Exploring Unsupervised Clustering of Seismic Noise Sources in Urban DAS Data: A Methodology Guide

Seismic measurements record the superposition of many seismic sources, with anthropogenic ones dominating frequencies above 1 Hz. While the anthropogenic seismic vibrations in urban areas are too small to influence daily human life, measurements in high precision physics experiments, such as those carried out at the Deutsche Elektronen-Synchrotron (DESY) particle accelerators in Hamburg can be negatively influenced by these vibrations. To gain insight into the seismic wavefield at DESY, distributed acoustic sensing measurements were started in the WAVE initiative (www.wave-hamburg.eu). The goal of this study is to utilize unsupervised machine learning tools to detect and identify different anthropogenic seismic noise sources. Two different approaches were tested: the seismic measurements are clustered using a temporal average of one second on time-frequency representations and a deep embedded clustering technique. For the first method, the clustering methods fuzzy-c-means, Gaussian mixture model (GMM), hierarchical clustering and hierarchical density-based spatial clustering of applications with noise (HDBSCAN) were used. The clustering performance of all methods was compared using car signals on a short DAS fiber section as our ground truth data. Furthermore, the usage of spectrograms and continuous wavelet transforms was compared on the ground truth data set, with the continuous wavelet transform giving better results. In a next step, the best-performing clustering methods GMM and HDBSCAN of the temporal average and deep embedded clustering were applied to the entire 12 km fiber to cluster seismic noise sources. Based on the results, the respective advantages and disadvantages of the different approaches were determined. The study was concluded with a "recipe'' on how to approach unseen DAS data based on scientific objectives and physical properties of interest, paving the way for an optimized DAS data analysis.