Physics-Guided Covariance Regularization Network for Robust DOA Estimation Under Low-SNR and Finite-Snapshot Conditions

Abstract

Direction-of-arrival (DOA) estimation under low signal-to-noise ratio (SNR) and finite-snapshot conditions remains challenging in array signal processing. Under such conditions, the sample covariance matrix may deviate substantially from the true covariance matrix. This deviation degrades the estimation performance of conventional model-driven methods and limits the effectiveness of existing deep learning approaches. To address this problem, a Physics-Guided Covariance Regularization Network (PGCR-Net) is proposed. The Physics-Guided Covariance Regularization (PGCR) module performs Hermitian-preserving covariance augmentation followed by Positive Semi-Definite Projection (PSDP), thereby improving covariance diversity while maintaining positive semi-definiteness. A Symmetry-Aware Feature Fusion (SAFF) module is further developed to exploit symmetric and antisymmetric structures in covariance-derived feature tensors, thereby enhancing structure-aware feature representation for DOA estimation. Experimental results show that PGCR-Net achieves improved estimation accuracy and robustness, outperforming conventional subspace-based estimators and representative deep learning baselines, particularly under low-SNR, finite-snapshot, and array-model mismatch conditions.