Study of ensemble learning-based fusion prognostics

In this paper we explore the effectiveness of ensemble learning in the failure prognosis field by MLP neural network. An effective ensemble should consist of a set of learners that are both accurate and diverse. In the training stage, we use the Adaboost.R2 technique to train several weak learners (multi layer perceptron network-MLP) to increase the diversity of the individual models. In the prediction stage, we focus on the design of the fusion of the weak learner ensemble. In contrast with the traditional static weight allocation method based on the overall data set, we propose a dynamic weight allocation method, based on the performance of an individual weak learner on the subset. The idea behind this method was to use a test sample's neighbors to estimate the accuracy and bias of the individual weak learner when the learner is used to make the prediction. An improved hyperrectangle neighborhood defining method is proposed in this paper. Some experiments using MLP neural network as a weak learner on a NASA turbofan engine degradation simulation dataset were carried out. The preliminary empirical comparisons showed higher performance of the novel ensemble learning methodology for the RUL estimation of engineering systems.