Probabilistic model selection for prognostics of thermal Creep in advanced reactors

A critical challenge with prognostics of passive component degradation in advanced reactors is the sensitivity of materials degradation to operating temperatures, load conditions, and microstructure of the material. The dependence of degradation accumulation on these (and other) environmental factors can also change over the lifecycle of the degradation mechanism. These types of nonlinearities also make it difficult to capture the growth phenomena using a single empirical model, and challenge the ability to predict failure progression, especially closer to endof- life (where abrupt material failure can occur). These challenges indicate the need for tools and techniques for prognostic model selection. This paper presents a Bayesian model selection approach to select the appropriate creep degradation model at any given time, using relevant sensor measurements reflecting the material degradation state. The model selection approach, based on reversible jump Markov chain Monte Carlo (MCMC) methods, is integrated with Bayesian particle filter-based prognostic framework. The proposed approach is evaluated using synthetic data, representing measurements from materials subjected to high-temperature creep degradation. Creep phenomena occurring under extreme temperatures and loading environments is expected to be one of the common causes of material degradation in advanced reactor passive components. We begin with a set of models that capture the three distinctive stages of creep growth pattern; namely, primary stage, secondary stage, and tertiary stage. The proposed approach is then used to infer the present state of material degradation from the measured sensor response, select an appropriate damage growth model, and predict the remaining useful life (RUL) of the component. Results to date indicate the feasibility of the approach and show that the use of model selection can improve the accuracy of the prognostic result for advanced reactor passive components. The results also indicate that within the Bayesian prognostic framework, the accuracy of the model selection may be impacted by the noise levels assumed in the process and measurement models.