Predicting time-to-failure of water pipelines

Abstract

Water distribution networks (WDNs) are designed to consistently provide users with access to clean water. However, pipe failure affects the water quality and service provision of water mains. To keep the level of service, the pipe network undergoes frequent repair and rehabilitation, especially in dense urban areas. Researchers extend detailed studies to understand pipe failure phenomena in depth. However, developing a failure prediction model based on past failure data has been challenging. Generally, the current deterioration models primarily focus on the pipe condition assessment and pipe failure rate. Computation of time-to-failure for water pipelines is limited. Time-to-failure can be used as the prime metric for pipe rehabilitation and repair decision-making. Various static and dynamic parameters impact the accurate time-to-failure (TTF) computation for water distribution pipelines. This renders the precise evaluation of failure time for pipes rather tricky. Therefore, the current study aims to develop an efficient TTF prediction modeling regime for water pipes considering three objectives: 1) Study the factors affecting the TTF of water pipelines, 2) Develop and validate models for predicting failure timing in water pipelines and 3) Automate the validated models. For the first objective, the time-to-failure influencing factors were identified and prioritized from past models using systematic review and fuzzy-AHP ranking. Primary findings reveal that only some models exist focusing on failure age or time-to-failure metrics for water pipeline failure prediction. Study of factors revealed several findings: i) data access and quality issues dictate use of pipe length and pipe diameter as the predictors, ii) the effect of weather features on failure time of water pipelines is unexplored, and iii) relevant effect of failure features is unquantified and spatially variant. For the second objective, clustering and topic modeling revealed the current state-of-the-art and intricate gaps in developed models for efficient time-to-failure prediction. The review found that the best practice for failure time prediction using past databases needs to be more specific and developed. Thus, several statistical and machine learning-based models were developed, tested, and optimized for accurate time-to-failure prediction. Survival hazard models predicted a probabilistic time window for failure occurrence with a cross-validation concordance index above 0.8 for both generic and failure-specific models after optimization. Year of installation, rainfall, and annual average daily traffic (AADT) were observed to be the most influencing features in all the models, apart from pipe length and pipe diameter. For predictive maintenance, however, machine learning-based deterministic models were developed. For 1st failure prediction, the Random Forest and CatBoost was observed most accurately, and for 2nd failure prediction, CatBoost and Support Vector Machines-based models were observed to have the highest accuracy initially. Further, feature selection using a random forest-integrated genetic algorithm was employed to remove redundant and unnecessary features, improving model performance. Secondly, feature engineering was used to develop a weather index for failures from 2010 to 2021, which further improved the modeling performance. For 1st failure prediction, a testing accuracy of 1<MAE<1.5 and R2 > 0.9 for both generic and material-specific models was achieved. For 2nd failure prediction a testing accuracy of 0.5<MAE<1.5 and R2 > 0.8 for both generic and material-specific models. For water pipelines with sequential failures, it was considered beneficial to use a spatially informed time-to-next failure prediction using deep learning-based frameworks. Both adapted tabular attentive deep learning model (tab-net) and the spatiotemporal LSTM model have testing MAE<0.9 and R2>0.8. The results assert the use of feature engineering and spatially informed deep learning models for high-accuracy time-to-failure predictions. The models are integrated with a Python-programmed web-based GUI for automated usage and application in the water industry. The outputs of the study make an original contribution to the knowledge and practice of water pipeline failure prediction, management of urban water networks, and water pipeline rehabilitation planning. The study has practical applications: (1) The developed models and shortlisted factors can be readily adapted to other water networks, (2) The work allows for both prediction of consecutive failure times (3) Enabling automated prediction of failure time using the developed web application. Using predictive knowledge of failure timing, water practitioners can overcome data limitations and systematic uncertainty in pipeline maintenance and rehabilitation decision-making. Using the modeling capabilities, urban water management decisions can have a strong technical basis and confidence, resulting in a more robust, adaptable, and sustainable system.