Abstract
Prediction of the remaining useful life (RUL) of critical components is a non-trivial task for industrial applications. RUL can differ for similar components operating under the same conditions. Working with such problem, one needs to contend with many uncertainty sources such as system, model and sensory noise. To do that, proposed models should include such uncertainties and represent the belief about the system’s state in a probabilistic form. In this work, a Bayesian approach is proposed for predicting the RUL of critical components. The approach is divided into two main parts, online and offline. In the offline part, the approach builds k-nearest neighbours classifier (kNN) for different datasets according to their end of life (EOL) values. On the other hand, the online part is similar to the offline apart from the use of Bayesian online state estimator. Bayesian online state estimator is used to represent the uncertainty of the approach about the health status. The approach starts by extracting trends that represent the health evolution of the critical component and uses these trends to build offline models of the critical component. Then, the approach uses these models to predict the RUL from new online data and assigning uncertainty value to it. The approach can be applied to a system with variable operating conditions, however, the prediction horizon will span between the minimum and maximum RUL values available in the training dataset.
