A neuro-fuzzy self built system for prognostics: a way to ensure good prediction accuracy by balancing complexity and generalization

In maintenance field, prognostics is recognized as a key feature as the prediction of the remaining useful life of a system allows avoiding inopportune maintenance spending. However, it can be a non trivial task to develop and implement effective prognostics models including the inherent uncertainty of prognostics. Moreover, there is no systematic way to construct a prognostics tool since the user can make some assumptions: choice of a structure, initialization of parameters... This last problem is addressed in the paper: how to build a prognostics system with no human intervention, neither a priori knowledge? The proposition is based on the use of a neuro-fuzzy predictor whose architecture is partially determined thanks to a statistical approach based on the Akaike information criterion. It consists in using a cost function in the learning phase in order to automatically generate an accurate prediction system that reaches a compromise between complexity and generalization capability. The proposition is illustrated and discussed.