A DEEP LEARNING-BASED SEISMIC EVENT WAVE VELOCITY COMPUTATION AND SOURCE LOCATION DETERMINATION

Abstract

The mining industry demands an accurate predictor for microseismic event location to avoid widespread disasters such as rock burst which are induced by the human activity or natural disasters. The purpose of the thesis is the deep learning application to calculate the microseismic location and by tracking the velocity. The data used in the project has two origins: STanford EArthquake Dataset (STEAD) and the School of Mining and Geology lab generated data extended by the synthetic dataset generated using the Fast Marching Method (FMM) mimicking the lab experiments. The result of the thesis is the velocity approximation resulting in better microseismic event location. Multiple deep learning approaches were explored for improved seismic event source location. The frequency-time spectrum information obtained after application of the continuous wavelet transform (CWT) / short-time Fourier transform (STFT) of the single receiver input was fed into the convolutional neural network (CNN). The other two models (CNN and temporal convolutional network (TCN)) use multi-receiver raw waveform input. It is demonstrated that the proposed approach with the TCN results in more accurate microseismic source identification compared to other tested methods. The lab generated data together with FMM based synthetic dataset allowed to build a model for velocity prediction.