Prediction of Sewer Pipelines Using Machine Learning Techniques: A Case Study on the City of Hamilton Sewer Network

Several physical, environmental, and operational factors contribute to the deterioration of wastewater pipelines during their service lives. Failure of this critical infrastructure would have repercussions on the economy, society, and the environment. Assessing sewers is most commonly performed by the use of closed-circuit television (CCTV) that helps in gaining knowledge of the structural and operational state of the pipelines. Identifying pipelines to be inspected annually is considered a main step toward a successful and cost-effective CCTV program. Generally, the initial process follows a desktop risk management approach that combines the condition of sewers and its criticality that cluster pipelines in short-, medium-, and long-term inspection intervals. Predicting sewers' conditions would help select critical sewers that are most likely to fail, where these sewers will be prioritized for CCTV inspections to discern their conditions and plan for required restorations, if needed. Accordingly, the current study employs data mining algorithms, specifically Extreme Gradient Boosting (xgboost) and logistic regression (LR), to predict each pipeline's condition based on the Pipelines Assessment Certification Program (PACP) and Water Research Centre (WRc) ratings. The models are implemented on the city of Hamilton sewer network that consists of combined and separate systems. This study attempted to address the prevalent class imbalance in pipe inspection data sets by employing a diverse range of hyperparameters. Overall, XGBoost produced more satisfying results while still falling short of acceptable average performance. Finally, the results were imported into ArcGIS to better depict the expected sewer condition. Comparing these algorithms shows that these techniques can be further improved by utilizing additional data sets collected from various municipalities across North America.