Improving Computational Efficiency of Prediction in Model-based Prognostics Using the Unscented Transform

Model-based prognostics captures system knowledge in the form of physics-based models of components, and how they fail, in order to obtain accurate predictions of end of life (EOL). EOL is predicted based on the esti- mated current state distribution of a component and ex- pected profiles of future usage. In general, this requires simulations of the component using the underlying mod- els. In this paper, we develop a simulation-based pre- diction methodology that achieves computational effi- ciency by performing only the minimal number of sim- ulations needed in order to accurately approximate the mean and variance of the complete EOL distribution. This is performed through the use of the unscented trans- form, which predicts the means and covariances of a distribution passed through a nonlinear transformation. In this case, the EOL simulation acts as that nonlinear transformation. In this paper, we review the unscented transform, and describe how this concept is applied to efficient EOL prediction. As a case study, we develop a physics-based model of a solenoid valve, and perform simulation experiments to demonstrate improved com- putational efficiency without sacrificing prediction accu- racy.