On the use of the unscented transform for failure prognostics

Dealing with uncertainty is of uttermost importance in the failure prognostics task. The difficulties in estimating current health state and the unknown future operational conditions are among the reasons that make this consideration so important. Random variables (RV) are the most usual way to represent such uncertainty. In the case of health state of a monitored system, or its trend, one usual way of propagating the uncertainty into the estimate of the remaining useful life (RUL) is based on Monte Carlo (MC) simulations. Despite being adequate for this task, MC simulations present some drawbacks: the calculations are computationally intensive and there is no consistent way of defining an adequate number of samples to simulate. This paper presents and evaluates alternative means to perform this task using the unscented transform (UT). The UT is a method that yields the probability distribution of a RV which results from a nonlinear transformation of another RV. If the latter is associated with the current health state of a monitored system, or its trend, and the former represents the RUL, the UT can be applied for failure prognostics. The advantages of using such approach are the necessity of a much smaller number of samples compared to MC and the straightforward definition of this number. This work presents novel applications of the UT for prognostics and performance evaluations that indicate the adequacy of this approach.