Tremor or train? An attempt to discern tremor and noise using a deep convolutional neural network

Lindsay Chuang, Zhigang Peng, Lijun Zhu, & James H. McClellan

Submitted August 8, 2019, SCEC Contribution #9383, 2019 SCEC Annual Meeting Poster #073

Deep tectonic tremors are often characterized as non-impulsive, low signal-to-noise ratio, long-duration, and lacking in high frequency energy, which can be confused with background or anthropogenic noise. A standard technique to avoid misclassification of tremors and noise is to exanimate coherencies of seismic waves and propagation velocities in particular frequency bands (i.e. 2-8 Hz) recorded at multiple stations in a regional seismic network. However, this approach can fail when coherent anthropogenic sources (e.g. train, traffic, wind farm etc.) are also present and produce tremor-like signals. For example, a few weak tremor bursts have been reported along the Anza section of the San Jacinto Fault (SJF) in southern California (Hutchison and Ghosh, 2017). Whether those identified tremors are truly from tectonic sources, or from freight trains nearby (Inbal et al., 2018) is still in debate. Clarifying the source origin of these signals can help us better understand the characteristics of both tremors and noise as well as regional tectonics in the Anza area. In this study, we build a binary tremor/noise classifier based on a deep convolutional neural network (CNN), and use this classifier to categorize tremor-like signals in Anza. The CNN is composed of multiple convolutional layers with average pooling followed by two fully connected layers. The model is first trained on a ten-year history (2001-2011) of tremor and noise waveforms along the Parkfield-Cholame section of the San Andreas Fault which were recorded by the 13-station borehole High-Resolution Seismic Network (HRSN). The noise windows are picked from 1 to 6 minutes before reported tremor signals. All the data are broken down to one-minute-long non-overlapping segments with mean values removed, and filtered to 2-8 Hz. The whole dataset (~320,000 windows) is then used to train the CNN with an 80/20 training and validation random split. Training accuracy reaches 85% and 90% after 17 and 50 epochs, which is comparable to the single station accuracy (above 86.6%) achieved by KNN in other areas. Overfitting is observed after 50 epochs. We plan to incorporate residual layers as well as dropout layers in our CNN to improve the model performance and reduce overfitting. We expect that the CNN can learn the general features of tectonic tremors at Parkfield and eventually can be applied to distinguish between tremor and train signals in Anza as well as other areas.

Citation
Chuang, L., Peng, Z., Zhu, L., & McClellan, J. H. (2019, 08). Tremor or train? An attempt to discern tremor and noise using a deep convolutional neural network. Poster Presentation at 2019 SCEC Annual Meeting.


Related Projects & Working Groups
Seismology