Comparative Analysis of Machine Learning and Survival Analysis Combinations for Water Main Failure Prediction

Abstract

Water main failure prediction models are a common tool in support of the asset management plans of utilities. This paper focuses on the development and evaluation of models that incorporate survival analysis in combination with machine learning algorithms. To understand the potential advantages of these models, five representative algorithms are applied to data from four different utilities. The Cox proportional hazards (Cox-PH) algorithm, a common survival analysis technique, is considered to provide a baseline measure of model performance. Other algorithms investigated include: extreme gradient boosting (XGBoost), random survival forest (RSF), neural multitask logistic regression (NMTLR), and XGBoost survival embeddings (XGBSE). XGBoost is a pure machine learning algorithm whereas the other algorithms embed survival analysis within a machine learning framework. Our findings suggest that the models that embed survival analysis within machine learning may predict water main failures more accurately than traditional survival analysis or machine learning models. All models are developed using data sets with high degrees of censorship and varying characteristics; thus, these approaches may be beneficial for a wide range of utilities, including those with limited break data. In particular, models based on the RSF algorithm are found to perform consistently well for the data sets investigated herein. To support decisions related to the repair or replacement of water mains, many utilities develop models that predict which water mains are most likely to break. Survival analysis (a branch of statistics) and machine learning models have been frequently applied in previous studies on water main failure prediction. However, when applied on their own, there are limitations to both of these types of approaches. More recently, several studies have focused on the potential for combinations of survival analysis and machine learning to overcome these limitations. In this study, three algorithms that combine machine learning and survival analysis are compared with a common survival analysis method and a machine learning algorithm. Two of the selected algorithms have not previously been applied in water main failure prediction, and this study exhibits their capabilities in this field. The algorithms are applied to data sets from four different utilities. Results from the analysis indicate that combinations of survival analysis and machine learning outperform a traditional survival analysis algorithm and a machine learning algorithm for the data sets included in this study.