Ensemble Machine Learning Framework for Sewer Pipe Condition Assessment

This study presents an ensemble machine learning framework to support sewer pipe condition assessment and inspection prioritization. Using 4,802 closed-circuit television (CCTV) inspection records from a US municipality and associated pipe attributes, the framework predicts four ordinal condition classes ranging from good to severe. The proposed approach achieved an overall accuracy of 74.2%, with classwise receiver operating characteristic–area under curve (ROC–AUC) values between 0.82 and 0.96, indicating reliable discrimination across condition states. Model outputs are translated into a risk-ranked list of pipe segments to guide follow-up inspections and maintenance planning. To enhance transparency, feature-importance analysis was conducted, identifying pipe age, material type, and slope as the most influential predictors of deterioration. The framework supports proactive asset management by integrating predictive performance with interpretability, providing a scalable decision-support tool for optimizing inspection resources.