ABSTRACT:

Locating seismic events in the field and acoustic emissions (AE) in the lab provides vital information about the spatial evolution of cracks during rock fracturing processes. Seismological source localization techniques require picking the P-wave arrival time of the received waveforms, velocity structure, and several stations to cover the event in a proper triangulation pattern. In this research, we first use the Akaike information criterion (AIC) to pick up the P-wave arrival time, and then using a constant velocity model and 8 nano-30 AE sensors we try to locate the spatial distribution of the individual microearthquakes from the minimization of the residuals. Supervised machine learning (ML) algorithms were trained and tested to the extracted features from the waveforms registered by the closest station to estimate the locations of the source events. The results have shown that ML methods can locate the source of an AE event using the waveform detected by a single station without any prior knowledge about the source, the medium that the signal passes, and the velocity model.

1. Introduction

Machine learning (ML) and deep learning (DL) have provided insight and solutions to some seemingly intractable problems which provides hope of finding approximate solutions to problems that are now deemed unsolvable like most of the problems encountered in seismology and earthquake studies (Mignan and Broccardo, 2019). Perhaps the most interesting benefit of ML in earth science studies is its ability to identify and extract new information from data as well as recognizing new patterns, structure, relationships in datasets that conventional techniques would not be able to reveal (Bergen et al., 2019).

Machine learning has been used for earthquake detection and location (Perol et al., 2018; Ross et al., 2018), detecting seismic arrivals (McBrearty et al., 2019), continuous monitoring of Cascadia subduction zone (Rouet-Leduc et al., 2019), detection of similarities between slow and fast slip events (Hulbert et al., 2019), and prediction of earthquake time to failure in the laboratory (Corbi et al., 2019; Rouet-Leduc et al., 2017). Despite promising results, there are still a lot of uncertainties related to the applicability of ML in resolving these problems. Lack of enough clean and reliable data could be one of the reasons. That's why running laboratory experiments under controlled conditions could be a very good beginning for verifying the available ML algorithms and developing new ones.

This content is only available via PDF.
You can access this article if you purchase or spend a download.