Data-driven Early Diagnosis And Residual Useful Life Prediction Of Slowly Varying Small Fault

Once catastrophic failure occurred in a large-scale complex systems,it will result in huge economic losses and casualties.Different from the planned maintenance scheme,the condition based maintenance based on early fault diagnosis and residual useful life(RUL)prediction can not only reduce unnecessary downtime losses,but also can well ensure the safety operation of the system.The research background of this paper is to secure the safety operation of industry process and to save the maintenance costs.To achieve this goal,early diagnosis of slowly varying small faults is a fundamental step.The advanced feature extraction theory and prediction technology are used to develop a set of RUL prediction methods.This research result of thesis is helpful for the maintenance operator make a maintenance decision.The theme of this thesis is to use some data-driven fault feature extraction tools to develop a set of RUL prediction methods.Feature extraction tools used can be listed as follows: wavelet filtering(WF),principal component analysis(PCA),designated component analysis(DCA),deep learning(DL)etc.Availability of the fault data,knowledge of the fault symptom relation,statistical distribution of the observation data and nonlinearity of the observation data etc.,are the possible prior information needed to develop RUL prediction methods on four aspects: system-level and component-level RUL prediction based on nonlinear fitting of statistical characteristic,RUL prediction based on machine learning of deep characteristics.The main innovations are follows as:(1)In the case of Gaussian observation data,wavelet filtering and principal component analysis are combined to extract the insignificant feature of early slowly varying fault.Based on the fault precursor developed in the first step,a real-time RUL prediction model using exponential nonlinear fitting method is established to realize system-level RUL prediction.When the knowledge of fault symptom relation is available,designated component analysis can be used to extract fault feature of some key components and component-level RUL prediction method can be developed.(2)When the observation data is nonlinear,the statistical distribution and the fault symptom is unknown,deep learning is used to extract the potential insignificant fault feature involved in theobservation to realize early fault diagnosis.Then machine learning method is used to establish a direct mapping relationship between the high-dimensional fault feature extracted by the deep neural network(DNN)and the actual RUL.