Bayesian Experiment Design for the Development of an Epoxy Resin Degradation Model

This paper presents a novel approach to design of highly time-consuming aging experiments of electrical insulation system (EIS) materials, specifically epoxy resin, through Bayesian methods. The aim is to find a suitable model with as few experiments as possible. This is achieved by evaluation the information gain for each potential new experimental condition, and select the most informative conditions for measurement. If a model structure is not entirely known, multiple candidate models can be designed and the experiment design procedure quickly discriminates between them. The method is compared with the conventional approach of factorial experiment design and Response Surface Methodology (RMS). The aging of the epoxy samples is then hygrothermal, using temperature and humidity as two factors. This aging setup is not well studied, as the experiments are very demanding, but of high importance for the diagnostics of electrical machines. The results then show that the Bayesian approach is able to find models whose predictions match the validation measurement better than conventional methods.