Using deep learning for phase detection and event location on hydraulic fracturing-induced seismicity

Student thesis: Master's ThesisMaster of Science by Research (MScR)

Abstract

Induced seismicity is a significant concern during fluid injection projects such as hydraulic fracturing for shale gas, enhanced geothermal systems and wastewater injection. With downhole microseismic monitoring, operators can obtain large seismic datasets to detect hydraulic fracturing induced seismicity (HFIS). Deep learning models like convolutional neural networks (CNNs) can offer rapid event detection in these large datasets. Rapid event detection can be useful for risk management strategies. CNNs have already displayed success in detecting regional earthquakes. Here, we examined whether a CNN pre-trained on regional earthquakes can also detect HFIS within high frequency continuous downhole data. We used data from a shale gas site at Preston New Road, UK, to assess the CNN model. The catalogue of the site, which contains over 23,000 events (-2.839 ≤ Mw ≤ 1.155), was generated using the coalescence microseismic mapping (CMM) method. Using confusion matrices, we evaluated the model’s ability to pick P and S-phases on single stations. To assess multi-station performance, we compared event catalogues and locations determined by the model and CMM method. We found that model performance declines with decreasing Mw. The model often misses small Mw < -2 events but detects new events not previously catalogued (230 new events within one hour). The model detected many new events during periods of high seismicity during injection. We infer that the CMM catalogue is more complete during less seismically active periods as the CNN model did not detect many new events. This study indicates that the pre-trained CNN offers the potential of detecting most events that the CMM detects (87.7%) in addition to more events during very active periods. The CNN produces these results more efficiently so it is promising, however, it requires further retraining with a dataset that represents the HFIS to improve phase detection and accurate picking.
Date of Award21 Jan 2021
Original languageEnglish
Awarding Institution
  • University of Bristol
SupervisorMax Werner (Supervisor)

Keywords

  • Earthquakes
  • Machine Learning
  • Seismology
  • Observational Seismology
  • Hydraulic Fracturing
  • Induced Seismicity

Cite this

'