A Machine Learning-Based Interpretable Glass-Box Model for Pipeline Failure Mode Prediction Considering Quantitative Coupling Factors

Occasional pipeline failure can lead to catastrophic consequences, as pipelines serve as critical components of infrastructure in the US. Thus, accurate prediction of pipeline failure modes is crucial to the pipeline system. Existing studies apply qualitative, quantitative, and semi-quantitative methods to identify key factors and assess pipeline vulnerabilities. However, the studies fail to adequately quantify the complex coupling effect between factors and their partial dependencies on failure outcomes. To address these gaps, this paper aims to analyze the influences of key contributing factors on various pipeline failure modes by quantifying the nonlinear relationships among factors and considering their interactive effects. A machine learning-based interpretable glass-box model is developed to predict pipeline failure mode through the XGBoost algorithm. Coupled factors are considered in the model through quantitative correlation analysis and XGBoost-based internal interpretable feature ranking methods. Historical crude oil pipeline incident data from the US Pipeline and Hazardous Materials Safety Administration (PHMSA) is selected to train and test the model. The results offer deeper insights into the underlying relationships among influencing factors and provide valuable guidance for future pipeline design, maintenance, and risk mitigation strategies, promoting a more resilient pipeline infrastructure system. Furthermore, the study contributes to the safety, long-term reliability, and economic stability of human society by ensuring the efficient delivery of essential resources.