Toward Integrating a Machine Learning Signal Classifier in the ShakeAlert Earthquake Early Warning System

Abstract:

Machine learning techniques are rapidly gaining popularity as means to quickly and accurately classify large amounts of data. In the field of seismology, machine learning has been used in a wide range of applications from detecting tiny earthquakes (Rosset al., 2019) to denoising seismic signals (Zhu et al., 2019).

In their recent study, Meier et al. (2019) demonstrated the feasibility of using machine learning to distinguish between earthquake and spurious noisy signals with a high degree of accuracy. Here we apply a modified version of the machine learning model developed in that study to the EPIC real-time earthquake early warning algorithm in order to more accurately classify incoming triggers. EPIC is one of the two algorithms currently running on the ShakeAlert production system. Though recent improvements have reduced the number of false triggers coming into the system, there have still been 10 M4.0+ false alerts over the past year (2019-01-01 through 2019-12-31). By applying the machine learning methodology from Meier et al., (2019) we hope to drastically reduce the number of false triggers. Furthermore, with higher confidence in the accuracy of the EPIC triggers, we will be able to explore the possibility of detecting earthquakes with fewer stations (the algorithm currently requires 4 station triggers to create an alert).

Once this implementation proves successful, we will create a separate classifier waveform processor module that can be used by any of the existing or future ShakeAlert algorithms to classify incoming waveforms.

In this presentation, we will also highlight other new research related to the EPIC earthquake early warning algorithm and discuss its performance over the past year.