نوع مقاله : مقاله پژوهشی
نویسندگان
1 استادیار، گروه مهندسی معدن، واحد سیرجان، دانشگاه آزاد اسلامی، سیرجان، ایران
2 دانشجوی دکتری تخصصی، گروه مهندسی معدن، واحد سیرجان، دانشگاه آزاد اسلامی، سیرجان، ایران
3 سرپرست واحد عملیات، امور معدن شرکت معدنی و صنعتی گهرزمین، سیرجان، ایران
4 کارشناس واحد حفاری و انفجار، امور معدن شرکت معدنی و صنعتی گهرزمین، سیرجان، ایران
چکیده
کلیدواژه‌ها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Precise blast pattern design and the prediction of fragmented particle size play a key role in optimizing extraction performance and reducing operational costs in open-pit mines. The aim of this research was to develop an integrated framework for predicting fragmented particle size (D80) and refining blast design parameters at the Gohar-Zamin iron ore mine. To this end, using machine learning tools and hybrid models together with incorporating fuzzy inference based on data from 19 blasting events and data augmented via a random data augmentation technique, predictive models were developed. To prevent errors and bias in model learning caused by data leakage from related samples, modeling was performed using the Group K-Fold cross-validation method based on initial block identifiers, and data augmentation was applied only within the validation folds. Dozens of base, ensemble stacking, and weighted-average models were trained. Model evaluation results demonstrated that the stacking model (SVR-CatBoost) with R_CV^2= 0.7398, R_Full^2= 0.8678, R_Test^2= 0.8299, and an over fitting gap of 0.1281 offered the best balance between accuracy and stability, and was consequently selected as the final, stable, and reliable model. The weighted average model (SVR-CatBoost) ranked second, with an R_CV^2=0.7438 and an over fitting gap of 0.1187. Despite high R_CV^2values, some base models performed poorly on the test data, confirming the importance of simultaneous evaluation using test data. Additionally, the fuzzy model demonstrated poor performance, recording a coefficient of determination (R²) of 0.497. Based on the results of sensitivity analyses, using linear regression, SHAP, Sobol, and Morris methods as well as intelligent modeling, the blast design parameters for 22 test blocks were adjusted, with distinct modifications applied to the schist and conglomerate rock types. Accordingly, for schist waste rock, the crushing of every 100,000 tons of rock resulted in a 30% reduction in the number of blast holes and a 40% reduction in specific charge; drilling and blasting operating costs decreased by 7.5% for ANFO and 0.5% for Emulate. Similarly, for conglomerate waste rock, the number of holes by 23%, specific charge by 30%, and operating costs by an average of 22% per hole was reduced.
کلیدواژه‌ها [English]
)