Two-Stage DOA Estimation with Multi-Channel Attention Residual Network and Maximum Likelihood: A Multi-Feature Fusion Approach

Le résumé fourni par la source

Abstract As a pivotal research direction in array signal processing, Direction-of-Arrival (DOA) estimation has been extensively applied in radar systems, wireless communications, and target tracking scenarios. Conventional subspace-dependent algorithms (represented by MUSIC and ESPRIT) tend to exhibit notable performance deterioration under adverse scenarios including low signal-to-noise ratio (SNR), insufficient snapshots, and coherent signal environments. Meanwhile, pure deep learning (DL)-based estimation schemes are restricted by grid discretization errors in grid-based inference and limited diversity of spatial feature representations. To tackle these limitations, this paper presents a two-stage DOA estimation framework that integrates a Multi-Channel Attention Residual Network (MCARN) with Maximum Likelihood (ML) optimization. In the initial coarse estimation stage, the MCARN leverages six complementary feature inputs—namely the real and imaginary components, amplitude and phase spectra of the array covariance matrix, fourth-order cumulants, as well as the beam spectrum within the angular span of [−60°, 60°]—to fully mine spatial and statistical characteristics, enabling robust coarse DOA prediction and effectively shrinking the subsequent search domain. In the refinement stage, the ML estimator employs the coarse estimation results as prior information to conduct high-precision local searching, which removes grid quantization errors and further enhances estimation accuracy. Numerical simulation results validate that the proposed method surpasses classic subspace approaches and state-of-the-art DL-based estimators in both estimation accuracy and robustness under low SNR and single-snapshot conditions. Furthermore, it attains an excellent tradeoff between computational efficiency and estimation performance, exhibiting promising practicability for real-time engineering applications.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.