Classification of Urban Seismic Noise Using Unsupervised Machine Learning

Session: Recent Development in Ultra-Dense Seismic Arrays With Nodes and Distributed Acoustic Sensing (DAS)
Type: Oral
Date: 4/30/2020
Time: 04:45 PM
Room: 110 + 140
Description: 

Ambient seismic noise is used extensively in seismic imaging, but few studies have identified distinguishing features and physical sources of cultural noise, a class of noise induced by human activity. In this work, we examine time and frequency domain features of urban cultural noise from a spatially dense array in Long Beach, California. We use 161 hours of recordings from the array’s 5200 geophones, which exhibit emergent and impulsive signals due to extensive cultural noise. We use convolutional autoencoders, a class of machine learning techniques, to learn latent features from spectrograms of the data. The features are used to develop clustering models that differentiate noise sources into separable classes. These classes may reflect the physical source at which the noise signal was generated.  Known sources of cultural  noise in the Long Beach Dataset include: Vibroseis truck signals, air traffic, car traffic, industrial machinery, etc. Without labeled examples of each type of noise signal, the task of classification becomes an usupervised clustering problem. We demonstrate that autoencoder based feature extraction and deep embedded clustering models provide a promising method to classify cultural noise signals and identify their origins.

Presenting Author: Dylan B. Snover

Authors