Machine-learning-based prediction of pipe-sticking events in petroleum drilling operations using domain-informed features and class-balanced ensembles

Abstract

Pipe sticking is a critical drilling failure that can increase the non-productive time, operational risk, and well construction costs. Reliable and early identification of pipe-sticking conditions is therefore important for improving drilling decision support while reducing the likelihood of severe downhole incidents. In this study, we developed and evaluated a machine-learning framework for binary classification of pipe-sticking events using a drilling dataset containing 297 observations and 13 original drilling-related variables. Domain-informed feature engineering was used to construct additional ratio, difference, and interaction variables representing the relationships among the geometric, density, friction, and torque-related parameters. Class imbalance was addressed using a custom minority-class interpolation procedure, followed by mutual-information-based feature selection. Eleven supervised machine-learning classifiers were then evaluated using stratified cross-validation, and the leading models were optimised using grid-search hyperparameter tuning before being combined using soft voting. The gradient boosting classifier achieved an accuracy of 92.09%, precision of 95.93%, recall of 91.54%, F1-score of 93.62%, and area under the receiver operating characteristic curve (ROC-AUC) of 96.60%. The soft-voting ensemble achieved an accuracy of 92.34% and ROC-AUC of 97.83%. Feature-importance analysis was used to identify the annular clearance, density, friction, and torque-related variables among the dominant predictors. However, the limited number of confirmed sticking events and the single-source dataset constrain the generalisability of the results. The proposed framework therefore represents a preliminary data-driven baseline for pipe-sticking classification rather than a field-validated deployment system.