A Machine Learning Classification Approach to Geotechnical Characterization Using Measure-While-Drilling Data

Abstract

Bench-scale geotechnical characterization often suffers from high uncertainty, reducing confidence in geotechnical analysis on account of expensive resource development drilling and mapping. The Measure-While-Drilling (MWD) system uses sensors to collect the drilling data from open-pit blast hole drill rigs. Historically, the focus of MWD studies was on penetration rates to identify rock formations during drilling. This study explores the effectiveness of Artificial Intelligence (AI) classification models using MWD data to predict geotechnical categories, including stratigraphic unit, rock/soil strength, rock type, Geological Strength Index, and weathering properties. Feature importance algorithms, Minimum Redundancy Maximum Relevance and ReliefF, identified all MWD responses as influential, leading to their inclusion in Machine Learning (ML) models. ML algorithms tested included Decision Trees, Support Vector Machines (SVMs), Naive Bayes, Random Forests (RFs), K-Nearest Neighbors (KNNs), Linear Discriminant Analysis. KNN, SVMs, and RFs achieved up to 97% accuracy, outperforming other models. Prediction performance varied with class distribution, with balanced datasets showing wider accuracy ranges and skewed datasets achieving higher accuracies. The findings demonstrate a robust framework for applying AI to real-time orebody characterization, offering valuable insights for geotechnical engineers and geologists in improving orebody prediction and analysis