Explainable Rockburst Prediction Under Class Imbalance: An MISSA-Stacking Ensemble Model
摘要
In response to the limitations of single models and traditional intelligent optimization algorithms, as well as the imbalance of sample categories dataset in rockburst prediction, this study selected the rock stress coefficient, σθ/σc (the ratio of the maximum tangential stress of the surrounding rock to uniaxial compressive strength), the brittleness coefficient, σc/σt (the ratio of uniaxial compressive strength to uniaxial tensile strength), and the elastic energy index (Wet) as prediction indicators and established a sample database consisting of 199 rockburst engineering cases. The borderline-SMOTE method was adopted to oversample the rockburst database. Based on this, five classification models—support vector machine (SVM), random forest (RF), multilayer perceptron (MLP), logistic regression (LR), and extreme gradient boosting (XGBoost)—were used as base learners with XGBoost serving as the meta-learner. The model’s hyperparameters were optimized using the multi-strategy improved sparrow search algorithm (MISSA), which brings about the establishment of the MISSA-stacking rockburst grade prediction model. Rockburst predictions were conducted on the balanced dataset processed by oversampling, and comparisons of the model with other models were made. The results showed that based on accuracy, F1 score, Kappa coefficient, and AUC, the MISSA-stacking model outperformed the other models. Simultaneously, the SHapley Additive exPlanations method was employed to interpret the prediction results, which determined the order of prediction indicator influence as σθ/σc > Wet > σc/σt. Finally, the model was applied to 11 engineering cases, which verified its practicality and reliability.