Inspection schedule for prognostics with uncertainty management

Abstract

Condition monitoring data is an essential ingredient for prognostics and health management. To minimize unnecessary inspections or measurements, it is crucial to evaluate the value of data to be measured in advance and determine the inspection or data measurement schedule. For this purpose, it is important to predict how much prognostics performance will be improved by adding additional data. Motivated by this objective, this paper proposes a new method that determines the future data measurement schedule which can reduce the uncertainty in prediction to the desired level. The proposed method decomposes the prediction uncertainty into epistemic and aleatory uncertainty, which are caused by the uncertainty of model parameters and the noise in the data, respectively. Then, contributions of these uncertainties to the overall prediction uncertainty in the future are analyzed. The next measurement schedule is determined such that the level of reducible epistemic uncertainty in the prediction is the same as that of aleatory uncertainty. The proposed method is applied to two different prognostics approaches: the model-based and data-driven methods. Two examples showed that the total number of inspections is reduced by about 85% while keeping the same level of prediction uncertainty.