Matched-field source localization using sparsely-coded neural network and data-model mixed training

Abstract

Source localization is a basic problem in underwater acoustics. Many solving approaches have been developed, and the matched-field processing (MFP) is one of the mostly-studied. However, MFP is sensitive to the mismatch problem and performs well only when the knowledge of ocean environment is accurate. Machine learning learns directly from the observation and can be designed to learn a generic model suitable for different scenarios. In this paper, source localization is viewed as a machine learning problem and a matched-field source localization model is learned by training a sparsely-coded feed-forward neural network with mixed environment models and data. Sparsely-coded network can prevent the model from over-learning. Results on SWellEx-96 experiment show that the learned model achieves good positioning performance in source range estimation for varying sound-speed profiles (SSP). Compared with Bartlett matched-field processing, machine learning model is more robust and thus has potential advantages in underwater source localization.