Model-based and data-driven prognostics under different available information

In practical industrial applications, different prognostic approaches can be used depending on the information available for the model development. In this paper, we consider three different cases: (1) a physics-based model of the degradation process is available; (2) a set of degradation observations measured on components similar to the one of interest is available; (3) degradation observations are available only for the component of interest.The objective of the present work is to develop prognostic approaches properly tailored for these three cases and to evaluate them in terms of the assumptions they require, the accuracy of the Remaining Useful Life (RUL) predictions they provide and their ability of providing measures of confidence in the predictions. The first case is effectively handled within a particle filtering (PF) scheme, whereas the second and third cases are addressed by bootstrapped ensembles of empirical models.The main methodological contributions of this work are (i) the proposal of a strategy for selecting the prognostic approach which best suits the information setting, even in presence of mixed information sources; (ii) the development of a bootstrap method able to assess the confidence in the RUL prediction in the third case characterized by the unavailability of any degradation observations until failure.A case study is analyzed, concerning the prediction of the RUL of turbine blades affected by a developing creep.

► We developed particle filtering and bootstrap ensemble-based prognostic approaches. ► We applied the methods to the RUL prediction of creeping turbine blades. ► We analyzed the methods sensibility to the quality of the information available. ► Particle filtering is more information demanding than data-driven approaches. ► In general, the particle filtering provides slightly better prognostic results.