Bayesian network design for fault diagnostics of railway switches

Abstract

Besides detecting failures and predicting future health conditions of technical systems, fault diagnosis (i.e., fault identification) is a key challenge in the analytic part of prognostics and health management (PHM). In this context, Bayesian networks (BN) has proven to be an effective tool for diagnostic reasoning about faults and effects. Since it is possible to generate such models not only from data but also from expert knowledge or a combination of both (hybrid approach), Bayesian networks are well-suited for many applications and (technical) disciplines. This, in particular, holds for situations where common data-driven approaches (e.g., neural networks, deep learning) suffer from a lack of a reasonable amount of adequate training data. This contribution discusses the detailed design of a comprehensive Bayesian network for railway switches as to be used for fault diagnosis in context of corrective and/or predictive maintenance, for instance. The new model explicitly pursues the modular paradigm of object-oriented Bayesian networks (OOBN), and thus provides a maximum degree of flexibility when adapting it to different types of railway switches. Moreover, it contains Bayesian nodes that act as a kind of "ON/OFF switches" and allow to (de-)activate specific parts of the model without affecting its overall structure. This, in particular, is useful whenever the general Bayesian network comprises modules (e.g., point heater or back drive) that are not available to all switches in the field. Finally, the model benefits from a newly developed, innovative design principle for Bayesian networks which, based on a generalization of the idea of Boolean clusters, reduces (or potentially even completely avoids) the problematic effect of overconfidence in diagnostic reasoning.