A High Precision DOA Estimation Method Based on Convolutional Neural Network

Direction-of-arrival (DOA) estimation is a critical research area in array signal processing, with extensive applications in fields such as radar, sonar, communication systems, and aerospace [1]. With the widespread deployment of 5G massive MIMO and millimeter-wave systems, base stations must perform high-precision beam alignment and user localization in real time. Consequently, robust and low-complexity DOA estimation has become a critical enabling technology for modern wireless communications. Traditional subspace-based signal decomposition algorithms incur high-computational complexity, and performance degrades at low signal-to-noise ratios (SNRs). In contrast, deep learning (DL)-based approaches have recently emerged as a powerful alternative, demonstrating robust performance in low-SNR regimes while significantly reducing computational complexity. Existing DL-based DOA estimators frequently depend heavily on the training data distribution, struggle to resolve closely-spaced sources, and typically exhibit underperforming classical algorithms at high SNRs. This work focuses on enhancing estimation accuracy in complex scenarios, a critical challenge in many real-world applications.