Prediction model for indoor rock compression failure time based on ensemble learning and optimization algorithms

The prediction of rock failure, a key fundamental research for addressing mining safety issues (such as mine slope stability and rockburst), faces challenges with traditional methods due to their complex generalization and computational processes that struggle to describe the entire failure process. Consequently, 12 prediction models integrating ensemble learning and optimization algorithms were established to predict rock peak stress and failure time using strain, elastic modulus, density, mass, and confining pressure as inputs. Five-fold cross-validation was used to optimize hyperparameters, significantly improving the model’s generalization ability, robustness, and stability. Dataset was established through rock mechanics experiments, with strain increments configured at 0.008‰, 0.01‰, and 0.012‰ in the test set. The Cross-Validation optimized Particle Swarm Optimization eXtreme Gradient Boosting (CV-PSO-XGBoost) model performed best-under a strain increment of 0.01‰, its stress prediction achieved coefficient of determination R² = 0.904, mean absolute error (MAE) = 4.315, root mean square error (RMSE) = 5.435; while the failure time prediction demonstrated R² = 0.811, mean absolute percentage error (MAPE) = 7.842%, MAE = 30.343. Finally, SHapley Additive exPlanations (SHAP) analysis showed strain and stress significantly impact the model, with strain positively predicting failure time, aligning with traditional rock failure models, validating reliability. This study provides insights into the research on rock strata stability in mining.