Machine Learning for Seismic Bump Prediction: A Comparative Study of Different Algorithms and Features

Predicting seismic events is crucial to preventing rock rupture events in coal mines. This endeavor involves analyzing long-term historical relationships between seismic data, or patterns and trends from the length of time series observations, to understand changing stress conditions and predict future events. In this exploration, the goal is to evaluate machine learning (ML) models’ efficiency in predicting seismic events from historical data. More specifically, two metaheuristic algorithms, Coati Optimization Algorithm (COA) and Dwarf Mongoose Optimization Algorithm (DMOA), are used to further improve the Random Forest Classification (RFC) scheme. These additions created two hybrid schemes: Random Forest with Coati Optimization (RFCO) and Random Forest with Dwarf Mongoose Optimization (RFDM). Drawing on the experimental outcomes, RFDM was superior to the RFC model, which achieved an accuracy of 0.940 at its maximum during the All phase, with 0.962 accuracy, 0.965 precision, and 0.963 recall. However, both models were improved upon by the RFCO model, with results of 0.990 accuracy, precision, and recall across all three metrics. The results support the proposed hybrid method and demonstrate how improved forecasting of seismic bumps can improve safety in mining operations. Improvements in early warning systems will reduce the risk of rock bursts, protect miner safety, and assist sustainability in coal mining.