Life Stage Recognition And Life Prediction Of Rolling Bearings Based On Transfer Learning

Rolling bearings as one of the key components in the mechanical system,its working condition has a decisive impact on the safety and reliability of the entire mechanical system,once the failure will cause incalculable consequences.Such as space vehicles,rail vehicles,wind turbines and hydraulic turbines,its working condition and life depends mainly on the bearings.And because the rolling bearing working environment is generally more complex,vulnerable to high temperature,dust,frictional wear and lubrication failure,etc.,very easy to degradation.Therefore,accurately monitoring early failures and operating conditions,life stage identification and life prediction of bearings is a pressing challenge.The establishment of health indicator(HI)can reflect the running state of bearing,and then set the threshold of health indicator based on statistical learning method and normal operation data can realize the early fault monitoring of bearing.By using the powerful feature extraction ability of neural network in deep learning,combined with the corresponding classifier,the bearing life stage under the same working condition is effectively recognized.Through the feature transfer and matching ability of domain adaptive technology in transfer learning,combined with the corresponding classifier,the bearing life stage recognition under cross-working conditions is realized.By judging the bearing failure behavior,and then using transfer learning to match the failure features under the same working conditions and failure behavior,the remaining useful life(RUL)prediction of rolling bearings is realized more accurately.The bearing Prognostics and health management(PHM)system is developed through the related software development technology of BS architecture,as well as the related algorithms of bearing fault monitoring,state identification,life prediction and signal processing.The main contents of this thesis are as follows.(1)Aiming at the problem that it is difficult to characterize the features of early weak faults in the bearing operation stage,phase space reconstruction is used to reconstruct one-dimensional time series into multidimensional time series,and the nonlinear characteristics of vibration signals are mined from the multidimensional series.Then,a Phase Euclidean distance cross-correlation(PEDCC)health indicator was proposed based on cross-correlation function and Euclidean distance,and use Chebyshev’s inequality to calculate the alarm line of abnormal bearing condition monitoring based on normal samples for monitoring early bearing failure.In addition,aiming at the problem of poor generalization performance of existing health indicators,the amplitude changes of bearing envelope spectrum at different frequencies under various fault stages were analyzed,and a standard deviation weighted envelope maximum amplitude feature energy ratio(SFER)health indicator was proposed and used for bearing operating condition monitoring.(2)In view of the poor identification accuracy caused by the large amplitude difference of different bearing health indicators at the same life stage when health indicators are directly used to identify life stage,the method of directly using deep learning to identify life stage is investigated.By using the multi-scale feature extraction ability of multi-channel and different parameter convolutional layers,as well as the temporal feature extraction ability of long short term memory(LSTM)network,a multichannel convolutional long short term memory neural network(MCLNN)is built to accurately identify the life stage of bearings under the same working condition.In addition,aiming at the problem of large difference in the distribution of bearing degradation characteristics under variable working conditions and the problem of feature loss caused by the existing transfer learning methods in domain adaptation,a deep transferable convolutional autoencoder network(DTCAE)is proposed.After the feature extraction layer,the decoding layer is added to reconstruct the original data,and by measuring the reconstruction loss and reconstruction distribution difference,the features of the target domain can be retained as much as possible while reducing the distribution difference between the source domain and the target domain.(3)Aiming at the difficulty of bearing life prediction caused by the difference of feature distribution among different failure behaviors,a life prediction method based on transfer learning was studied.The relationship between health indicator and failure behavior is analyzed,and the method of determining failure behavior is proposed.The ability of attention mechanism to pay attention to important information and the ability of bi-directional long short time network(Bi-LSTM)to extract temporal features are studied,and the transferable multi-channel Bi-LSTM attention network(TMLAN)is proposed for the remaining life prediction of bearings.Based on the failure behavior determination method and TMLAN,a new life prediction framework is proposed to achieve accurate prediction of the remaining life of bearings.(4)To realize the practical engineering application of the theoretical research,a bearing PHM system with BS architecture is developed,which integrates data processing,signal analysis,early failure monitoring,fault diagnosis,life stage identification,degradation trend prediction and life prediction.Besides,the system also contains functions such as login and logout,user management,authority management and security protection.Finally,the practicability and effectiveness of the system are verified by simulating practical engineering applications.