Characterizing Acoustic Signals and Searching for Precursors during the Laboratory Seismic Cycle Using Unsupervised Machine Learning

Recent work shows that machine learning (ML) can predict failure time and other aspects of laboratory earthquakes using the acoustic signal emanating from the fault zone. These approaches use supervised ML to construct a mapping between features of the acoustic signal and fault properties such as the instantaneous frictional state and time to failure. We build on this work by investigating the potential for unsupervised ML to identify patterns in the acoustic signal during the laboratory seismic cycle and precursors to labquakes. We use data from friction experiments showing repetitive stick-slip failure (the lab equivalent of earthquakes) conducted at constant normal stress (2.0 MPa) and constant shearing velocity (10 μm=s). Acoustic emission signals are recorded continuously throughout the experiment at 4MHz using broadband piezoceramic sensors. Statistical features of the acoustic signal are used with unsupervised ML clustering algorithms to identify patterns (clusters) within the data. We find consistent trends and systematic transitions in the ML clusters throughout the seismic cycle, including some evidence for precursors to labquakes. Further work is needed to connect the ML clustering patterns to physical mechanisms of failure and estimates of the time to failure. Supplemental Content: Figures and text that describe the statistical features, sensitivity analysis of the moving windows, effects of the bandwidth parameter, and additional clustering results. PRECURSORS TO EARTHQUAKES Earthquake forecasting is an important problem for mitigating seismic hazard, and it can help illuminate the physics of earthquake nucleation. Forecasts could be based on physical models of the nucleation process or changes in fault-zone properties (so-called precursors) before failure. However, with current monitoring techniques and models of earthquake nucleation, we are far from forecasting earthquakes or even identifying reliable precursors despite long-standing interests in the problem (Milne, 1899; Marzocchi, 2018) and a broad range of related and direct observations ranging from landslides (Poli, 2017), to glacial motion (e.g., Faillettaz et al., 2015, 2016), geochemical signals (Cui et al., 2017; Martinelli and Dadomo, 2017), geodesy (Chen et al., 2010; Xie et al., 2016; Moro et al., 2017), and seismology (Antonioli et al., 2005; Niu et al., 2008; Rivet et al., 2011; Bouchon et al., 2013). The situation is somewhat better for labquakes. Laboratory friction experiments coupled with ultrasonic measurements have been used to document the approach to failure (Scholz, 1968; Weeks et al., 1978; Chen et al., 1993), with important recent advances in documenting precursors based on spatiotemporal changes in rock properties before failure (Pyrak-Nolte, 2006; Mair et al., 2007; Goebel et al., 2013, 2015; Johnson et al., 2013; Kaproth and Marone, 2013; Hedayat et al., 2014; McLaskey and Lockner, 2014; Scuderi et al., 2016; Jiang et al., 2017; Rouet-Leduc et al., 2017, 2018; Hulbert et al., 2019; Renard et al., 2018; Rivière et al., 2018). Laboratory observations of precursors before earthquakelike failure encompass a variety of measurements, including high-resolution images that illuminate the failure nucleation process. These include passive measurements of acoustic emissions (AEs) (e.g., McLaskey and Lockner, 2014; Goebel et al., 2015), active measurements of fault-zone elastic properties (e.g., Scuderi et al., 2016; Tinti et al., 2016), and direct observations, using x-ray microtomography (micro-CT), of damage evolution in the failure zone (Renard et al., 2017). The microCT work reveals microfracture patterns and the interplay between shear deformation and local volume strain (Renard et al., 2017, 2018). The AE studies show that the Gutenberg–Richter b-value decreases systematically during the laboratory seismic cycle (Goebel et al., 2013; Rivière et al., 2018). In addition, active source measurements of elastic wavespeed and travel time show systematic changes throughout the laboratory seismic cycle and distinct precursors to failure for the complete spectrum of failure modes from slow to fast 1088 Seismological Research Letters Volume 90, Number 3 May/June 2019 doi: 10.1785/0220180367 Downloaded from https://pubs.geoscienceworld.org/ssa/srl/article-pdf/90/3/1088/4686471/srl-2018367.1.pdf by cjm38 on 03 May 2019 elastodynamic events (Kaproth and Marone, 2013; Scuderi et al., 2016; Tinti et al., 2016). These studies include measurements for dozens of repetitive stick-slip failure events showing that elastic wavespeed and transmitted amplitude increase during the linear-elastic loading stage and decrease during inelastic loading. MACHINE LEARNING AND ACOUSTIC SIGNALS BEFORE FAILURE Recent developments in the application of machine learning (ML) to seismic data suggest a number of possible benefits for seismic hazard analysis and earthquake prediction. One approach shows systematic changes in event occurrence patterns and seismic spectra that could illuminate the earthquake nucleation process (e.g., Holtzman et al., 2018; Wu et al., 2018). Another approach, using laboratory data similar to those that we focus on in this article, has shown that supervised ML can predict stick-slip frictional failure events—the lab equivalent of earthquakes (Rouet-Leduc et al., 2017). These works show that the timing of failure events can be predicted with fidelity using continuous records of the acoustic emissions generated within the fault zone (Rouet-Leduc et al., 2017, 2018; Hulbert et al., 2019). Stick-slip failure events are preceded by a cascade of microfailure events that radiate elastic energy in a manner that foretells catastrophic failure. Remarkably, this signal predicts the time of failure; the slip duration; and for some events, the magnitude of slip. However, successful implementation of a supervised ML algorithm demands access to a large labeled training dataset. Unsupervised ML offers an alternative approach that can be applied when labeled data are not available. The purpose of this article is to explore the application of unsupervised ML to characterize acoustic emissions during the laboratory seismic cycle and search for precursors to failure. This approach differs significantly from previous work using supervised ML in which statistical features are used to build a function that maps an input (statistics of the acoustic signal) to an output (e.g., time to failure). Supervised ML involves a training stage followed by a stage in which the algorithm is tested against new observations. In unsupervised ML, the task at hand is quite different. In our case, the goal is to find structure (clusters) within the seismic signal and track its evolution throughout the seismic cycle. Clusters are characterized and identified within an n-dimensional feature space via an ML clustering algorithm. We use a mean-shift ML clustering algorithm (Cheng, 1995; Comaniciu and Meer, 2002) to assess statistical features of the acoustic signal and compare our results with those obtained using the commonly used kmeans clustering algorithm (Tan et al., 2006). We apply both clustering algorithms to 43 statistical features after conducting a principal component analysis (PCA). For comparison to our previous work, we perform a second analysis using only the variance and kurtosis of the acoustic signal identified as the most significant features in the supervisedML analysis (Rouet-Leduc et al., 2017, 2018; Hulbert et al., 2019). That is, they improved the accuracy of the ML regression analysis the most out of ∼100 statistical features. Our goal is to assess how robust these features are when attempting to identify precursors to failure via unsupervised ML. We acknowledge that using results from a supervised ML study as inputs to an unsupervised ML analysis may violate the truly unsupervised nature of the analysis. However, we argue that this approach is well warranted because it can help connect unsupervised and supervised ML approaches. Our work has the potential to improve the understanding of laboratory precursors and ultimately to improve methods for seismic hazard analysis. FRICTION STICK-SLIP EXPERIMENTS We use data from frictional experiments conducted in a biaxial deformation apparatus (Fig. 1a) using the double-direct shear configuration (e.g., Rathbun and Marone, 2010). Two layers of simulated fault gouge are sheared simultaneously within three forcing blocks that contain grooves perpendicular to the shear direction to prevent shear at the layer boundary. The grooves are 0.8 mm deep and spaced every 1.0 mm. The initial gouge layer thickness is ∼5 mm, and the nominal contact area is 100 × 100 mm2. The center forcing block (15 cm) is longer than the side blocks (10 cm) so that the friction area remains constant during shear. Our experiment used glass beads with particle diameters in the 104to 149-μm range to simulate granular fault gouge (Anthony and Marone, 2005). The gouge layers are bounded by cellophane tape around the edges, and a thin rubber jacket is placed around the bottom half of the Horizontal DCDT Multichannel PZT Blocks Vertical DCDT (a)