^{a}and Peter Zeiler

^{b}

^{a}simon.hagmeyer@hs-esslingen.de

^{b}peter.zeiler@hs-esslingen.de

This paper focuses on data-driven remaining useful life prediction using ensemble methods for prognostics and health management. An important factor for the performance of an ensemble method is the diversity within the ensemble. An effective neural network ensemble method that emphasizes the generation of diversity is negative correlation learning. It is argued that for both diagnosis and prognosis, the consideration of uncertainties has a substantial added benefit over a simple point estimate. For this reason, a prediction interval is derived for the ensemble method negative correlation learning using the delta method. In the delta method, the neural network is treated as a nonlinear regression model, which is approximated by a Taylor series. A look at the derived formula of the prediction interval, emphasizes that negative correlation learning behaves inversely to a regularization. Furthermore, the formula for a diversity parameter of zero is equal to the prediction interval of the regular multilayer perceptron.