Bond graphs for quantitative fault diagnostics to support Bayesian filtering-based failure prognostics

Predictive maintenance has a lot of potential in preventing catastrophic and expensive failures and in optimizing maintenance logistics. Bayesian filters can be used for prognostics as they can update physics-of-failure models, establishing a quantitative relation between asset usage and the remaining useful life. However, such filters require real-time direct condition measurements (e.g. crack length), while in many practical applications, only indirect condition measurements (e.g. vibrations), are available. A quantitative diagnostic algorithm that converts indirect into direct condition measurements is needed to create the required input for Bayesian filters. Such a diagnostic algorithm requires (labeled) degradation data, which are often unavailable for real-world applications. Therefore, simulation models can be helpful to generate synthetic degradation data. This paper proposes to use bond graphs for the generation of synthetic degradation data. Bond graphs are especially helpful due to their modularity and reusability. The creation of synthetic degradation data is demonstrated on a case study concerning outer race damage of a rolling-element bearing. An existing algorithm for fault size estimation is applied to the synthetic data generated by bond graphs. The fault size, derived by the quantitative diagnostics algorithm, is implemented in a prognostic framework based on a Bayesian filter. It is shown that the bond graph accurately creates synthetic failure data. Furthermore, it is demonstrated that existing diagnostic methods are insufficiently accurate to represent all damage shapes in bearings while bond graphs create vibration data for specific fault shapes. Therefore, bond graphs can improve quantitative diagnostics.