Data-Driven Rock Strength Classification from Investigative Drilling Using Machine Learning

ABSTRACT: Investigative drilling (ID) is a variant of measurement while drilling (MWD) technique that has been utilized in many site investigation projects in Australia. This paper presents a data-driven approach for classifying rock strength using ID data. First, a dataset containing drilling parameters, including depth, penetration rate, percussion pressure, feed pressure, and rotation pressure, was constructed from the ID data of a transmission line access track project in soft rocks. Two types of rock strength grading - very low, low, and medium, as well as low and medium - were then assigned to the ID dataset. Five commonly used machine learning algorithms with default hyperparameters were tested and compared, with K-nearest neighbors (KNN) and Random Forest emerging as the most effective algorithms. Hyperparameter tunings were further conducted for the two algorithms. The results show that five- fold cross-validation accuracies of 79.58% and 91.54% were achieved for binary and three-class classifications using KNN, and 90.06% and 80.02% for Random Forest, respectively. Specifically, according to the confusion matrix of the three-class classification, rocks with low strength can be effectively differentiated, while rocks with very low strength are the least distinguishable. Finally, a feature importance analysis was conducted, revealing that the percussion pressure and penetration rate are the two most influential parameters, whereas the rotation pressure is the least important in classifying rock strength. This study demonstrates the potential of machine learning as a powerful tool for interpreting drilling data of ID or other MWD techniques, facilitating future development of reliable models for fast and cost-effective prediction of rock strength grades.