A fusion prognostic approach based on multi-kernel relevance vector machine and Bayesian model averaging

Abstract

The fusion approaches with multi-model ensemble can present a better performance than the simple approaches with single model in Prognostics and Health Management (PHM). Bayesian Model Averaging (BMA) is a very useful ensemble method in these fusion approaches because of its ability of uncertainty quantification. A fusion model based on BMA and relevance vector machine (RVM) is presented in this paper. Multi RVM models with eight different kernel functions are constructed for the degradation of training data. Then, BMA is used to integrate these RVM sub-models into one framework for the reliability of prognostic. The main advantages of this method are: 1) to solve the lack of uncertainty management of existing fusion remaining useful life (RUL) estimation methods and 2) to improve the prediction performance by ensemble multi complex nonlinear sub-models based on BMA, while most of BMA applications are to integrate some simple linear sub-models result in low prediction precision. Finally, for the low multistep-ahead off-line model prediction precision, an online iterated training strategy is introduced for BMA algorithm to realize high prediction performance. The experiments and the results with the battery data sets from National Aeronautics and Space Administration (NASA) demonstrate the effectiveness and the reliability of our proposed ensemble prognostic approach.