Deep Learning for S-Wave Arrival Time Picking in Acoustic Emission Source Location Analysis
Qiquan Xiong, E. K. Johnson, C. S. Wu, Jesse Hampton · 2024
ABSTRACT: Traditional laboratory acoustic emission (AE) source location analysis predominantly relies on P-wave arrival times, limiting the number of locatable sources due to the requirement of four signals for solving four unknowns. This paper explores the feasibility of incorporating S-wave arrival times, which are often otherwise discarded, into source location analysis to reduce the fundamental requirement from four to two signals. We employ a U-Net deep learning model to pick S-wave arrival times on AE signals, trained using theoretical calculations based on P-wave and S-wave velocities. Despite only achieving an accuracy of 29.6% on the testing dataset, the model shows promising improvement towards the desired shape after only 100 epochs of training. Potential approaches for improvement, such as integrating error distributions into the analysis and employing a training→prediction→and manual picking cycle, are discussed. These findings suggest pathways for future work to enhance AE source location analysis through deep learning techniques, potentially unlocking new insights into rock fracture processes in laboratory settings. 1. INTRODUCTION Traditional laboratory acoustic emission (AE) source location analysis uses primarily P-wave arrival times for solving source locations [1-7] with a limited number of studies using S-wave arrival times. In the mathematical inverse process, the P-wave arrival times of AE signals and the coordinates of the sensor array are inputs to an optimization function. This optimization function is based on the time delay of wave propagation from source to sensor (Fig. 1(a)) and has four unknowns, i.e., three spatial coordinates and one source onset time. The source locations produced by the mathematical inverse process are also refined based on the physical principle [8] or mathematical (i.e., error or residual) features [9-11]. Upon this conventionally practiced procedure, at least four signals must be involved to solve the four unknowns [12]. Because energy releases of rock fractures are scale-invariant [13, 14], i.e., its frequency-magnitude distribution obeys the power law, sources with low magnitude that only trigger a few signal recordings are exponentially more prominent than those of greater magnitude. This leads to the number of locatable sources being one to three orders of magnitude lower than the total signals [15] with only a small percentage being used for source location [12, 16]. A majority of recorded signals will be discarded from the catalog and will not be considered in the subsequent advanced AE data analyses [17-21].