Authors: Sina Masoumzadeh Sayyar, Matti Kummu, Ilkka Mellin, Franz Tscheikner-Gratl & Tuija Laakso
Publication type: A1
Abstract
The focus on proactive wastewater asset management emphasises the importance of asset lifespan prediction. Although machine learning models show promise in predicting the current condition of the pipes, they fall short in forecasting their lifespan due to the time of failure being unknown (censored); either because the event has not yet occurred or because it is only observed after it has already taken place. This research presents an approach for sewer pipes’ lifespan prediction that focuses on preprocessing steps to train various machine learning algorithms on censored pipe inspection datasets without modifying the algorithms. By segmenting the prediction time horizon into shorter periods and using inverse probability of censoring to weight the training data, we achieved high accuracy. The best model achieved an F1-score of 0.83 for failure occurrence and an RMSE of 0.97 years for failure timing, which is an acceptable timing error for sewer asset management decisions. Analysis showed that event definition affects model performance and interpretation of predictions, and that it is easier to predict the utility’s decision-making than the pipes’ deterioration. This approach performs well with limited data, creating the opportunity for practical use of machine learning even for small utilities.