Automatic Waveform Quality Control for Surface Waves Using Machine Learning Techniques

Date: 4/24/2019

Time: 06:00 PM

Room: Grand Ballroom

Large seismic waveform data sets are common as a result of additional seismic station and seismic network deployments grow in frequency and size. Unfortunately, not all waveforms have sufficient signal-to-noise characteristics and contribute useful information when employed in specific analysis techniques. For this reason, waveform data are usually examined before they are included in analyses sensitive to noisy signals. Quality control of waveform data, however, requires substantial time and effort, that can be a significant burden with large datasets. In some cases, data quality control becomes the most time-consuming part of the analysis. We screened roughly 400,000 surface-wave waveforms to assign quality levels (A, B, C etc.) to each waveform as part of efforts to improve earthquake locations in remote regions. The complexity of surface wave signals makes reliable automation of the quality control process a challenge. In this study we describe how using machine learning algorithms such as logistic regression, support vector machine, K-nearest neighbors, random forest, and neural networks, we are able to achieve a test accuracy of over 90 percent using subsets of labeled waveforms that we gathered from previous studies. We also developed interactive visualization tools to examine waveforms the algorithms are not able to categorize correctly. Preliminary results show that many of the incorrectly categorized signals are in fact human errors (wrong labels) in the training and assessment data. The resulting machine learning models can be used to assess signal quality for investigations involving surface-wave waveforms.

Presenting Author: Chengping Chai

Authors