Using a Negative Binomial Regression Model with a Bayesian Tuner to Estimate Failure Probability for Sewerage Infrastructure

Abstract

The replacement and maintenance of subsurface assets, such as water and wastewater pipes, is of great interest to water utilities because these infrastructure networks require large amounts of investment over time. Each asset requires investment relatively rarely, but the number of assets is so enormous that the flow of money is large. Therefore the accurate estimation of the deterioration and aging process of these assets is critical to the efficient and sustainable allocation of investment spend. The development of failure models is difficult for various reasons: short spans of data (very little longitudinal data), very sparse failure rates, inaccuracy of observational data, and accuracy and availability of potential predictor data. Technical difficulties also arise such as variability and noise, censoring effects, overdispersion, and, throughout the exercise, the large volume of data usually involved. In this paper, a new regression approach is formulated that maintains a rigorous statistical approach while still being practical and easy to apply. In addition, the formulation involved permits the individual pipe history to be used in an elegantly simple Bayesian update. The example provided refers to work done for Yorkshire Water Services in estimating probability of blockage failure for all their sewerage assets.