Value-Based Bayesian Optimization of Preventive Maintenance Programs

The EPRI Preventive Maintenance Basis Database (PMBD) has become a standard in the industry to develop, validate, or examine the impact of custom changes to maintenance strategies for common power plant equipment. The PMBD provides failure modes and an indication of frequency of occurrence. Recent feedback from PMBD users has made it clear that including a “Cost Module” to work with PMBD data would be a useful addition to the PMBD program and allow users to view the cost impacts associated with alternate custom maintenance strategies. This paper presents a methodology for the merging of maintenance information extracted from PMBD with cost estimates and additional expert-provided reliability data to estimate a maintenance cost distribution. Additional expert information includes missing data and PM type: monitoring, wear-rate reducing (e.g. oil change), or life-restoring (e.g. refurbishment). The cost distribution is calculated via Monte Carlo simulation and is dependent on the PM plan currently considered. Value-based optimization of the PM plan is performed through Bayesian optimization of the mean PM cost by varying the various PM frequencies. Bayesian optimization iteratively uses Gaussian Process Regression (GPR) to fit a non-parametric meta-model to a noisy objective function. As a part of GPR it is necessary to fit a covariance function that describes the spatial correlation or smoothness of the objective cost function. The meta-model with the covariance function effectively produces a built-in sensitivity analysis for the optimization as well.