Novel Sewer Defect Prediction Leveraging Advanced Machine Learning (ML) Models

A novel approach to sewer network assessment is presented that uses artificial intelligence (AI)/machine learning (ML) to predict infiltration and structural defect occurrences in each pipe instead of estimating the traditional criteria-based overall pipe condition or likelihood of failure. A comparative analysis of four decision tree-based ML models, and their use in predicting the defect locations in sewer networks, is presented. The models are developed using data from closed-circuit television (CCTV) inspections coupled with additional pipe information and inspection reports. The ML approach uses such information from two utilities to create utility-specific defect prediction models. The class imbalance in the data, due to more defects than nondefects, is addressed with three methods, and the hyperparameters, settings that define the model architecture, are optimized via a repeated stratified k-fold cross-validation grid search. The performance of the models is assessed using the area under the receiver operating characteristics (AUC-ROC) and precision recall (AUC-PR) curves. LightGBM-based models, with the cost-sensitive learning method for addressing class imbalance, show the best performance overall when predicting either types of defects for both utilities. The best performing model achieves an AUC-ROC of 0.79 and an AUC-PR of 0.62. For the two utilities investigated, an application of SHapley Additive exPlanations (SHAP) shows that the most important features for indicating both types of defects are "pipe location" and "pipe age."