An Efficient Repeating Signal Detector that Uses Machine Learning to Improve Detection of Induced Seismicity

Abstract

Induced seismicity has become an increasing issue with the increase in both hydraulic fracturing and disposal of the leftover wastewater. Although regulatory management strategies such as traffic light systems aid in mitigating the impact of induced seismicity on society, they do not forecast when or where future cases of induced seismicity are likely to occur. Detecting earthquakes prior to reaching regulatory thresholds is a challenge for both operators and regulators, such that advanced methods for detection are needed. Previous methods to identify induced seismicity include Matched Filter Analysis (MFA; Caffagni et al., 2016), Fingerprint and Similarity Thresholding (FAST; Yoon et al., 2015) and Fast Matched Filtering (FMF; Beauce et al. 2017). While all these methods are effective, they require long computation times or heavy computational requirements and multiple sensors to produce viable results. We recently developed a computationally efficient Repeating Signal Detector (RSD), which utilizes a form of machine learning to identify similar waveforms in continuous seismic data using a single seismometer. Instead of relying on a priori templates, RSD identifies repeating signals above a specified signal-to-noise threshold and then grouping based on frequency and time domain characteristics, resulting in a significantly faster processing time than other current approaches. Recent work has also focused on distinguishing repetitive cultural noise from repetitive seismicity, as both types signals can be confused by automatic detection methods. We will discuss recent efforts to apply RSD in areas of induced seismicity that employ strategies to limit repetitive noise.

AAPG Datapages/Search and Discovery Article #90373 © 2019 AAPG Eastern Section Meeting, Energy from the Heartland, Columbus, Ohio, October 12-16, 2019