Explainable Machine Learning for Well Integrity and Casing Failure Risk with Sparse Failure Labels

The integrity of the well is critical to efficient and safe oil and gas operations because few records exist of casing failure, and conditions in the downhole are complicated. In this study, an explainable machine learning (XML) framework is introduced to estimate the risk of casing failures in limited labelled wells. Which is based on measurements of pressure, temperature, annular pressure, cement bond quality and downhole vibration but integrates semi-supervised learning with gradient-boosted trees to predict the likelihood of failure. SHAP (Shapley Additive Explanations) values gives explanations on the feature level, and these values can be used to determine how operational parameters can contribute to risk predictions. Experimentation with synthetic and field-derived datasets shows that the XML framework has an accuracy of 91%, a precision of 0.87, a recall of 0.89, and a F1-score of 0.88. The critical predictors such as pressure differences in the annulus, low bond integrity of cement, and high vibration were quantitatively related with high risk of failures, whereas the risk scoring at the time gave a premonition to the known failures 2-6 months before the recorded failures. The improvement in the prediction performance and interpretability of the framework is evident when it is compared to the baseline models such as logistic regression and standard random forest. In low failure label environments, the technique fills the gap between risk assessment based on data and operational decision making by integrating the strong prediction with actionable information that enables operators to focus more on tracking, preventive actions, and to save on time on non-productive activities.