A hybrid prognostic framework: Stochastic degradation process with adaptive trajectory learning to transfer historical health knowledge

Abstract

Remaining useful life (RUL) prediction is crucial to supporting intelligent maintenance and health management of safety–critical products. Although advanced data-driven approaches such as deep neural networks are effective in processing high-dimensional non-linear health features, their application to field RUL prediction confronts with two challenges: (a) adaptivity of the lifetime parameter learning process is often restricted, and (b) prediction of multi-source uncertainties is almost analytically intractable. This paper addresses such challenges by devising a tractable, global adaptive model-data-interaction prognostic framework, where a non-linear stochastic degradation model governed by self-adaptive trajectory pattern is constructed to transfer historical health knowledge. In particular, a joint parameter learning framework is established under the structure of a multi-branch Bayesian network, such that to simultaneously learn: (a) degradation model parameters, and (b) network hyper-parameters. Additionally, the key control parameters of the degradation process are updated adaptively leveraging multi-dimensional sequential Bayesian learning. An efficient interpolation algorithm is further proposed to alleviate computation burden of RUL distributions. Case studies conducted on both turbofan engines degradation data and field train bearing vibration data demonstrate the superior model performance compared to existing methodologies.