BYU

Abstract by Kira Howarth

Personal Infomation


Presenter's Name

Kira Howarth

Degree Level

Undergraduate

Abstract Infomation


Department

Physics and Astronomy

Faculty Advisor

Tracianne Neilsen

Title

Effect of signal to noise ratio on a convolutional neural network

Abstract

Identifying the location of a sound in water is complicated by the variability of the ocean environment. We find that the source location and type of seabed is a problem that can be solved with machine learning.  A Convolutional Neural Network (CNN) can be used to accurately approximate both the range and seabed type, even with the addition of noise. The CNN model trained with synthetic pressure time series. The trained CNN gives nearly perfect results when applied to the validation set—a portion of the synthetic signals set aside for validating the CNN. The generalizability of the CNN is tested on real world data; the real world data comes from a 2017 study of the New England mud patch. These results are also promising when appropriate levels of background noise are introduced into the training data. When noise levels similar to what would be seen in the ocean are included, the CNN predictions on the real data are similar, if not better, to those seen on training data with no noise. From the results, it is clear that a CNN can predict both range and seabed type, even with noise.