A ResNet-Aided Two-Stage DoA Estimation Technique for Multiuser SIMO Systems

Direction of arrival (DoA)- or angle of arrival/departure (AoA/AoD)-estimation is a fundamental technology to perform beam alignment and reconfigurable intelligent surfaces (RIS) systems operated in a mmWave band, since it does not require pilot overheads. Traditional methods like MUSIC can suffer from estimation errors seriously in doubly selective fading channels in multiuser RIS and/or vehicular-to-everything (V2X) systems, although they can achieve accurate enough performance in a single source scenario. Supervised learning-based DoA estimation, employing deep neural networks (DNN) and convolutional neural networks (CNN), outperforms the traditional methods in a low signal-to-noise ratio (SNR) regime by effectively learning features from trainig data while suppressing noises inherent in observed data. However, these methods still have mean absolute error (MAE) floor problems. To address this issue, we propose a novel two-stage DoA estimation framework composed of a classification-based coarse estimator and a regression-based fine estimator using a wide residual network (ResNet). Simulation results confirm superiority of the proposed technique for multi-source DoA estimation problems in doubly selective fading channels. Specifically, the new approach reduces MAE from 0.24 to 0.044 at SNR of 30dB, which corresponds to an improvement of 82% over a conventional CNN classification method. In addition, at a tolerance error of 0.1 degree, the cumulative distribution function (CDF) of the estimates obtained by the proposed technique is significantly improved to 0.9 from 0.2 obtained by the conventional method.