Rockburst multi-factor coupling prediction and quantitative analysis of influencing Factors’ SHAP values based on TPE-FDM-XGBoost model

Accurately predicting rockburst and clarifying its influencing factors is a crucial support for ensuring the safe and efficient construction of underground projects, deepening the theory of geological disasters, and promoting engineering technology innovation. It has an irreplaceable value in fields such as underground resource development and transportation infrastructure construction. This research conducts multi − dimensional explorations around rockburst prediction. Based on the XGBoost integrated learning algorithm, a TPE-FDM-XGBoost model is constructed, by combining feature derivation model (FDM) and tree-structured Parzen estimator (TPE), the rockburst grades of 301 rockburst events are predicted. At the same time, the Shapley Additive explained (SHAP) analysis is comprehensively used to reveal the feature contribution mechanism, and the model prediction logic and key factors are deeply analyzed. The results show that through FDM and TPE, the Accuracy and F1 Score of the TPE-FDM-XGBoost model for rockburst category prediction are 0.9836 and 0.9834 respectively, which are far better than 0.6066 and 0.6090 of the basic XGBoost model. The collaboration of FDM and TPE realizes the dual gains of data dimension increase and model tuning, significantly improving the model generalization ability and stability. Through collaborative FDM, a cross-scale prediction framework of “data driven-mechanism constrained” is constructed, upgrading single mechanical and energy features to multi-factor coupling indicators. The SHAP analysis clarifies that the σ θ stress, w et elastic energy index and their synergistic feature Feature3 (σ θ × w et ) are the key factors for prediction. The synergy of high σ θ and high w et is the key triggering factor for intense rockburst, while the combination of low feature values is an important basis for judging No rockburst. The quantitative analysis of SHAP values effectively captures the collaborative mechanism from rockburst stress triggering to energy release, clearly presenting the continuous evolution process of rockburst risk from stability to out-of-control driven by the change of feature values, connecting the data-driven model with the physical mechanism of rockburst occurrence. The combination of feature engineering and physical meaning is the key direction to further improve the model performance and interpretability, providing precise guidance for the accurate prevention and control of rockburst and practical engineering applications such as rockburst prevention and control in mines and tunnels.