An Ensemble Deep Learning-Based Acoustic Emission Picking Model Reveals Migratory Foreshocks on Large-Scale Laboratory Fault

Abstract

Acoustic emissions (AEs) occurring during the frictional stick-slip experiment are crucial for predicting fault failure and understanding the nucleation mechanism of laboratory-scale earthquakes. However, detecting and picking their primary arrival phase (P-phase) in a complete manner remains a challenging task, especially when processing large data sets from a multi-channel continuous AE recording system throughout the laboratory earthquake cycles. In this study, we propose an Ensemble Deep Learning (EDL)-based supervised model named “AEbagging,” which comprises two individual feature extraction Deep Neural Networks serving as an encoder, and an EDL-based decoder for laboratory earthquake detection and phase picking. We conduct biaxial stick-slip experiment on a 0.85 m saw-cut granodiorite fault. By applying our model to the eight stick-slip events, we demonstrate its powerful capability in event detection and phase-picking. With the stacked pre-seismic AE sequences from the model-generated AE catalogs, we not only observe a precursory decrease in the $b$-value but also a migratory foreshock process from the lower right segment of the fault toward the upper left end during the nucleation phase. We discuss the possibility that such migratory foreshock activities are related to stress heterogeneities induced by fault surface roughness. This work demonstrates that the AEbagging model not only contributes to a better understanding of the spatiotemporal evolution of seismicity during laboratory fault instability, but also has potential for broad application in disciplines ranging from engineering to geophysics.