APPLICATION OF MACHINE LEARNING TO THE PREDICTION OF WAVE VELOCITY IN A GIVEN MINE GROUND CONDITION

Abstract

This thesis explores the use of machine learning (ML) to predict wave velocities in mining environments, aiming to improve mining safety by reducing seismic risks like rockbursts. It challenges traditional, less accurate methods with an innovative approach that combines laboratory models and ML algorithms for more precise predictions. The study constructs physical models to replicate mine conditions and generate data for training ML models, from simple linear regression to complex deep neural networks. In a comprehensive analysis of predictive modeling techniques for seismic wave velocities, it was discovered that Linear Regression and Gradient Boosting outperformed, with an R-square value of 0.83, showcasing a balanced reduction in bias and variance. In contrast, the K-Nearest Neighbors (KNN) method's lower effectiveness implied that its proximity-based assumptions might be less relevant in seismic contexts, while the Deep Neural Network (DNN) model notably struggled, evidenced by a negative R-squared value of -0.81, wich is not possibble because Rsquare ranges between 0 and 1. It indicates substantial overfitting likely due to the complexity of the model and limited data. Among the models evaluated, Linear Regression emerged as the most fitting, owing to its simplicity, interpretability, and high accuracy, effectively avoiding overfitting and proving reliable for predicting seismic wave velocities. The findings advocate for future acquisition of more extensive datasets to potentially enhance the performance of complex models like the DNN, but within the current dataset's constraints, Linear Regression is identified as the superior predictive model for this purpose. Study firmly establishes ML's role in advancing seismic risk assessment in mining, opening avenues for future research in integrating ML with seismic data analysis.