Research on high-resolution DOA estimation methods for closely spaced signals in low SNR

Estimation of direction of arrival (DOA) for closely spaced sources under low signal-to-noise ratio (SNR) conditions remains a critical challenge. Classical subspace-based algorithms often fail in such scenarios, while existing deep learning models lack effective mechanisms to explicitly exploit the specific geometry of conformal arrays. To address this challenge, a novel multi-architecture fusion network termed the Graph-Filtered Regression Network (GFR-Net) is proposed. The network first leverages a Multi-Layer Perceptron (MLP) module to classify input signals into distinct SNR levels, thereby adapting to the modeling requirements of signals across different SNR regimes. Next, a Graph Neural Network (GNN) is employed to construct a spatial filtering network, which enhances the structural integrity of the signals by capturing inherent spatial correlations. Finally, a Convolutional Neural Network-Long Short Term Memory (CNN-LSTM) hybrid network performs spatiotemporal modeling on the filtered signals, enabling accurate DOA regression estimation. Simulation results demonstrate that the proposed GFR-Net significantly improves estimation accuracy and angular resolution for closely spaced sources under low SNR conditions, outperforming conventional subspace-based and representative deep learning methods in terms of robustness and reliability.