A Toeplitz Prior-Based Deep Learning Framework for DOA Estimation With Unknown Mutual Coupling

In this paper, we explore the problem of direction-of-arrival (DOA) estimation with unknown mutual coupling using a deep learning (DL) framework which is based on Toeplitz prior. First, for source number estimation, we model it as a multilabel classification task and build a source number detection network (SNDN) to learn relevant information in the real sample covariance matrix. Next, taking full advantage of the Toeplitz structure, an ideal covariance reconstruction network (ICRN) is proposed to recover the ideal covariance matrix free from mutual coupling and noise interference. Furthermore, we design a database to store the parameters of ICRN after training on different numbers of sources, and its role is to load the corresponding parameters for ICRN according to the detection results of SNDN. Finally, the DOAs can be easily estimated from the restored covariance matrix by the MUSIC. The simulation results show our proposed approach not only outperforms the existing classical methods, but in some cases its DOA estimation accuracy can even exceed the Cramér-Rao Lower Bound.