Remaining Useful Life Prediction Based on Nonlinear Tweedie Exponential Dispersion Process Considering Random Effects, Covariates, and Imperfect Maintenance

Stochastic process models can capture the stochastic dynamics, and they have been widely applied in component degradation modeling. However, the model uncertainty may lead to inaccurate and unreliable prediction of the remaining useful life. In addition, time-varying working conditions and maintenance should be incorporated into the degradation model to accurately determine the component degradation level. To solve the aforementioned problem, a generalized degradation model, based on the nonlinear Tweedie exponential dispersion process (TEDP) considering random effects, covariates, and imperfect maintenance, is developed. First, the nonlinear TEDP model with random parameters is proposed to construct the component degradation model. Considering the time-varying operating environment, the effect of the continuous and discrete covariates on the component degradation is modeled based on the proportional hazards model. The effect of the imperfect maintenance is also incorporated into the component degradation model. Then, mean-field variational inference, stochastic gradient variational inference, and expectation–maximum algorithms are used together to estimate the unknown parameters in order to improve the mathematical interpretability of the developed models. Finally, simulation studies and two real cases on gallium arsenide (GaAs) laser and hydraulic pump are used to demonstrate the effectiveness and validity of the proposed model and algorithm.