Adapting Oversampling and Hyperparameters Tuning to Improve the Performance of Prediction Model for Sewers

Regular inspections are crucial for ensuring effective system operation and for planning needed maintenance. Closed-circuit television (CCTV) is widely employed in North America to examine the internal conditions of sewage pipes. Due to the extensive inventory of pipes and associated costs, it is often not practical for municipalities to conduct inspections on each sanitary sewage pipe section. According to the American Society of Civil Engineers (ASCE) Infrastructure Report published in 2021, combined investment needs for water and wastewater systems are estimated to be $150 billion during 2016–2025. Therefore, new solutions are needed to fill the trillion-dollar investment gap to improve the existing water and wastewater infrastructure for the coming years. Machine learning-based prediction model development can be an effective method for predicting the condition of sewer pipes. In this research, sewer pipe inspection data from several municipalities were collected, which included variables, such as pipe material, age, diameter, length, soil type, slope, and Pipeline Assessment and Certification Program (PACP) score. These sewer pipe data exhibit a severe imbalance in pipes’ PACP scores, which is considered the target variable in the development of models. Due to this imbalanced data set, the performance of the sewer prediction model is poor. This paper, therefore, aims to employ oversampling and hyperparameter tuning techniques to treat the imbalanced data and improve the model’s performance. Utility owners and municipal asset managers can utilize the developed models to make more informed decisions on future inspections of sewer pipelines.