Exploration of Data Interpretation Methods and Machine Learning–Based Failure Prediction Models for Plastic Gas Distribution Pipelines

Gas pipelines are a vital part of North America’s energy infrastructure, and analyzing reported incidents is essential for assessing their reliability. Previous studies lack comprehensive data interpretation of contributing factors in pipeline failures. This study addresses that gap by evaluating incident data from the Pipeline and Hazardous Materials Safety Administration for gas distribution systems. The data set was analyzed using multiple correspondence analysis (MCA) and clustering, revealing patterns such as mechanical puncture as the predominant failure mode in excavation damage, equipment failure, and incorrect operation typically causing rupture in younger pipes and pipe or welding material failure leading to leaks in older pipes. A predictive model based on extreme gradient boosting using MCA-transformed features achieved 72% accuracy and a Cohen’s Kappa of 0.55. The model shows practical potential for identifying high-risk segments and supporting asset management for plastic gas pipelines, offering valuable tools for utility owners and consultants.