Predictive and Spatial Analytics for Planning Inspections of Sewer Infrastructure

Edward McBean; Richard Harvey
Aging sewer pipes inevitably deteriorate to a point where raw, untreated wastewater leaks out of the pipe and into the surrounding soil and nearby sources of groundwater. Municipalities can utilize closed-circuit television (CCTV) inspection technology to identify individual pipes within a sewer network in bad structural condition. Although CCTV inspections provide essential information on pipe condition, they are expensive and often limited to small portions of an entire sewer network. Consequently, any threat to the environment posed by uninspected pipes remains unknown. Predictive analytics can leverage existing inspection datasets so that reliable predictions of condition are available for individual pipes not yet inspected. The predictive capability of the “random forests” data mining algorithm is demonstrated using a case study of sanitary sewer pipe condition data collected by a municipality in Ontario, Canada. A comparison of class-imbalance learning strategies (undersampling and threshold adjustment) is carried out to evaluate the potential to increase predictive accuracy for bad condition pipes representing the minority class of interest. Threshold adjustment is found to provide an optimal level of performance for the classification problem – with the trained predictive model achieving a false negative rate of 18%, false positive rate of 29% and an excellent area under the receiver operating characteristic (ROC) curve of 0.81 (considered to be excellent given the nature of the inspection dataset). An analysis of the predictive capabilities of the random forests algorithm trained using a dataset one-third the size of the one originally available to the case study municipality indicates the algorithm has utility for municipalities outside of the case study area. Network spatial analytics are implemented to visualize clusters of bad condition pipes in the sewer system. Visualization of pipe condition in this manner is found to be an effective tool for screening candidate pipes for inspection and for conveying the necessity of inspection to members of a municipality responsible for approving sewer inspection-related budgets.
Exfiltration; Inspection; Predictive Analytics; Random Forests; Wastewater
Download | Back to Issue| Archive