Robust DOA Estimation Using Complex-Valued Residual Attention Networks With Low-Rank and Sparse Prior

With the widespread application of multi-sensor systems, direction of arrival (DOA) estimation in complex electromagnetic environments is crucial for target detection and localization. Under non-ideal conditions, the received signals are easily affected by uncontrollable factors such as coherent signals and variations in signal power. In this letter, a novel DOA estimation method based on the complex-valued residual attention convolutional neural network (CRA-CNN) is proposed. A complex-valued residual network integrated with an attention mechanism is introduced to extract key features from the signal covariance matrix, significantly enhancing feature representation and discrimination. Notably, a novel loss function combining low-rank and sparse prior constraints is designed to enhance sensitivity to essential features while suppressing redundancy and noise. Simulation results demonstrate that CRA-CNN improves both the accuracy and robustness of DOA estimation.