A Machine Learning-based Microseismic Event Location and Wave Velocity Prediction

Human-induced seismicity from underground mining can have a huge effect on the safety, productivity, and operating costs of mines. Hence, the mining industry demands an accurate predictor for microseismic source locations in order to mitigate unexpected catastrophes such as rockbursts. The aim of the research is to use machine learning models to predict real-time seismic wave velocities in the constantly degrading underground mine environment. In underground mining, the rockmasses and voids are constantly changing in quality, geometry, and content with mining. Data for this study is generated by a research team at the laboratory of the School of Mining and Geosciences at Nazarbayev University. The research on the application of machine learning is also conducted at the School of Engineering and Digital Sciences in the same university. The main findings of the data analysis were that due to the heterogeneity and constant change of the mining environment, the seismic wave velocity is not constant, hence the traditional methods of calculating the event source location using a constant velocity model in the seismic monitoring systems will often lead to inaccurate results. The accuracy of determining seismic event source locations depends on the seismic wave velocity assumed in the algorithm for the event source location. Thus, the ability to predict real-time seismic wave velocity allows the better location of seismic events. Several machine learning approaches have been explored for the prediction of seismic wave velocities under various conditions. Linear Regression models were trained from the data obtained in the laboratory to predict the wave velocity. The deep learning approach was also used to achieve higher precision, in particular, Deep Artificial Neural Networks (ANN) have been applied. Compared with Linear Regression which could only learn linear relationships between the features and the target, Deep ANN could learn complex relations due to the activation function in each layer.