Shakealert Testing and Certification: Future Developments

Date: 4/25/2019

Time: 06:00 PM

Room: Grand Ballroom

Earthquake early warning algorithms must undergo rigorous real-time and offline testing before being accepted into the ShakeAlert production system.  The ShakeAlert Testing and Certification platform attempts to simulate how the production system will perform by deploying the proposed algorithms, configurations, earthworm rings/modules, and ActiveMQ messaging on separate, nearly identical hardware.  For real-time testing, the test servers are fed the same West Coast input waveform data as the production servers, and the algorithms’ performance is compared between the test and existing production systems after a minimum of two weeks.  For offline testing, four simultaneous instances of the algorithms are run on one machine with a historic test suite of earthworm tankplayer files (format used for replaying data into an earthworm ring with realistic timing).  These tankplayer files include data from West Coast earthquakes (M4.1 to M7.2), teleseisms, regional events, and problematic data sets. 

There are several areas of ongoing development in ShakeAlert testing and certification intended to extend and improve its capabilities.  One key area is assessing the performance of the system for large magnitude events.  These events are infrequent on the West Coast, so to augment our large magnitude data set, M5+ events from Japan are being incorporated. This will include the 2011 M9.0 Tohoku-Oki and 2016 M7.0 Kumamoto earthquakes.  Evaluation of large magnitude event alert timelines will also be developed and implemented to help quantify ShakeAlert warning times.  Another key area of development is refining how ground motion predictions are assessed for different algorithm results.  Code is already in place for classifying alerts into True Positives, False Positives, True Negatives, and False Negatives; as well as for calculating overall quality factors for a grid of points where the alerting threshold and target threshold are equal and the minimum alert time is >0 seconds.  Additional configurable features, are being added to make the calculations more flexible.

Presenting Author: Deborah E. Smith

Authors