Research On Extraction Of Multi-Domain Degradation Characterization Indicators And Life Prediction Of Rolling Bearing

With the increasing mileage of high-speed railway continuously,safety and reliability about EMU have become more and more important.The rolling bearing is a crucial component related to the safety of high-speed trains,and its health status has a direct impact on driving safety.Accurate extraction of bearing degradation indicators is the key to the identification of working mode,the judgment of defect point,and life prediction about bearing.The thesis introduces briefly the bearing structure features,failure modes and load features.Besides,the vibration features and degradation state of the bearing are mainly discussed.In this thesis,the bearing degradation characterization indicators are effectively constructed with the bearing test data.Based on these indicators,the degradation trend prediction and residual life prediction of the bearingare studied by IMS bearing data.The main tasks are as follows:1.A denoising preprocessing study combining synchrosqueezing stockwell transform and singular value decomposition algorithm is proposed.Aiming at the complex noise characteristics of bearing vibration signals,the thesis proposes an improved denoising preprocessing method based on stockwell transform and singular value decomposition.Combined with synchrosqueezing transform method,the time-frequency coefficient matrix based on synchrosqueezing stockwell transform is acquired.The differential spectrum method accurately determines the threshold.The result of the experimental signal shows that the method effectively eliminates noise and obtains a good reconstructed signal.2.Based on the principle that the sensitivity of different characteristic indicators to different types and degrees of bearing defects is inconsistent,36-dimensional indicators are extracted in the four domains including time domain,frequency domain,wavelet packet domain and entropy domain.In order to avoid errors that caused by different bearings,a relative treatment is used.Due to the repeatability of information and the superposition of noise between the indicators,the partial correlation analysis is used,and the parameter indicators that can not effectively characterize the defect type and degree are eliminated.Finally,the characteristic parameters of the mixed domain are composed of 19 dimensions.3.Based on kernel principal component analysis,the thesis builds the bearing degradation performance characterization indicators.Since the principal component analysis method can not obtain the nonlinear relationship of the data,the thesis introduces the radial basis kernel function,so the data is mapped to the space for principal component analysis,and based on over 85%of the contribution rate,two kernel principals are obtained,which effectively reduces the computational complexity.The experimental data shows that the obtained kernel principal can effectively judge the bearing of 10 different defect types and degrees.After the partial correlation analysis method,the final result is obviously better.Compared with the principal component analysis method,the kernel principal component analysis method can effectively extract the nonlinear relationship between parameters,and the proposed indicators effectively characterize bearing degradation.4.The bearing life prediction based on genetic algorithm,empirical mode decomposition and BP neural network is proposed.Aiming at the instability of the kernel principal trend,the thesis gets a kernel principal trend extraction by empirical mode decomposition.In the analysis of the bearing life cycle degradation state,five stages in the degradation process are effectively judged through trend change.The stages are composed of normal period,slight degradation,recovery period,severe degradation and sharp degradation.Due to the degraded characteristics of different stages,life prediction on account of multiple BP neural network is presented,and the thesis will use genetic algorithm to optimize BP neural network parameters.The results show that the algorithm effectively reduces the error and improves the prediction accuracy.