Modeling Pipe Break Data Using Survival Analysis with Machine Learning Imputation Methods

Abstract

The development of asset life estimation tools based on historical data is essential to the effective management of pipeline assets. One tool that may assist with asset management is survival analysis. However, left-truncated break records pose a challenge in the practice of survival analysis to obtain sound inferences and predictions. In this study, we propose a data-driven approach that integrates machine learning imputation methods with survival analysis. To demonstrate the proposed methodology, we perform a case study using ductile iron (DI) water distribution pipes from an anonymized utility in the midwestern United States. Two artificial neural network (ANN) models are developed as imputation methods to calibrate the survival curves and mean time to first failure (MTTF) estimates from the Weibull proportional hazards model (WPHM). Results show that the MTTF estimation bias is reduced from 14.3% to 2.1% by using imputation as a preceding procedure. Empirical findings show that despite the limited accuracy of imputation models, the use of imputation methods can still improve the survival analysis results and mitigate the impact of left-truncated break records.