Long term prediction approaches based on connexionist systems - A study for prognostics application

Data-driven approaches are increasingly applied to machine prognostics. More precisely, connexionist systems like neural networks and neuro-fuzzy systems benefit from a growing interest. Indeed, their approximation capability makes them as powerful candidates to achieve the prediction step of prognostics. Nevertheless, prognostic implies to be able to perform multi step ahead predictions whereas many works focus on short term predictions. Following that, the aim of this paper is to review and discuss the connexionist-systems-based approaches to ensure long term predictions for prognostics. The paper emphasizes on univariate time series forecasting. Five connexionist-systems based approaches are pointed and formalized, namely: the iterative, direct, DirRec, parallel and MISMO approaches. Their performances are analyzed according to three types of criteria: those one of prediction accuracy, of complexity (computational time) and of implementation requirements. In addition, simulations are made among 111 times series prediction problems in order to reinforce the discussion. These experiments are performed by using the exTS (evolving extended Takagi-Sugeno system). Finally developments are applied on a real engine fault prognostics problem in order to validate conclusions on a real world case and to point out some best practices for prognostics applications.