Deep Learning-Based DOA Estimation: Enhancing Angle Resolution and Efficiency

DOA (Direction of Arrival) estimation is an array signal processing technique used to determine the direction of signals received by a sensor array. It is commonly applied in fields such as radar, wireless communications, and audio processing. Traditional DOA estimation techniques are limited by the Rayleigh criterion, which results in multiple signals within the main lobe beam being indistinguishable. These techniques have gradually been replaced by modern super-resolution DOA estimation techniques. MUSIC (Multiple Signal Classification) and ESPRIT (Estimation of Signal Parameters via Rotational Invariance Technique) are the two most common super-resolution DOA estimation algorithms. They can surpass the limitations of the Rayleigh criterion and improve the angular resolution. In the past decade, deep learning technology has rapidly developed and achieved significant advancements in various fields. In recent years, deep learning has also been extended to array direction finding, where neural networks are utilized to learn the features of array signals and estimate the target angles based on the output results. Such techniques eliminate the covariance matrix decomposition process involved in super-resolution algorithms, reducing computational complexity. This paper focuses on deep learning-based DOA estimation algorithms to address certain issues present in traditional super-resolution algorithms and deep learning-based algorithms, with the aim of improving angle estimation performance. The main work of this paper are as follows: 1. Introduction to the theory of array signal processing, including the array signal model. Then, the MUSIC algorithm for super-resolution is introduced, including the theoretical derivation and simulation results. Through the simulation results, it is found that the MUSIC algorithm has limitations in low signal-to-noise ratio and small angle separations, providing room for improvement in future algorithms. 2. The algorithm flow of the deep learning-based DOA estimation under the known number of sources condition is described. A deep learning regression-based DOA estimation algorithm is proposed, and the performance difference between the proposed algorithm and existing algorithms is verified through simulation experiments. The proposed algorithm reduces the time required for training and testing, and exhibits good estimation accuracy in low signal-to-noise ratio and small angle separations.