XGBoost algorithm-based prediction of safety assessment for pipelines

Abstract

Pipeline safety is closely related to people's lives, environment and economic development. The traditional methods for pipeline inspection are laborious and very expensive, such as non-destructive examination (NDE) and engineering assessment (EA). In recent years, machine learning (ML) methods have been proven useful to solve pipeline assessment problems. These methods have mainly focused on a characteristic analysis of weld defects or pipeline life curves, and have not provided systematic and effective assessment methods for the pipeline industry. In this study, after comparing with other reference ML models, XGBoost algorithm-based predictive model is proposed for safety assessment of pipelines. First, two models based on XGBoost are established and tuned to create relationships between 9 parameters (e.g., internal diameters of pipelines) and NDE results/EA results. Then grid search and various statistical metrics (confusion matrices, AUC-ROC and confidence levels) are applied to optimize and assess these models. In addition, the prediction interpretability of XGBoost is analyzed to show its usefulness and applicability in industry. The results show that both of these two models satisfy the required predictive performance. The XGBoost model based on the EA results is finally selected thanks to its superiority by achieving an accuracy of 98.5%, AUC-ROC of 99% in testing, higher ratio of certain prediction (91.8%) and other metrics (precision of 98.8% and recall of 98.7%). In the analysis of model interpretations, a maximum operating pressure (MOP) is found to be the most important feature to affect the safety of pipelines. The results also show that when the current ML method based on the historical results of NDE and EA is established and proven stable, it can accurately predict the future risk of new pipelines in the absence of NDE and EA. In the studied research region, up to 91.8% of the funds and manpower spent on NDE and EA can be saved based on this ML method. Overall, the proposed XGBoost model provides an alternative to help industry evaluate pipeline welds with very few excavations and manual inspections. As more data from other regions is collected and used to train ML models, this method can be widely applied in industry.