Passive Source Ranging Using Residual Neural Network With One Hydrophone in Shallow Water

Abstract

The source ranging problem can be regard as a classification problem in machine learning. The paper used a deep neural network (ResNet18) as a deep learning model to estimate the source range based on a single hydrophone in the shallow water. The simulation data generated by the acoustic propagation model were used as the training data. The trial data from the SACLANT experiment (1993) as test data have demonstrated the performance of the method. The results indicate that a single hydrophone in the shallow water environment is applicable to predict the source range when choosing an appropriate deep learning model. The analyzation of a shallow water sea trial data shows that the average of the range estimation for samples is 5.44 km. And the mean square error and the mean absolute percentage error of ranging were 0.036 km2 and 1.5308%, respectively.