Probability Estimates of Short-Term Rockburst Risk with Ensemble Classifiers

Rockburst has become one of the most serious threats to the safety of workers, equipment, and excavations in deep underground engineering. However, short-term rockburst risk prediction remains an unsolved problem. This study aims to propose five ensemble classifiers to estimate the probability of short-term rockburst risk. These ensemble classifiers adopted the logistic regression, naive Bayes, Gaussian process, multilayer perceptron neural network, support vector machines, and decision tree as base learners, and used the average-based, accuracy-based, precision-based, recall-based and F₁-based combination rules, respectively. A total of 91 rockburst samples collected from the tunnels of Jinping II hydropower station, which included seven microseismic indicators, were used to verify the feasibility of the proposed ensemble classifiers. The comprehensive performance of each ensemble classifier was compared and evaluated using the accuracy and macro average of the precision, recall and F₁ metrics. In addition, the effects of different combinations of indicators on the prediction results were analyzed. Because of the favorable predictive performance, the proposed ensemble classifiers were applied to predict the short-term rockburst risk in different locations of the same project. The probability of each risk level was calculated, and then the final short-term rockburst risk was determined based on the highest probability. The results show that the comprehensive performance of the proposed ensemble classifiers is better than each base learner, and the accuracy-based and the F₁-based ensemble classifiers can be preferentially selected to predict the short-term rockburst risk. The highest accuracy and macro average of the precision, recall and F₁ metric values are 0.8667, 0.8901, 0.8661 and 0.8779, respectively. The proposed ensemble classifiers can provide valuable guidance for predicting the short-term rockburst risk.