پيشبيني نوع و پارامترهاي ژئومكانيكي سنگها با استفاده از پايش عمليات حفاري و بهرهگيري از روشهاي يادگيري ماشين

Given the critical importance of accurately characterizing rock geomechanical properties in the design an‎d execution of drilling an‎d mining operations, an‎d considering the time-consuming, costly, an‎d operational limitations of conventional laboratory testing methods, this study adopts machine learning as a fast an‎d data-driven approach for estimating these properties. The main objective is the prediction of rock mechanical properties, including uniaxial compressive strength (UCS), hardness (Schmidt), an‎d the abrasivity index (Cerchar), as well as the classification of rock types, based on drilling data. The model inputs consist of rotational speed (RPM), weight on bit (WOB), penetration rate (PR), an‎d load-on-bit electrical current (LOB(A)). To conduct this study, operational data related to the drilling process were extracted an‎d analyzed from the southern wall of the Sarcheshmeh copper mine, located in Rafsanjan County, Kerman Province. Concurrently with drilling operations, rock samples were collected, an‎d laboratory tests for determining uniaxial compressive strength, hardness, an‎d abrasivity were performed at the Department of Mining Engineering, Isfahan University of Technology. Drilling parameters were recorded during the drilling operations. By integrating laboratory measurements with field drilling data, a dataset comprising 2037 records was compiled. The data were first subjected to descriptive an‎d statistical analyses, an‎d correlations among parameters were eva‎luated. Outliers were removed using a interquartile range (IQR) method, an‎d the data were scaled through min–max normalization. For modeling purposes, XGBoost, Ran‎dom Forest, ANFIS, an‎d PSO-BP algorithms were employed. The results indicate that tree-based models, particularly XGBoost an‎d Ran‎dom Forest, outperform ANFIS an‎d PSO-BP in both geomechanical property prediction (with accuracies of 0.99 an‎d 0.95 for UCS prediction, respectively; 0.94 an‎d 0.83 for hardness prediction; an‎d 0.99 an‎d 0.97 for abrasivity prediction) an‎d rock-type classification (both with an accuracy of 0.98). Overall, the findings demonstrate that integrating drilling data with machine learning techniques provides a rapid, accurate, an‎d cost-effective alternative to laboratory testing, an‎d can play a significant role in optimizing drilling design, selec‎ting appropriate drilling tools, an‎d enabling intelligent control of the drilling process.

راحب باقرپور

مهربد خشوعي اصفهاني

محمود بهنيا , نادر فتحيان پور