A Modified Sparse Bayesian Learning Method for High-Accuracy DOA Estimation with TCN Under Array Imperfection

Abstract

Array imperfection may cause performance degradation to direction-of-arrival (DOA) estimation in practice. Most DOA estimation methods overlook the array imperfection by regarding the array manifold as a piece of precisely prior knowledge. Although previous works suggest some simple calibration processes, limitations of array errors like amplitude and phase deviation (AP) and antenna position perturbation (PP) may still lead to manifold mismatch against high-precision. The application of neural network (NN) methods in DOA estimation has demonstrated improved robustness but is limited in handling complex array errors. In this paper, a Transformer-based calibration network (TCN) is designed to capture global sequence information effectively and generate steering vectors of grid points. Then a framework based on modified root-sparse Bayesian learning (RSBL) is proposed to iterate calibration and estimation steps alternately. Extensive experiments show that the proposed method can achieve better performance in different array imperfections, including AP and PP, than other existing methods. When weak array imperfection exists, the proposed method keeps the average error below 0.5 degrees while MUSIC, OMP, and RSBL reach the highest above 2.7 degrees.