This study was undertaken with the goal of mitigating operational risks in mining arising from the ore grade estimation process. The objective was to compare various methods and introduce an appropriate approach with minimal error for estimating the iron grade in the Sheytoor deposit.The Sheytoor iron ore deposit is located in Yazd province, 78 kilometers from the city of Bafq, and 10 kilometers southeast of the Gazestan village. The dataset used in this research includes 57 boreholes with a total length of 15,104.7 meters. The total number of samples analyzed for iron grade is 1,552 with the average grade of 34.29%, in the range of 1.1 (minimum iron grade) and 64.5 (maximum iron grade). Each sample is characterized by four parameters, including location (Easting, Northing and depth), and iron grade. Before performing the estimation process, pre-processing steps were taken, including core samples compositing, outlier detection and removal, examining out-of-range data, identifying elliptical ore grade heterogeneity, and determining threshold limits for kriging. The four methods used for grade estimation, include ordinary kriging, indicator kriging, multilayer perceptron neural networks, and support vector regression. To compare the performance of employed estimation methods, the accuracy of each method on test data were computed.Validation results, revealed that the multilayer perceptron (MLP) method has the best performance in the testing phase with an RMSE value of 0.151. Furthermore, for additional validation analysis, the method of cross-validation was employed. Two parameters, the coefficient of determination in regression and the RMSE value, were eva‎luated as performance criteria for each method in cross-validation. Based on the coefficient of determination in regression and RMSE, the MLP method with a regression coefficient of 0.78 and an RMSE value of 0.151 demonstrated the best performance among all the methods. For a more comprehensive comparison, histograms and statistical information for each method before and after estimation were presented for the test data. Based on the analyses of these plots and derived descriptive statistics, the multilayer perceptron neural network method exhibited the best performance compared to other methods used in this study.

نادر فتحيان پور

احمدرضا مختاري , فرهاد محمدتراب