Live Lagged Correlation for Event Detection

Session: Waveform Cross-Correlation-Based Methods in Observational Seismology [Poster]
Type: Poster
Date: 4/30/2020
Time: 08:00 AM
Room: Ballroom
Description: 

We demonstrate a real-time earthquake early detection system based on the lagged correlation on live data. Live data are generated and transmitted without being stored. Arbitrarily Lagged correlation on live data across disparate signal sources can identify significant events at the arbitrary spatial location.

We consider computing lagged correlation on live data, which is challenging because the computation must be “in sync” for the detected events to be useful. Consider hundreds of sensors that can receive arbitrarily lagged, but correlated signals propagated through a media from an event. For example, when an earthquake happens, seismometers across the region of the earthquake (if not the world) will receive signals at varying times at varying speeds. In a monitoring system that absorbs live data from hundreds of sensors, computing lagged correlation becomes challenging because of high-speed data, a large number of stations and the necessity for accurate correlation values.

We propose a technique to compute lagged correlation on live data efficiently by combining two operations: filtering and cross-correlations. Due to the nature of the seismometer, the signal generated by other sources of vibrations will also be captured by seismometer and will cause inaccurate detection of earthquakes. Such noise is unavoidable, so the noise filtering is a necessary predecessor of any other analysis. Precisely, we achieve an order of magnitude speed-up by exploiting filters in correlation computation.

We demonstrate that our technique is suitable for live seismic monitoring and event detection. Correlation detectors are better than absolute-value (STA/LTA) based single station detectors in having low false positives.

In summary, with our techniques and the widely used messaging system Kafka, we built a working system to cross-correlate hundreds of streams in real-time at 100Hz speed using a conventional workstation. Besides, we can see from our visualization system in real-time where the event happened and how it spread across the seismometers.

Presenting Author: Sheng Zhong

Authors