Simultaneous damage diagnosis of casing using interpretable multi-label classification

Casing damage poses a significant challenge, severely compromising well integrity and production efficiency in oilfields. However, most existing studies typically focus on predicting only a single type of casing damage, thereby overlooking the complex reality in which multiple damage types, such as buckling, shrinkage deformation, and corrosion perforation, can simultaneously occur under actual field conditions. To address this limitation, this study innovatively treated casing damage detection as a multi-label classification problem and proposed a binary relevance based on optimal base classifier combination (BR-OC) model. The proposed model significantly improved the multi-label prediction performance by integrating the advantages of individual base classifier models and optimizing the classifier combination. Initially, a comprehensive multi-label dataset was constructed by incorporating four critical categories of influencing factors: geological, engineering, development, and corrosion, resulting in 32 input variables and three output labels. Subsequently, seven machine learning models were evaluated and strategically combined on a dataset containing 284 wells, yielding an optimal multi-label classification model (MLC_XGB-LGBM-DT). The proposed model notably outperforms traditional binary relevance algorithms and other established multi-label classification methods, achieving a Hamming loss of 0.111, micro-recall of 90.5%, micro-F1 score of 0.8, and subset accuracy of 75.4% on the test set. Finally, a comprehensive model analysis workflow was developed, featuring Shapley additive explanations (SHAP)-based interpretability, sensitivity analysis, and the optimization of controllable factors. The integration of interpretability and optimization techniques provides valuable insights into engineering practices, facilitating the identification of controllable parameters to effectively mitigate casing damage risks.