The Similarity Matrix Profile, an efficient method for detecting both low and high signal to noise ratio seismic events in very long time series

Nader Shakibay Senobari, Gareth Funning, Zachary Zimmerman, Yan Zhu, & Eamonn Keogh

Submitted August 15, 2018, SCEC Contribution #8705, 2018 SCEC Annual Meeting Poster #063

We propose here an alternative, efficient method for seismic event detection for continuous data– the similarity Matrix Profile (MP). The MP is essentially a report of the index (i.e. location in the time series) and correlation coefficient (CC) value of the nearest neighbor for any subwindow of continuous data. Several algorithms for computing the MP efficiently on GPU cards and/or clusters have been developed in recent years so that it is now possible to calculate the MP for time series containing up to 1 billion data points (equivalent to 579 days of waveform data at 20 Hz). The MP is an efficient way of detecting both low and high signal to noise ratio events, as the nearest neighbor CC values for background noise tend to be relatively low. Therefore, the MP approach, with an appropriate threshold, can be used to detect earthquake swarms, aftershocks and foreshocks, low-frequency earthquakes (LFEs), repeating events and more.

We demonstrate the MP approach for two data sets, 580 days each of continuous recorded seismic data. Our first test is centered on the 2004 Mw 6.0 Parkfield earthquake using a nearby borehole station, We detect ~16 times more Parkfield aftershocks using the MP approach with a conservative detection criterion (CC>0.9); the number of detected events per day fits the Omori-Utsu law almost perfectly. Individual LFEs and non-volcanic tremor (NVT) are also detectable in the MP as their CC values are 0.8-0.9, compared with background noise (~0.6).we also find multiple new repeating families of seismic events that were activated 2-3 weeks prior to the mainshock. For our second test, we use data from a USArray station located near Mapleton, OR that is affected by NVT. Most events here with CC>0.9 are LFEs, based on their shapes, durations and frequency contents. We observe two periods of increased LFE activity in August to November 2006 and June to October 2007, broadly in agreement with the published literature. However our detection shows a more gradual increase in detected LFEs, rather than a sharp onset, as seen in that earlier studies. In all cases we observe abrupt changes in MP values associated with P-wave arrivals, suggesting that the method could be used to facilitate automated phase picking.

Key Words
Seismic data mining, foreshocks and aftershocks

Citation
Shakibay Senobari, N., Funning, G., Zimmerman, Z., Zhu, Y., & Keogh, E. (2018, 08). The Similarity Matrix Profile, an efficient method for detecting both low and high signal to noise ratio seismic events in very long time series. Poster Presentation at 2018 SCEC Annual Meeting.


Related Projects & Working Groups
Seismology