Multi-Source Sensing and Interpretable Ensemble Learning for Cross-Validated Performance-Weighted Evaluation of Open-Pit Blasting Performance

Comprehensive evaluation of open-pit blasting is challenging because performance indicators originate from different sensing sources and are predicted with different levels of model performance. This study proposes a multi-source sensing and interpretable ensemble learning framework for model-performance-weighted blast-performance evaluation. Data from 70 production blasts were organized into 12 input variables and 10 indicators spanning fragmentation quality, muckpile morphology, safety and adverse effects, and operational efficiency. Random forest, XGBoost, LightGBM, and CatBoost models were developed separately for each indicator and compared using five-fold cross-validation, while SHAP was applied to interpret the selected models. The indicator-specific models achieved cross-validated R2 values of 0.7812–0.9012; these are selection-conditioned cross-validated estimates obtained under the same five-fold partition that was used to select the model for each indicator, and they are not unbiased estimates of generalization performance. Powder factor and uniaxial compressive strength were influential for fragmentation, whereas maximum charge per delay strongly affected peak particle velocity and muckpile displacement. Cross-validated R2 was then incorporated into AHP–entropy weighting as a relative model-performance coefficient that is not a confidence level, a probability of correctness, or an uncertainty estimate. In field application, engineering adjustment based on model feedback increased the comprehensive score from 65.4 to 87.5, improved the grade from III to II, and reduced D80 and PPV by 25.4% and 27.8%, respectively. The models are specific to the monitored mine rather than transferable, and the field implementation is an application case rather than an independent external validation. The framework provides an interpretable and model-performance-aware basis for site-specific blast-scheme comparison and iterative improvement.