Performance of Kernel-Based Methods in Analyzing Underwater Noise Level Statistics
Abstract
In this work we present an evaluation of kernel-based methods of estimating spectral density statistics compared to histogram-based methods based on evaluations of fine structures. The data is from the Monterey Accelerated Research System, where month-long windows of data from February 2019 to January 2021 were preprocessed to provide 30-second power averages for analysis. The data was processed in MATLAB where it was analyzed through traditional histogram and kernel-based methods to produce spectral probability densities (SPD). The structure presented in the respective SPDs for each method was compared to determine where the methods converge, what differences appear, and which method performs best. Performance was analyzed through a statistical analysis of physical noise sources, specifically windspeed at the surface of the water. We developed a simple classifier that can categorize ambient noise data as being associated with high wind speed or low wind speed, which demonstrates a distinct difference between days with similar noise level characteristics and different noise level characteristics. This classifier was optimized in the 200-400 Hz range, where the statistics for the dataset for accuracy, precision, recall, and F-value were 0.6302, 0.6317, 0.9936, and 0.7706, respectively, with a windy vs. non-windy threshold value set at 65 dB re 1 muPa/square root of Hz. We found that for the chosen time duration of analysis window, the classifier had similar performance regardless of method.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 2022
- Accession Number
- AD1173481
Entities
People
- Alexandra M. Smith
Organizations
- Naval Postgraduate School