DOACNN models off-grid error into labels to achieve effective end-to-end off-grid DOA estimation. As compared to the representative approaches, the proposed method not only can realize high-precision off-grid DOA estimation in the presence of MC but also has noticeable advantages in terms of resolution and efficiency.