Subspace-based DOA estimation method for coherent signals assisted by transformer

Direction of Arrival (DOA) estimation plays a crucial role in array signal processing, with broad applications across various fields. Traditional Model-Based (MB) DOA estimation algorithms, however, are constrained by their reliance on prior knowledge and tend to perform poorly in non-ideal conditions, such as when dealing with coherent sources or limited snapshot numbers. In contrast, Data-Driven (DD) approaches using deep learning offer novel insights and capabilities for enhancing DOA estimation. This paper introduces a novel hybrid MB-DD framework for estimating the DOAs of multiple coherent narrowband signals captured by Uniform Linear Arrays (ULA). Specifically, the proposed method, TransSubNet, employs an autoencoder module centered around a Transformer Encoder to generate a proxy covariance matrix, which is then fed into classical subspace algorithms, such as root-MUSIC, to estimate the DOAs of the received signals. By integrating the strengths of both MB and DD methods, TransSubNet utilizes the Transformer-based network to improve the covariance matrix estimation process, enhancing the robustness and accuracy of subspace-based DOA algorithms. Our architecture achieves high-precision DOA estimation even under challenging conditions, such as low snapshot numbers, low Signal-to-Noise Ratios (SNRs), and 1-bit quantized signals. Experimental results show that TransSubNet outperforms existing algorithms in the domain, delivering superior performance across a range of scenarios.