Date of Award


Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy in Electrical Engineering (PhD)

Administrative Home Department

Department of Electrical and Computer Engineering

Advisor 1

Timothy Havens

Advisor 2

Andrew Barnard

Committee Member 1

Tony Pinar

Committee Member 2

Lan Zhang

Committee Member 3

Miles Penhale

Committee Member 4

Laura Brown


Arctic acoustics have been of concern in recent years for the US navy. First-year ice is now the prevalent factor in ice coverage in the Arctic, which changes the previously understood acoustic properties. Due to the ice melting each year, anthropogenic sources in the Arctic region are more common: military exercises, shipping, and tourism. For the navy, it is of interest to detect, classify, localize, and track these sources to have situational awareness of these surroundings. Because the sources are on-water or on-ice, acoustic radiation propagates at a longer distance and so acoustics are the method by which the sources are detected, classified, localized, and tracked. These methods are all part of sound navigation and ranging (SONAR). This dissertation describes algorithms which will better SONAR results without modification of the sensors or the environment and the process by which to arrive to this point. The focus is to use supervised machine learning algorithms to facilitate such technological enhancements. Specifically, neural networks analyze labeled experimental data from a first-year, shore-fast, shallow and narrow water environment. The experiments were conducted over the span of three years from 2019 to 2022, mostly during the months from January to March where ice formed over the Keweenaw Waterway at the Michigan Technological University. All experiments were conducted to analyze a passive acoustic source; that is, the source was non-cooperative and did not send any localizing pings for active SONAR. The experiments were recorded using an underwater pa-type acoustic vector sensor (AVS). The data and analysis were done intermittently to update any upcoming experiments with discrepancies found in the analysis to create a more generalized algorithm. The work in this dissertation focuses on two topics for passive SONAR: localization and classification. Because of the ``black box" nature in machine learning, tracking the target source is an extension of localization and thought of as the same goal within machine learning. To introduce and verify the complexity of the testing environment, an underwater acoustic simulation is shown with Ray tracing and bathymetry data to compare with the experimental results used in machine learning. The focus of the algorithms is to produce the best results for the experiments and compare the results with traditional methods, such as a simulation or a linear Gaussian localization with a Kalman filter. Experiments studying neural network types have shown that the Vision Transformer (ViT) produces excellent results. The ViT is capable of analyzing acoustic intensity azimuthal spectrogram (azigram) data and localizing a moving target at high accuracy, and the ViT is capable of classifying multiple acoustic sources with the acoustic intensity magnitude spectrogram at high accuracy as well.

Creative Commons License

Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.