Ensemble learning evaluation of mechanical property for mining waste cemented backfill

Abstract

Processing the large amount of waste rock generated from mining into backfill materials enables the resource utilization of waste materials. Additionally, the mechanical properties of backfill materials influence the stability of overlying strata. Therefore, this study established a large-scale dataset of coal gangue-based and tailing-based backfill materials. Three novel ensemble learning models were developed to evaluate the nonlinear effects of 29 dimensional factors on the mechanical properties of backfill material. Additionally, factors like material type, preparation method, and measurement error can cause discrepancies in mechanical properties, affecting evaluation accuracy. Consequently, this paper constructed SVR fusion models using three ensemble learning methods to enhance robustness on differentiated data. It investigated the impacts of ensemble methods, ensemble levels, and perturbation levels on the model. The study found that the Bagging-ensembled SVR performed optimally, with R2, MAE, and RMSE of 0.951, 0.369, and 0.560, respectively, exhibiting better robustness on perturbed datasets. This study contributes to resource utilization in mining waste.