Bayesian Networks for Remaining Useful Life Prediction

Abstract

Remaining useful life (RUL) prediction is a critical task in the field of condition-based maintenance. It is important to perform RUL prediction in a statistical sound way. However, it is not straightforward to properly combine multiple information sources about an asset, such as available statistics, measurements, derived features, and prior knowledge in the form of mathematical models and relations, including their uncertainties. Bayesian networks (BNs) are a means of graphically representing all statistical information in a comprehensible way and allow for correctly combining all information. BNs allow for inference in all directions, thereby not merely providing a RUL prediction with explicit uncertainty, but select the most informative features, diagnose which degradation mechanism is manifest if multiple mechanisms exist, provide decision support in the form of optimal condition-based maintenance points when combined with a cost model. BNs also explicitly quantify the model uncertainty arising from the scarcity of the training data. We illustrate these benefits on two realworld industrial examples: solenoids and bearings. We also provide a method to correctly include the effect of changing operating conditions.