Adaptive Direction-of-Arrival Estimation Using Deep Neural Network in Marine Acoustic Environment

Deep learning is widely used for target detection and direction-of-arrival (DOA) estimation due to its powerful data fitting capability. However, limited by different environments and number of sound sources, it is difficult to be applied to complex underwater environments. We propose a two-stage approach called beam network for underwater acoustic DOA estimation. In the first stage, local beam patterns with different data augmentation methods, carrying the essential information required for target detection, are used as the input feature to our model. In the second stage, an adaptive convolutional neural network (CNN) is proposed to construct a classification model. Only single-source data are required for model training, and data from multisources can be tested. Furthermore, the model is suitable for arrays with different numbers of hydrophones in different geometrical arrangements. The performance of the proposed method is evaluated by comparing with mainstream DOA estimation algorithms, such as conventional beamforming (CBF), multiple signal classification (MUSIC), minimum-variance distortionless response (MVDR), and sparse Bayesian learning (SBL). In three simulation scenarios and two sets of recorded data from different marine environments, the proposed method has higher directivity and lower angular root-mean-squared error (RMSE).