Deep learning based 2D-DOA estimation using L-shaped arrays

Abstract

Direction-of-arrival (DOA) estimation problems arise in many applications such as wireless communication and localization. Recently, a number of deep learning (DL) based methods have been studied for one dimensional (1D) DOA estimation with relatively fewer studies for 2D-DOA estimation. In this study, we propose a low-complexity DL based method to estimate both elevation and azimuth DOAs of sources along with their pairing. To this end, first a classification neural network is proposed to estimate both elevation and azimuth DOAs of multiple sources. Next, a residual classification network is introduced to estimate the pairing between the estimated DOAs. In particular, we consider two perpendicular linear arrays which are located on z-axis and x-axis, respectively. The first proposed network uses the Sample Covariance Matrix (SCM) of the former array to estimate elevation angles and utilizes the latter array to estimate the corresponding azimuth DOAs. The second network, is responsible for estimating the pairing between the estimated elevation and azimuth DOA sets. Numerical simulations demonstrate the enhanced performance of our proposed 2D-DOA estimator scheme, surpassing existing 1D and 2D deep learning (DL) methods. Notably, our approach closely approaches the Ziv–Zakai bound (ZZB), particularly in low signal-to-noise ratio (SNR) and low-angle-difference scenarios, even in the presence of multiple highly correlated signals. Moreover, our complexity analysis validates the superiority of the proposed method.