Improving data-driven prognostics by assessing predictability of features

Within condition based maintenance (CBM), the whole aspect of prognostics is composed of various tasks from multidimensional data to remaining useful life (RUL) of the equipment. Apart from data acquisition phase, data-driven prognostics is achieved in three main steps: features extraction and selection, features prediction, and health-state classification. The main aim of this paper is to propose a way of improving existing data-driven procedure by assessing the predictability of features when selecting them. The underlying idea is that prognostics should take into account the ability of a practitioner (or its models) to perform long term predictions. A predictability measure is thereby defined and applied to temporal predictions during the learning phase, in order to reduce the set of selected features. The proposed methodology is tested on a real data set of bearings to analyze the effectiveness of the scheme. For illustration purpose, an adaptive neuro-fuzzy inference system is used as a prediction model, and classification aspect is met by the well known Fuzzy Cmeans algorithm. Both enable to perform RUL estimation and results appear to be improved by applying the proposed strategy.