Deep Learning for DOA Estimation: Novel Neural Network Architecture for Correlated Sources

Abstract

Direction of Arrival (DOA) estimation is a critical aspect of signal processing in various applications, particularly in wireless communication. This thesis presents a novel approach aimed at improving DOA estimation performance under challenging conditions such as coherent sources and low signal-to-noise ratio (SNR), specifically focusing on scenarios with two coherent sources utilizing a Uniform Linear Array (ULA). The research highlights the limitations of existing DOA estimation schemes, particularly the increase in estimation error as the angle of incoming signals deviates from the center within the range of (-90°, 90°). Additionally, it addresses the need for enhanced performance under low SNR conditions. The proposed solution introduces a Cascaded Neural Network (CaNN) architecture, consisting of two stages of neural networks. The first stage comprises an Enhanced SNR (ESNR) classifier, designed to enhance performance across various SNR levels. The second stage involves an angle estimator neural network, which aims to improve performance across different angle ranges. To mitigate the challenges posed by coherent signals, the thesis proposes the use of a spatially smoothed auto-covariance matrix, which is fed into both the SNR classifier and angle estimator blocks. A comprehensive performance evaluation is conducted, comparing the proposed CaNN approach with existing schemes such as Spatial Smoothed-Multiple Signal Classification (SS-MUSIC) and ESNR CaNN. The results demonstrate the superiority of the proposed CaNN in terms of DOA estimation accu- racy across different angles and SNR ranges over traditional DOA estimation techniques. Overall, this research contributes to advancing DOA estimation techniques, particularly in scenarios with coherent sources and low SNR, by leveraging the capabilities of neural networks within a cascaded architecture.