Model-based prognostics of gear health using stochastic non-linear dynamical models

This paper presents a probabilistic approach towards the estimation of damage process dynamics, which is then used for prediction of the remaining useful life of the system. The dynamics of damage propagation can be modelled as a non-linear stochastic dynamical state-space model, where system states are not directly accessible and all information about them has to be inferred from the output data. The paper employs an iterative procedure for calculating the maximum-likelihood estimate of model parameters, based on the expectation-maximisation algorithm together with the unscented Kalman filter for the estimation of system states. After a new feature value is obtained, the algorithm updates the model parameters and a more accurate prediction can thus be made. The algorithm has been used to predict the remaining useful life of a single-stage gearbox system. Feature values have been computed from the vibration signals and were processed using Hilbert transform. Several test runs of the system have been performed and the data collected from them was used for the validation of the algorithm.