Signal as Token: Robust DOA Estimation in Complex Environments Aided by Antenna-Based Attention

Direction of Arrival (DOA) estimation, with applications in various fields, is a widely-researched problem. However, the lack of adaptation DOA estimation methods in the presence of various complex environments remains a significant challenge. Traditional approaches rely on precise modeling and assumptions, which limit their flexibility in real-world scenarios. Learning-based methods, including machine learning and deep learning, have partially mitigated these issues. Nonetheless, they often lack a comprehensive analysis across diverse scenarios and suffer from limited mathematical interpretability. In this paper, we propose an approach that harnesses the power of the Transformer to tackle the DOA estimation problem. By employing the Transformer in the DOA estimation task and introducing an antenna-based attention mechanism tailored for DOA estimation, we rigorously demonstrate through mathematical derivations that the output of antenna-based attention corresponds to the pseudo-Singular Value Decomposition (pseudo-SVD) of the covariance matrix. Leveraging this mechanism to capture more profound feature information within the received signals, leads to highly accurate DOA estimation. Furthermore, our proposed method exhibits good adaptability in the presence of low signal-to-noise ratios (SNRs), a limited number of snapshots, array errors, coherent sources, and broadband sources. Rigorous experiments conducted on synthetic and real-world datasets validate the effectiveness and generalization ability of our method. Additionally, our proposed method has also been proven effective in solving the problem of estimating the number of signal sources.