A Comparison Study of Methods for Parameter Estimation in the Physics-based Prognostics

Prediction of remaining useful life of a system is important for safety and maintenance scheduling. In the physics-based prognostics, the accuracy of predicted remaining useful life is directly related to that of estimated model parameters. It, however, is not a simple task to estimate the model parameters because most real systems have multivariate model parameters, which are often correlated each other. This paper mainly discusses the difference in estimating model parameters among different prognostics methods: the particle filter method, the overall Bayesian method, and the incremental Bayesian method. These methods are based on the same theoretical foundation, Bayesian inference, but they are different from each other in the sampling scheme and/or uncertainty analysis process. A simple analytical example and the Paris model for crack growth are used to demonstrate the difference among the three methods in terms of prognostics metrics. The numerical results show that particle filter and overall Bayesian methods outperform the incremental Bayesian method. Even though the particle filter shows slightly better results in terms of prognostics metrics, the overall Bayesian method is efficient when batch data exist.