Deep Learning Based DOA Estimation with Single Vector Hydrophone in Shallow Water

Deep learning method based on residual neural network and deep transfer learning is proposed for direction of arrival (DOA) estimation with a single vector hydrophone in shallow water. The data received by a single vector hydrophone is simulated by normal wave theory of acoustic propagation model. We propose a residual neural network that contains 50 layers network (Resnet50), which is trained with high-quality pre-processing data. A fine-tuned training strategy is proposed to transfer Resnet50 to a new acoustic environment. Convolutional layer parameters are adjusted freely with a lower learning rate in transfer learning, while the fully connected network layer is initialized. The simulation and the experiment results show that the deep learning method has excellent performance compared with the active intensity method and convolutional neural network (CNN) with the different environment conditions or the source signal-to-noise ratio (SNR) decreases.