Evaluation of Artificial Intelligence Tool Performance for Predicting Water Pipe Failures

Over the past years, there has been a sustained interest in developing machine learning (ML) models that are sophisticated enough to capture the failure trends of water distribution systems and that are able to predict future breaks of the pipeline system. Given the limited budgetary resources of water pipeline owners, coupled with the deteriorated state of water networks, there is a vital need to deploy such tools to prioritize inspection and replacement of vulnerable regions, as well as to mitigate the chance of having catastrophic failures within the system. This study extends several ML algorithms that analyze the historical failures of water pipelines, with the goal to predict future breaks. The performance of each algorithm has been examined using various water networks as different case studies with varying network size and configurations. The developed models are all aimed to estimate the future likelihood of pipe failure by exploring historical failure patterns, surrounding attributes (e.g., environmental and demographic), as well as pipe characteristics. To improve the predictive power of the learning algorithms, several engineered features have been also created from raw data and tested to facilitate the learning process of each algorithm. While developing such models is by no means an insignificant task, an equally, if not more important emphasis should be put on how precisely these models are predicting actual failures. Additionally, the model variables should be defined wisely enough to ensure the uniqueness of each network has been captured and incorporated into the analysis. Lastly, it is crucial to evaluate the precision of the developed predictive models to evaluate the level of reliability a utility can expect by deploying it, as well as the further improvement needs of the predictive algorithm itself. Accordingly, this paper will review the analyses performed, the outcomes of this study, and discuss plans to improve upon the analyses to ensure that maximum usefulness of the model can be achieved.