The rockburst hazard is a primary geological disaster endangering the environment in underground engineering. Due to the complexity of the rockburst mechanism, traditional methods are insufficient to predict the rockburst hazard objectively, especially when dealing with an imbalanced dataset. To address this issue, the hybrid models of PSO-BPNN-AdaBoost and PSO-BPNN-XGBoost were developed to predict rockburst hazards in this study. First, a rockburst dataset with 266 cases was constructed, containing six indicators: the maximum tangential stress, uniaxial compressive strength, uniaxial tensile strength, elastic deformation energy index, tangential stress index, and brittleness coefficient of strength. Then, the original dataset was oversampled using the synthetic minority oversampling technique (SMOTE) for dataset balancing. Subsequently, the PSO-BPNN-AdaBoost and PSO-BPNN-XGBoost models were constructed and evaluated to have the best accuracies of 0.901 and 0.851, respectively. Finally, the developed models were applied to predict the rockburst hazard in the Daxaingling Tunnel, the Cangling Tunnel, and the Zhongnanshan Tunnel shaft. The results indicate that the obtained rockburst hazard levels are consistent with engineering records, and the developed PSO-BPNN-AdaBoost and PSO-BPNN-XGBoost models are reliable for rockburst prediction.
Read full abstract