Fault Time-to-Failure Forecasting using LSTM and Unsupervised Classification

Abstract:

When a rock is subjected to stress it deforms by creep mechanisms that include formation and slip on small-scale internal cracks. Intragranular cracks and slip along grain contacts release energy as elastic waves called acoustic emissions (AE). AEs are thought to contain predictive information that can be used for failure forecasting. Previously, we developed a method using unsupervised classification and a Long Short-Term Memory (LSTM) network to forecast labquakes using AE waveform features. Our data were generated in a laboratory setting using a biaxial shearing device and a granular fault gouge that mimics the conditions around tectonic faults. In particular, we analyzed the temporal evolution of AEs generated throughout several hundred laboratory earthquake cycles. We used a Conscience Self-Organizing Map (CSOM) to perform topologically ordered vector quantization based on waveform properties. The resulting map was used to interactively cluster AEs according to damage mechanism. We compared the AEs between two sensors in order to gain insights about cluster meaning. Finally, we used an LSTM network to test the predictive power of the AE clusters. By tracking cumulative waveform features over the seismic cycle, the network was able to forecast the time-to-failure (TTF) of the fault. However, we found that the network struggled to forecast TTF for unusually long or short seismic cycles. To address this, we further developed the learning methodology by replacing LSTM with an attention network, adding shear stress as a network input, creating a custom loss function, and incorporating a combination of meta networks and multitask learning. We also show preliminary results from using our forecasting method on aftershocks from the 2019 Ridgecrest sequence.