From real data to Remaining Useful Life estimation: an approach combining neuro-fuzzy predictions and evidential Markovian classifications

This paper deals with the proposition of a prognostic approach that enables to face up to the problem of lack of information and missing prior knowledge. Developments rely on the assumption that real data can be gathered from the system (online). The approach consists in three phases. An information theory-based criterion is first used to isolate the most useful observations with regards to the functioning modes of the system (feature selection step). An evolving neuro-fuzzy system is then used for online prediction of observations at any horizons (prediction step). The predicted observations are classified into the possible functioning modes using an evidential Markovian classifier based on Dempster-Shafer theory (classification step). The whole is illustrated on a problem concerning the prediction of an engine health. The approach appears to be very efficient since it enables to early but accurately estimate the failure instant, even with few learning data.