Abstract
Conventional direction-of-arrival (DOA) estimation methods generally rely on the white Gaussian noise assumption, making them ineffective in hybrid noise scenarios. This paper proposes a deep neural network based on sparsely-gated mixture-of-experts (SMoE) mechanism for underwater DOA estimation in hybrid noise environments. The model first transforms the complex covariance matrix of the array signal into two real covariance matrices via the method of separating real and complex parts. Subsequently, a CNN is employed to extract spatial features from the array signal. Finally, a SMoE mechanism is utilized to handle hybrid noise environments, which dynamically adapts to diverse noise conditions via sparse expert activation. With our staged training strategy, the model achieves 0.94[Formula: see text] RMSE at 0 dB SNR with six noise types, outperforming conventional approaches. This work provides a feasible solution for DOA estimation in hybrid noise environments.