Optimizing the Maintenance Planning of Infrastructure Assets
Advisor: Professor Khaled El-Rayes
Abstract
Large organizations often have diverse facility portfolios. This aspect is notably true for organizations like the Department of Defense, which manages hundreds of thousands of facilities encompassing far-reaching applications, from business administration to weapons testing and evaluation. Ensuring these facilities meet operational objectives is no easy task, especially as infrastructure budgets continually face downward pressure to reduce costs. Considering such constraints, large organizations must find better ways to manage facility operations. This process includes crucial aspects like modeling facility degradation and decision support systems for improving maintenance activities.
A primary objective of this research was to build on the U.S. Army Corps of Engineers degradation model as applied to constituent components within facility infrastructure. The extended Weibull model used by the Corps of Engineers has a pedigree spanning decades and is applied across various assets, including pavements, dams, buildings, and other facility types. Inspection information gleaned from the preceding decades provides a wealth of information that can be leveraged to improve this approach. This research developed a method based on a novel likelihood function capable of selecting model parameters to improve the condition forecasts. A parametric failure likelihood function was created that incorporates all the parameters of the extended Weibull model. A distribution of empirical failure times is collected and the likelihood function is used to select the parameter values that best match the failure time distribution. This enables the definition of model parameters from empirical data. When data is sparse, machine-learning methods were developed to advance the applicability of this method by estimating failure ages from non-failed equipment inspection observations.
The research further broadened the use of the new likelihood function by leveraging it to translate the previously deterministic extended Weibull model to a probabilistic model. A Markov chain Monte Carlo method is used, where the extended Weibull model can produce a range of plausible condition forecasts, rather than a single prediction. This range of forecasts is more amenable to risk-based decisions as it helps capture the uncertainty in degradation. A case study demonstrates how this approach can be used for improved, risk-based cost estimation of component repair.
With improved condition forecasts and repair estimates, the research continued with the development of new decision optimization frameworks. A mixed-integer linear program is created to generate a repair plan to meet a user-defined facility health goal at minimal cost. An alternate linear programming framework seeks to maximize facility health while constrained by a user-defined budget. This latter approach is also extended to multi-year repair planning and applied to component- and facility-level decisions. The focus on linear programming affords the
developed models more scalability, which is particularly important when developing optimal maintenance plans for large-scale portfolios of real property assets
