Rock blasting fragmentation prediction based on feature engineering and interpretable machine learning

Accurate prediction of blasting fragmentation is crucial for optimising mining blasting parameters and controlling costs. Traditional empirical formulas fail for nonlinear relationships due to linear assumptions, whereas existing machine learning models suffer from empirical feature selection and locally optimal hyperparameters. This study proposes a GRA-HOA-LSBoost model integrating Grey Relational Analysis (GRA) for core feature selection, Hiking Optimisation Algorithm (HOA) for global hyperparameter optimisation, and Least Squares Boosting (LSBoost) for capturing nonlinear relationships via residual fitting. Validated with 110 measured datasets from 9 global mines, the model achieves test set metrics: MAE = 0.0363, MSE = 0.0027, RMSE = 0.0523, R2 = 0.9072, outperforming conventional machine learning models. This model equips engineers with a reliable quantitative tool to optimise blasting schemes efficiently under complex geological conditions. Ablation experiments confirm GRA-HOA synergistic effects; SHAP analysis reveals the model’s focus on blasting process parameters, and 5-fold cross-validation verifies its accuracy. The proposed model provides a scientific tool for engineering parameter regulation, with significant implications for improving blasting efficiency and reducing costs.