3D Deep Residual Convolutional Neural Network for Underwater Acoustic Source Localization Using Local Acoustic Intensity Field

Abstract

The development of underwater robot technology provides a powerful means to obtain the local spatial distribution of an acoustic source radiated field. The local acoustic field interference characteristics can be used to locate the target source. In this paper, a deep neural network was established by taking the local acoustic intensity field as input, and the interval prediction and confidence distribution of the target source position as output. To extract multiscale features and improve network degradation problems, a 3D deep residual convolutional neural network structure was constructed. The convolution kernel was specially designed to keep the 3D spatial detail features of the local acoustic intensity field. The training and validation data sets are generated by the BELLHOP model. The performance of the trained network was evaluated using the validation set. The results show that the designed deep neural network can accurately estimate the azimuth, distance and depth interval of the target source relative to the local acoustic intensity field, and provide the confidence distribution of each interval.