Robust Direction-of-Arrival Estimation Based on Cascaded Deep Neural Networks

This paper proposes a cascaded deep neural network framework specifically designed to tackle the challenges of Direction of Arrival (DOA) estimation in environments characterized by antenna array imperfections (channel amplitude/phase errors) and low signal-to-noise ratio (SNR). The framework comprises two primary modules: Firstly, the Covariance Reconstruction Module, which utilizes a Convolutional Autoencoder (CoAE) network to reconstruct sample covariance, effectively mitigating the interference induced by array imperfections. Secondly, the DOA Estimation Module, leveraging a two-dimensional Convolutional Neural Network (CNN), takes the output from the CoAE network as its input. This module harnesses the robust nonlinear mapping capabilities of the CNN to achieve high-precision DOA estimation under low SNR conditions within a multi-label classification framework. Simulation results indicate that the proposed model significantly reduces DOA estimation errors and demonstrates notable improvements in performance and robustness compared to methods in the literature.