- Main
Advancing Water Distribution System Resilience: Data-Driven Failure Prediction and Large-Scale Pipeline Mechanics
- Chiu, Shih-Hung
- Advisor(s): Soga, Kenichi
Abstract
Urban water distribution networks are essential to public health, economic activity, and community resilience, yet utilities must simultaneously manage aging buried assets and deploy new pipelines capable of surviving extreme environmental hazards. These challenges arise at different stages of the infrastructure life cycle and therefore require complementary approaches. Existing pipelines generally lack direct condition-monitoring systems and must be assessed using incomplete historical records, whereas newly installed systems can be evaluated through controlled experiments that reveal their structural behavior and failure mechanisms.For existing infrastructure, conventional asset-management approaches commonly rely on pipe age and previous failures, which provide only a partial representation of deterioration. This study combined left-truncated and right-censored survival analysis with ensemble machine-learning models to identify degradation drivers and support near-term rehabilitation prioritization. Previous failure history emerged as the strongest predictor of subsequent failure, while operating pressure, pipe diameter, material, and geotechnical exposure also influenced deterioration. Material-specific analyses further showed that these relationships varied across pipe types. Tree-based machine learning models, evaluated over a defined capital-planning horizon using decision-oriented ranking metrics, outperformed the utility's existing heuristic and identified a subset of first-time failures by incorporating physical, operational, environmental, and spatial information. Spatial evaluation showed that some unpredicted failures occurred near highly ranked pipes, suggesting that pipe-level risk rankings could also support coordinated corridor- or neighborhood-scale rehabilitation planning.For newly installed infrastructure, isolated component tests and full-scale system experiments were used to determine how hazard-resilient segmented pipelines accommodate extreme ground deformation. Tension, compression, four-point bending, and biaxial tests on ductile iron and iPVC joints showed that performance was governed primarily by localized restraint mechanics rather than nominal pipe strength alone. Locking-segment orientation, groove geometry, gasket-tooth configuration, restraint stiffness, and fitting design substantially altered joint stiffness, deformation compatibility, leakage behavior, and failure location. Small-diameter ductile iron specimens generally failed through localized bell or spigot fracture, whereas large-diameter systems were more frequently controlled by displacement or ejection of segmented locking components. The iPVC tests revealed a related trade-off between flexibility and hydraulic resilience: flexible restraints accommodated greater rotation but could permit earlier leakage, while rigid restraints maintained sealing longer but promoted abrupt pipe-wall fracture. Distributed fiber optic sensing resolved localized wedging, asymmetric restraint engagement, cross-sectional distortion, and strain redistribution that were not evident from discrete measurements alone. In the full-scale split-basin experiment, most of the imposed fault displacement was accommodated through joint movement, with deformation progressing through sequential joint activation. The system response was governed by the combined effects of longitudinal friction along the pipe barrels, localized passive resistance at enlarged bells, and restraint engagement after available joint travel was exhausted. Finite-element analyses further showed that reproducing the staged response required explicit representation of friction–displacement coupling, bell anchoring, and evolving joint kinematics.Overall, this work establishes a life-cycle framework for improving the resilience of urban water infrastructure. For aging networks, statistically rigorous and operationally aligned models transform incomplete historical records into actionable rehabilitation priorities. For newly installed systems, continuous strain measurements and full-scale testing reveal the localized mechanics that control structural and hydraulic performance under extreme deformation. Together, these findings support a transition from reactive management based primarily on age and observed failures toward more proactive asset prioritization, joint design, and system-level resilience assessment.