Degradation prognostics of industrial assets involves estimating their remaining useful life (RUL) by projecting their current health and operating conditions and assessing the associated uncertainties. This is usually done using physics-based simulations or data-driven models. While each approach has strengths, they can fall short especially when the numerical code is time-consuming and when the available degradation data is sparse. To address these issues, we developed an offline data fusion method that combines kernel-based sensitivity analysis with an iterative Bayesian update of influential computer model inputs. To reduce the computational cost, we use a field surrogate modeling strategy and an aggregation method to reduce the metamodeling bias in the posterior distributions. After presenting the method, I will show how it reduces the RUL prediction uncertainties on a clogging prognostics problem for steam generators in nuclear power plants.
Joint work with: Vincent Chabridon (EDF R&D), Emmanuel Remy (EDF R&D), Mathilde Mougeot (ENS Paris-Saclay), Didier Lucor (LISN), Merlin Keller (EDF R&D), Julien Pelamatti (EDF R&D), Michaël Baudin (EDF R&D).
