Towards predictive modeling of time to sewer pipe failure : a preliminary exploratory study combining statistical analysis and AI techniques

Abstract

Urban sewer infrastructure is under increasing strain from population growth, network expansion, and aging. Consequently, pipe failures endanger public health and environmental safety. Traditional infrastructure management primarily relies on reactive maintenance. However, recent advances in predictive modeling have enabled proactive approaches to condition assessment. Most models assess current pipe conditions instead of predicting failure time, creating a significant gap in proactive maintenance planning. To address this gap, this preliminary study investigates sewer failure time prediction using statistical analysis and machine learning. This study uses sewer data from Hong Kong's Drainage Services Department, supplemented with parameters from open-source databases. Statistical analysis revealed distinct bimodal failure patterns, with peaks occurring at 30-40 years and 57-60 years for concrete pipes, and at 30-40 years and 52-60 years for vitrified clay pipes. Regional analysis further identified failure patterns across Hong Kong's major districts. The study also examined and compared two advanced machine learning models, namely Random Forest and XGBoost, for failure time prediction. Key challenges influencing prediction accuracy were identified, including complex failure mechanisms, feature engineering constraints, and issues with historical data. This study provides a foundational framework for sewer failure time prediction and highlights key methodological challenges requiring resolution to improve prediction accuracy.