Envelope-Based, Real-Time Nested Grid Search: Estimates for Earthquake Early Warning

Date: 4/25/2019

Time: 06:00 PM

Room: Grand Ballroom

One popular EEW algorithm that has been tested in real-time is the Virtual Seismologist. It uses time frequency evolution of ground motion envelopes to provide estimates for earthquake source parameters. The goal of this method is to mimic the analysis that a human seismologist would perform in a reduced amount of time. With this goal in mind, we develop an envelope-based, real-time algorithm that will rapidly recognize incoming ground motion envelopes and assign the most probable magnitude, location, and origin time estimates to the observed earthquake. For our algorithm, we take on two approaches: (i) a standard grid search that will find the approximate parameters that best describe the incoming envelopes, and (ii) a historical search of envelopes from specific past earthquakes.

The first approach is the standard grid search. The grids are predetermined based on the first-triggered station: magnitude, latitude, longitude, and origin time. Observed P- and S-wave envelopes are created using polarization analysis (Ross et al 2014) for identification of phase separation. These observed envelopes are compared with the predicted Cua-Heaton (CH) ground motion envelopes. Source parameters are described probabilistically using the uncertainties of the CH ground motion prediction equations (GMPEs), and the grid points that maximize the likelihood are determined.

In addition to the grid search, we have the historical search as a second approach. Its intention is to find envelopes that match past events. It will check the outcome of the initial grid search, or, if available, find better estimates. To reduce searching time, many services, like GoogleMaps, use tree data structures. Similarly, we apply the kD tree search. We create a catalog of source parameters and their corresponding ground motion peak values. The traditional brute force approach would search through the whole catalog, but we construct a kD tree and search only a subset of it. Both approaches are done in real-time, updating the estimates with additional data from additional stations.

Presenting Author: Becky H. Roh

Authors