High frequency source depth localization using deep neural network

Abstract

Localization of high-frequency sources in underwater is a difficult problem because the propagation model is increasingly sensitive to environmental parameters as the frequency increases. A deep learning approach is proposed to estimate the depth of a high-frequency underwater source using a single hydrophone trained on real data. A residual neural network is trained by a spectrogram of measured signal and estimates the depth of the source as a regression problem. The method is applied to data collected during the shallow water acoustic variability experiment 2015 in the northeastern East China Sea and compared with the results of the frequency difference matched field processing which uses 16 sensors. It was also validated that the trained model can display the generalized results for the signals measured at other times or at similar settings but different source-receiver location. As a result, the neural network is able to more accurately estimate the depth of high-frequency source and shows the features found by the network from real data are effective for localization, while the model often fails to generate accurate replicas.