Autonomous feature discovery, representation and analysis for small SONAR targets using domain cognizant machine learning techniques

This thesis tackles difficult problems associated with target recognition in the shallow ocean with SONAR. The United States Navy has a vested interest in this technology for use in explosive mine detection and identification. SONAR data contains information about the environment and any targets, but efficiently extracting this information is difficult.

One key contribution of this thesis is the application of complex-valued machine learning techniques to SONAR data. Traditional SONAR processing techniques make use of a frequency representation for data, which is complex-valued. Acoustic analysis is typically done with decibels, so only the magnitude of the frequency data is used and the phase information, which contains sound arrival times, is excluded. Complex-valued machine learning, as presented in the thesis, makes full use of phase information and is shown to far outperform equivalent methods that only used magnitude. The inclusion of phase and its arrival time data allows machine learning models to make inferences about positions, shapes and other properties of the soundscape that affect sound traversal.

Extracting the environment and target information from SONAR is a tough problem that complex-valued machine learning helps approach. The thesis makes progress towards parameter extraction with autoencoders, which are a special type of neural network that produce low dimension encoding spaces for the data. The end goal of this autoencoder research is disentangling the encoding space to associate patterns with interpretable physical properties such as target geometry, location and material. In this work, clear emergent patterns tied to target position are observed.