Knowledge Elicitation for Predictive Maintenance Modelling with Bayesian Networks

Predictive maintenance is closely connected to condition monitoring, health and usage monitoring systems (HUMS), prognostic modeling and condition-based maintenance (CBM). It is concerned with identifying what maintenance actions should be chosen and when, based on the predicted state of the item of interest over time. It can be viewed as a logical development of CBM. The authors of this paper believe that dynamic Bayesian networks (DBNs) provide a useful computational approach to predictive maintenance modeling. However, DBNs require a large number of probabilities to populate a model. While these numbers can come from collected data or domain experts, one of the key challenges in creating prognostic models is dealing with a lack of the data required, particularly for new systems. Therefore, human expertise can serve as an initial source of knowledge and can come from designers, maintainers and users of older systems. In this paper, we discuss the various types of knowledge required to build such models, how it might be elicited and the practical considerations involved. .