SCEC Award Number 19177 View PDF
Proposal Category Individual Proposal (Integration and Theory)
Proposal Title Machine Learning Based Tremor Detection in Central and Southern California
Investigator(s)
Name Organization
Zhigang Peng Georgia Institute of Technology
Other Participants Yuling Chuang
Chenyu Li
Lijun Zhu
SCEC Priorities 1e, 3a, 3d SCEC Groups Seismology, CS, FARM
Report Due Date 04/30/2020 Date Report Submitted 11/16/2020
Project Abstract
The increasing volume of seismic data from long-term continuous monitoring motivates the development of algorithms based on convolutional neural network (CNN) for faster and more reliable phase detection and picking (e.g., Ross et al., 2018; Zhu and Beroza, 2019; Zhu et al., 2019). Inspired by the successes of the CNN approach on seismic phase picking, we developed a single station, CNN-based tremor classifier to separate tremor and noises (non-tremor) signals. We first trained the CNN model using a Parkfield dataset that composes 326,904 labelled tremor and noise data. The CNN model reached 93% training accuracy with 0.975 AUC value, suggesting higher ability on separating tremor and noise signals when compare to previously published CNN model. We further applied the classifier to pre-documented tremor and train noise signals in southern California’s Anza area in an effort to help clarifying the signal types. The results suggest that these signals are less likely to be events that are similar to Parkfield tremors. We also tested the model on five stations with tremor recordings in the year of 2016 in Taiwan. The test result suggests the CNN model can successfully recognize tremor signals at some stations with high accuracy. This suggests the CNN model has leaned a generic representation of tremors to a certain degree. However, the predictability is low at some stations, which suggests in-situ noises may significantly distort tremor signals and therefore bias the classification results.
Intellectual Merit This project develops a deep-learning based tool for tremor and noise classification. The model trained on Parkfield tremors has learned the generic representation of tremors, and showed the ability of recognizing tremors is other areas such as Taiwan. In an effort to clarify true event types of the pre-documented train and tremor signals, our current model predicts that they are less likely to be tremors. The model we developed in this project also has higher performance as a classifier, compare to other previously published CNN model for tremor/noise classification. By using shorter waveforms as inputs instead of long spectrograms, it also has higher potentials to turn the model into a real-time application.
Broader Impacts This project supported collaborations of two GT students. Lindsay Chuang and Cenyu Li from School of Earth and Atmospheric Sciences (EAS). Lindsay Chuang is a 3nd year Ph.D. student, and Chenyu Li is a 6th year Ph.D. student graduated in summer 2020. This project will be one component of their Ph.D. theses.
Exemplary Figure Figure 1. (a) The distribution of tremor and stations used in training (top left figure), and the Anza area (bottom right) where the model was applied to classify previous documented tremor and train noise. (b) A schematic view of the CNN model architecture.