Novel stacking models based on SMOTE for the prediction of rockburst grades at four deep gold mines

Abstract

Rockburst is a frequently encountered hazard during the production of deep gold mines. Accurate prediction of rockburst is an important measure to prevent rockburst in gold mines. This study considers seven indicators to evaluate rockburst at four deep gold mines. Field research and rock tests were performed at two gold mines in China to collect these seven indicators and rockburst cases. The collected database was oversampled by the synthetic minority oversampling technique (SMOTE) to balance the categories of rockburst datasets. Stacking models combining tree-based models and logistic regression (LR) were established by the balanced database. Rockburst datasets from another two deep gold mines were implemented to verify the applicability of the predictive models. The stacking model combining extremely randomized trees and LR based on SMOTE (SMOTE-ERT-LR) was the best model, and it obtained a training accuracy of 100% and an evaluation accuracy of 100%. Moreover, model evaluation suggested that SMOTE can enhance the prediction performance for weak rockburst, thereby improving the overall performance. Finally, sensitivity analysis was performed for SMOTE-ERT-LR. The results indicated that the SMOTE-ERT-LR model can achieve satisfactory performance when only depth, maximum tangential stress index, and linear elastic energy index were available.