Contrastive Learning-Based Deep Neural Network for Robust DOA Estimation

A two-stage cascaded direction of arrival (DOA) estimation framework is proposed for parameter estimation under non-ideal array conditions. The framework consists of two modules: contrastive reconstruction network (CRN) and DOA estimation network (DEN). To enhance robustness against array imperfections, CRN leverages self-supervised contrastive learning (SSCL) to recover covariance matrix, which is then fed into the DEN to perform the DOA estimation task. To further improve generalization, an augmentation strategy using non-ideal data, such as amplitude-phase perturbations and random sensor masking, is introduced during data training, where a two-stage training strategy is adopted to separately optimize the parameters of CRN and DEN. Simulation results demonstrate that the proposed framework achieves higher estimation accuracy and stronger robustness in non-ideal array conditions.