Predictive Analytics for Planning Inspections of Linear Water and Wastewater Infrastructure




Harvey, R. Richard

Journal Title

Journal ISSN

Volume Title


University of Guelph


Techniques currently available for modeling the deterioration of aging water and wastewater linear infrastructure tend to focus on providing municipalities with generalized estimates of condition at the network-wide level. These models have often been found incapable of reliably predicting the condition of individual pipes within a larger network. The primary goal of this research was to utilize existing data mining tools to make predictions of individual pipe condition that could effectively direct the inspection, maintenance and rehabilitation of critical infrastructure on an asset-by-asset basis. The municipalities of Guelph, Ontario, Canada and Scarborough, Ontario, Canada are provided as case studies. Portions of the sanitary sewer and stormwater networks in Guelph were inspected from 2008 to 2011 using closed circuit television (CCTV) technology. A combined application of predictive and spatial analytics effectively leverages information contained within the existing inspection dataset so that the potential threat posed by uninspected pipes can be suitably assessed. Methods are described for using class-imbalanced inspection datasets to train, tune and test support vector machines, decision tree classifiers and random forests. Decision tree classifiers were found to be a useful first step for extracting information from existing datasets as they illustrate the influence of pipe-specific attributes (e.g. year of installation and length) on structural condition. Support vector machines were outperformed by random forests that achieved excellent levels of predictive accuracy for what is, in reality, a difficult classification task. Ultimately, the proposed modeling methodology has the potential to significantly reduce the time and money spent identifying bad condition, uninspected pipes. An analysis of historical water main failures within Scarborough, Ontario, Canada indicates the majority of failures occur during the very cold winter months. Extensive installation of cement mortar lining and cathodic protection has extended the lifespan of aging water mains. Artificial neural networks are found capable of predicting the time to failure for individual pipes. Simulated failure scenarios indicate a return to high failure rates if cement mortar lining and cathodic protection are not extended to all candidate pipes within the distribution network.



Infrastructure, Water, Sewer, Stormwater, Data mining, Inspection, Management, Random Forests, Decision tree, Artificial Neural Network