DOA Estimation Method Based on Multi-Layer Neural Networks

Abstract

As a subject of extensive research in numerous scientific and engineering disciplines, the Direction of Arrival (DOA) estimation leads to the creation of numerous specialized algorithms. As deep learning continues to evolve, it has progressively reshaped the field of DOA estimation, bringing forth new opportunities. Traditional DOA estimation algorithms typically require manual feature selection and model design; however, deep learning inherently extracts features and patterns from data autonomously, thus providing more accurate and robust DOA estimation. Within deep learning architectures, the MultGRU algorithm is proposed as a novel solution for DOA estimation. This algorithm employs Gated Recurrent Unit (GRU) to attain accurate DOA estimation. The MultGRU adopts a two-level architecture: The first-level GRU network performs coarse localization by mapping the target DOA to large angular intervals (5° each within 0° ∼ 90°) using array covariance matrix features. The second-level employs some parallel subGRU networks with 5° overlap between adjacent subGRUs to subdivide the coarse intervals into 0.1° fine-grained intervals. A probability-weighted correction mechanism is integrated to optimize estimates by fusing predictions from adjacent intervals to effectively reduce estimation errors. Experiments are conducted to compare the MultGRU with other DOA estimation algorithms under different numbers of snapshots and sensors, varying Signal-to-Noise Ratios (SNR), and in dual-source scenarios. The DOA estimation accuracy of the MultGRU is improved by more than 29% overall, and its computational efficiency is higher than theirs.