The study investigates the application of machine learning techniques for classifying rockbursts among non-destructive tremors recorded in the Rydułtowy part of the ROW hard coal mine in the Upper Silesian Coal Basin, Poland. The mining environment is dominated by ultra-thick, high-strength sandstone strata, which significantly increase the likelihood of high-energy tremors. The interaction of geological/geomechanical, mining, technical/technological, and seismic factors is highly nonlinear, rendering deterministic analytical approaches insufficient for reliable rockburst identification. A dataset comprising 99 records, including 16 dynamic phenomena, was divided into training and testing subsets, with 75% of the data used to evaluate the discriminative power of the input variables and to train the machine learning models. Three parameters consistently exhibit the highest predictive relevance: peak particle velocity, seismic energy, and the rock mass bursting tendency index. Ten machine learning classifiers were evaluated using stratified 10-fold cross-validation. Ensemble-based models—particularly XGBoost, AdaBoost and Random Forest—demonstrated the most stable and accurate performance. The results indicate that machine learning models provide an effective computational framework for supporting rockburst hazard assessment in geologically complex mining conditions associated with ultra-thick sandstone strata.
Wojtecki et al. (Thu,) studied this question.