An interpretable rockburst prediction model based on SSA-CatBoost

Abstract

Rockburst is a common geological hazard in deep geotechnical engineering. In this study, the categorical boosting (CatBoost) was used to classify rockburst intensity, and the weight and threshold parameters of CatBoost was optimized by the sparrow search algorithm (SSA). Then, a database containing 357 rockburst cases from engineering projects around the world was established, and six features was selected as inputs for the model: the maximum tangential stress of rock σ_θ; the uniaxial compressive strength of rock σ_c; the tensile strength of rock σ_t; the stress ratio of rock σ_θ/σ_c; the brittleness ratio of rock σ_c/σ_t; and the elastic strain energy index of rock W_et. In the performance test of SSA-CatBoost, it reached an accuracy of 92.96% on the independent test set, and the prediction accuracy was improved by 16.04 percentage points compared to CatBoost. The SSA-CatBoost has better accuracy and generalization capabilities when compared to the existing rockburst empirical criteria and other machine learning models. The accuracy of the SSA-CatBoost reached 100% when the model was validated in three actual projects. In addition, interpretability analysis was conducted on the output results of the SSA-CatBoost model by the shapley additive explanations (SHAP). The input feature W_et that has the greatest impact on the model was identified, and the facilitating or inhibiting effects of different features on the output results are revealed. Research results indicate that the model can predict the levels of rockburst in geotechnical engineering.