Research On Remaining Useful Life Prediction For Non-stationary Degradation Process

Aiming at ensuring the reliability and safety of industrial devices when they are running and reducing maintenance costs,prognostics and health management(PHM)has been widespreadly concerned and provided a profound impact on industries in the last decades.Remaining useful life(RUL)prediction is critical to PHM,which would support the maintenance decision-making in real-time.To predict the RUL accurately some reasonable models should be investigated,with which the degradation rule could be characterized.However,because of the increasing complexity in industrial devices,it's not easy to develop a model based on physical degradation mechanism.In this situation,the data-driven prediction approach has shown its advantage whatever in academia or practice thanks to the development of sensing technologies.In operation,the devices suffer from the internal defects,the changes from environments,the various loads and so on,which would result in non-stationary degradation.Under such complex operational conditions,this thesis focuses on how to establish more reasonable degradation models and provide more practical solutions for RUL prediction.Firstly,we review some recent works about RUL prediction,specially introduce two kinds of prediction approaches based on the regression models with random coefficients and the stochastic process models.Then the related works about non-stationary degradation processes are discussed in detail.Three different degradation models based on wiener process are proposed for differen-t degradation situations,i.e.a degradation model with change point,a degradation model with known shocks and a hybrid degradation model with unknown shocks.According to the character-istics of each model,the estimation algorithms for the hidden states and the models' parameters are developed based on the Bayesian framework.The RUL prediction algorithms are given sub-sequently and some experiments are used to verify the accuracy of the predicted results.The degradation models proposed here can be used for the real-time condition monitoring and RUL prediction under the three conditions respectively,which would provide some prospective degra-dation information for predictive maintenance.At last,we summarize the thesis and give a brief account of the future research.