[Paper Review] Unsupervised classification of acoustic emissions from catalogs and fault time-to-failure prediction

Hope Jasperson, C. T. Bolton+3 more|arXiv (Cornell University)|Dec 1, 2019

Rock Mechanics and Modeling31 references5 citations

TL;DR

This study proposes an unsupervised machine learning framework to classify acoustic emissions (AEs) from laboratory-generated fault earthquakes and predict time-to-failure using waveform clustering and an event-based LSTM. By applying a Conscience Self-Organizing Map (CSOM) to extract damage mechanism clusters and training an LSTM on cumulative waveform features, the model achieves accurate prediction of fault failure time across seismic cycles.

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). Early research into AEs envisioned that these signals could be used in the future to predict rock falls, mine collapse, or even earthquakes. Today, nondestructive testing, a field of engineering, involves monitoring the spatio-temporal evolution of AEs with the goal of predicting time-to-failure for manufacturing tools and infrastructure. The monitoring process involves clustering AEs by damage mechanism (e.g. matrix cracking, delamination) to track changes within the material. In this study, we aim to adapt aspects of this process to the task of generalized earthquake prediction. Our data are 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 analyze the temporal evolution of AEs generated throughout several hundred laboratory earthquake cycles. We use a Conscience Self-Organizing Map (CSOM) to perform topologically ordered vector quantization based on waveform properties. The resulting map is used to interactively cluster AEs according to damage mechanism. Finally, we use an event-based LSTM network to test the predictive power of each cluster. By tracking cumulative waveform features over the seismic cycle, the network is able to forecast the time-to-failure of the fault.

Motivation & Objective

• To adapt nondestructive testing methods for acoustic emission (AE) clustering in rock mechanics to the challenge of generalized earthquake prediction.
• To identify distinct damage mechanisms in AE waveforms during repeated laboratory earthquake cycles using unsupervised learning.
• To evaluate the predictive power of AE clusters for forecasting fault failure time using deep learning.
• To establish a data-driven, topology-preserving method for AE classification that supports interpretability in failure mechanism tracking.

Proposed method

• Employing a biaxial shearing device with granular fault gouge to generate controlled laboratory earthquake cycles and collect AE waveforms.
• Applying a Conscience Self-Organizing Map (CSOM) to perform topologically ordered vector quantization of AE waveforms based on their features.
• Using the CSOM map to interactively cluster AEs according to underlying damage mechanisms such as matrix cracking or slip events.
• Extracting cumulative waveform features over the seismic cycle to represent temporal evolution of AE activity.
• Training an event-based Long Short-Term Memory (LSTM) network on clustered AE features to predict time-to-failure.
• Validating the predictive performance of each AE cluster using the LSTM model across multiple seismic cycles.

Experimental results

Research questions

• RQ1 Can unsupervised clustering of AE waveforms using CSOM effectively identify distinct damage mechanisms during fault slip cycles?
• RQ2 Which AE clusters exhibit the strongest temporal correlation with impending fault failure?
• RQ3 To what extent can cumulative AE features across a seismic cycle predict time-to-failure using an LSTM network?
• RQ4 How does the topological ordering of AE waveforms via CSOM improve interpretability and predictive accuracy compared to standard clustering?

Key findings

• The CSOM successfully produced a topologically ordered map of AE waveforms, enabling interpretable clustering by damage mechanism.
• Distinct AE clusters corresponded to specific failure-related processes such as crack nucleation and slip events.
• The LSTM model achieved high predictive accuracy in forecasting time-to-failure when trained on cumulative features from the most informative AE clusters.
• Temporal evolution of AE features within the seismic cycle showed consistent patterns preceding failure, which were captured by the LSTM.
• The predictive performance varied across clusters, with some clusters (e.g., those associated with critical damage stages) showing significantly stronger forecasting power.
• The integration of unsupervised clustering with deep learning enabled robust, data-driven time-to-failure prediction without requiring labeled failure events.