Research On Remaining Useful Life Estimation Methods For The Key Components Of Equipment Based On Data-driven Methods And Bayesian Theory

As the integration of mechanical equipment is getting higher and higher,its failure modes are more and more complex and any sudden operation failure may cause catastrophic accidents.For modern mechanical equipment,it not only need to improve the reliability of equipment at the design stage,but also need to ensure the high operational reliability of equipment at the service stage.In recent years,the remaining useful life prediction has gradually become one of the important technologies to ensure the reliable operation of equipment.Because it does not rely on failure mechanism models and can effectively characterize the uncertainty of prediction,remaining useful life prediction methods based on data-driven methods and Bayesian theory have been rapidly developed.Although this kind of remaining useful life prediction methods has received more and more attention,more researches in this field still need to be developed and improved,especially for the remaining life prediction of the key components under complex conditions.For example,a series of technical problems still remain to be solved in the remaining useful life prediction problems under multi-source information,multi-failure modes,multi-operating modes,multiple uncertain factors and so on.This dissertation conducts a systematic research on the remaining useful life prediction for the key components of equipment based on data-driven methods and Bayesian theory,and aims at developing some exploratory work in data-driven modeling and uncertainty quantification of prediction.The research content is mainly based on data-driven methods and Bayesian theory.By considering the multiple sources of information,the diversity of failure modes and the uncertainty of prediction,three difficult points are further studied,including predictive model construction with robust prediction under single source information and single failure mode,predictive model construction with model update under multi-source information and single failure mode,and predictive model construction with model solving under multi-source information and multi-failure modes.We hope this dissertation can provide guidance and reference for the remaining useful life prediction of the key components of equipment under different information sources,different failure modes and different operating conditions,and can be applied to the remaining useful life prediction for the key components of aircraft,high-speed rail and nuclear power units in the future.Based on this idea,the following studies are carried out in this dissertation:(1)A remaining useful life prediction method is proposed by an improved Bayesian filter for the key components of equipment under single source information and single failure mode,which realizes the predictive model construction with monitoring data of single state quantity and robustly predicts the remaining useful life of single key component under single failure mode.First of all,an improved unscented Kalman filter is established in the proposed method based on Kalman filter and linear adaptive strategy to construct a predictive model of the degradation indicator of single key component,which can adaptively adjust the noise term of the model.Then,considering the uncertainty of the model or algorithm parameters,the established predictive model is used to robustly predict the remaining useful life of single key component.In other words,when the initial values of the model or algorithm parameters fluctuate,the prediction result of the proposed method just fluctuate within a small range.Finally,the single condition and multi-condition degradation problems under single source information and single failure mode are analyzed respectively to verify the effectiveness of the proposed method.(2)A remaining useful life prediction method is developed by data fusion and Bayesian theory for the key components of equipment under multi-source information and single failure mode,which realizes the construction and update of the predictive model using multi-source condition monitoring information,and predicts the remaining useful life of single key component under single failure mode.Firstly,based on the principal component analysis method,a health indicator is established by fusing historical monitoring data of multi-source state quantities,which is used to characterize the degradation degree of single key component.Secondly,based on the negative time scale modeling method,a data-driven state model is constructed to characterize the time-varying law of the health indicator,in which the life of the key component is used as the parameter of the state model.Then,a Bayesian update model of the state model parameters is established by the Bayesian theory,and real-time monitoring data of equipment and Markov Chain Monte Carlo method are used together to update the model parameters and predict the remaining useful life of the key component.Finally,the single condition and multi-condition degradation problems under multi-source information and single failure mode are utilized to verify the effectiveness of the proposed method.(3)A remaining useful life prediction method is studied based on Bayesian deep learning for the key components of equipment under multi-source information and multi-failure modes,which realizes the construction and efficient solution of the predictive model under the Bayesian framework,and can directly use the multisource condition monitoring information of equipment to predict the remaining useful life distribution of the key components under multi-failure modes.Firstly,based on the end-to-end modeling method,the remaining useful life predictive model is established by a suitable deep convolutional neural network,and the parameters of the predictive model are taken as random variables under the Bayesian framework;Secondly,based on the historical monitoring data of some same or similar equipment,a non-parametric variational inference algorithm is used to efficiently solve the posterior distribution of the parameters of the predictive model;Then,based on the real-time monitoring data of the equipment under test,the predictive model can provide the remaining useful life distribution of the key components to quantify the uncertainty of the prediction result.Finally,the single condition and multi-condition degradation problems under multi-source information and multi-failure modes are analyzed respectively to verify the effectiveness of the proposed method.(4)An improved remaining useful life prediction method is constructed for the key components of equipment under multi-source information and multi-failure modes,which can integrate the performance index function of a predictive model to improve its prediction accuracy.Firstly,a deep convolutional neural network is used to directly construct the remaining useful life predictive model under the Bayesian framework;Secondly,based on the performance index function of the predictive model,a flexible optimal target function is constructed to solve the predictive model,which can adapt to different remaining useful life prediction problems;Then,in order to efficiently solve the model under the above optimal target function,a non-parametric variational inference algorithm is developed to solve the posterior distribution of the parameters of the predictive model;Finally,the single condition and multi-condition degradation problems under multi-source information and multi-failure modes are studied respectively,and it is verified that the improved method can obtain a predictive model with higher precision.