An adaptive unscented Kalman filter –based method for RUL prediction via nonlinear degradation modeling

Remaining useful life (RUL) prediction based on degradation path is regarded as a core component in prognostics and health management (PHM). Current methods in literature for nonlinear degradation mainly rely on linearization modeling, and a few nonlinear degradation methods based on nonlinear modeling often utilize Wiener process with increment modeling and linear Kalman filter structure, which would make the results biased under strong nonlinearity. This paper proposes an adaptive unscented Kalman filter (UKF)-based prognostic method for RUL prediction in nonlinear degradation systems. To be specific, a general nonlinear state-space model is established for nonlinear degradation path directly to describe the dynamics and nonlinearity of the degradation process instead of increment modelling or linearization. Then, nonlinear UKF structure is used to update the degradation parameters forward once newly observed data is obtained. To capture the uncertainty and variability during degradation process adaptively, a novel Bayesian paradigm, which consists of an improved unscented Rauch-Tung-Striebel Smoother (URTS) and expectation maximization framework, is derived to update the hidden variables and other unknown parameters recursively. Finally, two real-life nonlinear degradation cases consisting of a bearing degradation and a hard disk drive degradation are used to verify the superiority of the proposed method.