467 - A Novel Sewer Condition Assessment Approach that Leverages Advanced Machine Learning (ML) Models

A novel approach to sewer network condition assessment that uses artificial intelligence (AI)/machine learning (ML) is presented in this paper. Defect locations in sewer networks were predicted using multiple ML models that were developed using data from CCTV inspections coupled with additional pipe information including pipe surroundings. Most of the existing condition assessment methods include a factor derived from the CCTV operator assigned Pipeline Assessment and Certification Program (PACP®) codes in the inspection reports to prioritize maintenance and may include additional factors in a linear weighted function. The ML approach used the CCTV videos and reports from two utilities and analyzed them using computer vision (CV) to create the input datasets. Those datasets were used in multiple decision-tree ML algorithms that focused on predicting infiltration and inflow (I&I) and structural defects. The class imbalance in the data was addressed with three techniques and the hyperparameters were optimized via a parallel cross-validation grid search. The performance of the models was assessed using the Area Under the Curve (AUC) of Receiver Operating Characteristic (ROC) curve. The LightGBM algorithm for I&I prediction with SMOTE to address class imbalance achieved AUCs of 0.80 for utility A and 0.83 for utility B. For structural defect prediction, LightGBM with class weights to address imbalance performed best, achieving AUCs of 0.73 for utility A and 0.76 for utility B. The SHAP (SHapley Additive exPlanation) tool showed that most important feature was “pipe location” for utility A and “pipe age” for utility B.