Fusion of heterogeneous data for robust degradation prognostics

• This paper provides a full methodology for fusion of heterogeneous data for a computationally intensive degradation simulation model. • The proposed modular approach performs offline data assimilation in two steps. Firstly, it employs a Bayesian model updating (BMU) step using kernel sensitivity analysis techniques to rank and evaluate the time-varying importance of input variables and then provides a specific Bayesian technique to sample from the data-informed posterior distributions using Monte Carlo Markov chain (MCMC) techniques. Secondly it makes use of Ensemble Kalman smoothing methods for full state updating and subsequent uncertainty reduction. • The proposed BMU method overcomes the curse of dimensionality for high-dimensional MCMC posterior estimation by iteratively updating the marginal distributions of individual influent input variables assuming independent marginals and by measuring the data-informed posterior compared to the prior by computing the Kullback-Leibler divergence. • The proposed method allows to make robust probabilistic predictions of the remaining useful life of an asset by propagating data-informed posteriors of the input variables in the computer simulation model and comparing it to non-informed prior distribution. • The proposed method is well suited for expensive-to-evaluate simulation models by integrating a surrogate modeling step and proposes a way to integrate the induced metamodeling bias with a Monte Carlo aggregation approach. • Benefits of our method are showcased first on a controlled application to Paris-Erdogan’s law for crack growth propagation with a fictitious material and then on a challenging prognostics application for steam generators clogging in nuclear power plants. Assessing the degradation state of an industrial asset first requires evaluating its current condition and then projecting the forecast model trajectory to a predefined prognostic threshold, thereby estimating its remaining useful life (RUL). Depending on the available information, two primary categories of forecasting models may be used: model-based simulation codes and data-driven (machine learning) approaches. Combining both modelling approaches may enhance prediction robustness, especially with respect to their individual uncertainties. This paper introduces a methodology for fusion of heterogeneous data in degradation prognostics. The proposed modular approach acts iteratively on a computer model’s uncertain input variables by combining kernel-based sensitivity analysis for variable ranking with a Bayesian framework to inform the priors with the heterogeneous data - and adds a Kalman based smoothing step for reducing uncertainties on the prognostics horizon. Additionally, we propose an integration of an aggregate surrogate modeling strategy for computationally expensive degradation simulation codes. The methodology updates the knowledge of the computer code input probabilistic model and reduces the output uncertainty. As an application, we illustrate this methodology on a toy model from crack propagation based on Paris law as well as a complex industrial clogging simulation model for nuclear power plant steam generators, where data is intermittently available over time.