Improving failure modeling for gas transmission pipelines: A survival analysis and machine learning integrated approach

This study proposes a methodology to model gas transmission pipeline failures using historical pipeline failure data. Censoring occurs frequently in the dataset, and overlooking it may lead to biased predictions. To address this issue, the statistical Cox model, a survival analysis and machine learning integrated model, i.e., RSF, are introduced in this study, along with other machine learning models (ANN, SVR, RF, and XGBoost), primarily for comparison. The Cox and RSF models provide insights into the influence of covariates on pipeline failure, informing decisions regarding pipeline construction, inspection, and maintenance activities. The findings indicate that the statistical Cox model overestimates failure age due to its limited ability to capture failure nonlinearity, while other machine learning models underestimate failure age because they cannot handle dataset censoring. In contrast, the survival analysis integrated machine learning method, RSF, outperforms other methods for modeling gas pipeline failures. The findings have practical implications for effectively managing reliability and mitigating risks associated with gas transmission pipelines to ensure safety. Moreover, the proposed methodology can potentially be applied to other pipeline systems and various types of systems, provided certain requirements are met.