Operating Condition Assessment And Remaining Life Prediction Of Rolling Bearings Based On The Life-cycle Data

As critical support components in the rotating machines,the design reliability in the early stage and the operating condition of rolling bearings in use have great impacts on the personal safety,production efficiency and economic benefits.Hence,the operation risk and the maintenance difficulty will geatly be reduced if the reliability of the batch is evaluated to check whether the quality satisfies the operating conditions before putting into use.While in operation,Prognostics and Health Management makes the maintenance more initiative through predicting the performance state trend of mechanical parts,and avoids the passive maintenance such as scheduled maintenance and condition maintenance in the past.Based on the life-cycle data of rolling bearings,the thesis expands the theoretical framework of health condition evaluation.Firstly,the probability statistical analysis is carried out on the life data;in further,the fault monitoring in the run,the fault location and remaining life prediction are investigated with artificial intelligence algorithms based on the fault features extracted from the vibration signals.The type-II generalized logistic distribution and Tukey's g-and-h distributon are introduced firstly for statistics analysis of the bearing lifetime data due to the fault tolerant ability.For the distinct characteristics of the two distributions,their parameters are estimated through the particle swarm optimization algorithm but with different objective functions,namely the negative maximum likelihood function of the type-II generalized logistic distribution and the mean square error between the quantiles of Tukey's g-and-h distribution and the actual values.Comparisons are performed among normal distribution,Weibull distribution,type-II generalized logistic distribution and Tukey's g-and-h distribution with the lifetime data of bearing 104 and bearing 7208,and the influences of different distributions for the bearing reliability are discussed.Moreover,the model of the P-S-N curve concerning accelerated life data are established with the location-scale model based on the type-II generalized logistic distribution and Tukey's g-and-h distribution,and the analytical expression based on the type-II generalized logistic distribution and the quantile expression based on Tukey's g-and-h distribution are obtained,respectively.The experimental results show that the fitting accuracy of Weibull distribution,type-II generalized logistic distribution and Tukey ' s g-and-h distribution are close,but type-II generalized logistic distribution is more flexible in applications,while Tukey ' s g-and-h distribution is suitable for quantile life calculation and prediction.A combined approach based on orthogonal locality preserving projection as a manifold learning algorithm and support vector data description is proposed for the performance monitoring of the rolling bearing with healthy samples only.Firstly,the statistic features including time domain features and Kolmogrov-Smirnov test statistics are extracted from the vibration signals.The features are in itially selected according to the correlation to the root mean square and standardized,then orthogonal locality preserving projection is applied to reduce the dimensionality.Finally,support vector data description is utilized to perform the information fusion and the performance degradation indicator is obtained to implement the operating condition monitoring.Two groups of rolling bearing life-cycle data respectively from the Intelligent Manufacturing Systems Center in the University of Cincinnati and Hangzhou Bearing Test & Research Center are used to validate the proposed method and the results show the effectiveness of the proposed method for the early fault monitoring.Besides,the bandwith of the Gaussian kernel function in the support vector data description is optimized with ergodicity based on the life-cycle data,which could maximize the correlation between the performance degradation indicator and time.This degradation indicator with the optimized bandwith will lay the foundation for the following remaining life prediction.In view of the non-stationary and nonlinear signals in early fault stage of rolling bearing,as well as the existence of strong noise and harmonic interference,adaptive local iterative filtering is introduced to remove the influences of the background noise and the non-stationary signals,and then the fault vibration signals could be analyzed at different scales.The simulation has shown that adaptive local iterative filtering could effectively relieve the shortcomings of mode mixing in EMD.In the following,the intrinsic mode function with the most abundant fault information by adaptive local iterative filtering is selected based on the principle of maximum kurtosis.From the comparison of Hilbert transform,Teager energy operator,the analytical energy operator and the high order analytical energy operators from two to four multipes by simulation,the advances of the four-multiple high order analytical energy operator in enhancement of the fault impulse transient signals and efficient suppression of the noise is illustrated.In the next,the envelope demodulation of the intrinsic mode function with the maximum kurtosis is carried out with the fourth multiple analytical energy operator,and the envelop spectrum analysis is performed in the following with the fast Fourier transform.Hence,the early fault diagnosis of the rolling bearing is realized by comparison of t he fault frequency and the fault characteristic frequency which is an analytical solution.Finally,the superiority of the proposed method is illustrated through the frequency analysis of abnormal points of the test bearings in the state monitoring,while the accuracy of the state monitoring method is verified in further.To modify the traditional multi-fault classification methods for rolling bearing with three steps,namely,feature extraction,feature reduction and fault classification,an approach based on the multi-scale entropy jointing with kernel sparse representation classification algorithm is proposed.Kernel sparse representation classification algorithm could select features independently with respect to different types of faults,so as to better perform fault classification.Its advantage lies in that it could improve the classification accuracy through the higher dimensionality of training samples.In addition,it avoids the problems of strategy of classification in support vector machine and the requirement of large samples in artificial neural network.Finally,the validity and the accuracy of the proposed method are verified by the data of fault bearings in the impeller pump and the data of artificial fault bearings from CWRU.The results show that at the condition of small samples,the classification accuracy of kernel sparse representation classification for bearing fault types could be advanced along with the increase of the feature dimensionality.The ensemble learning algorithm is investigated for remaining useful life prediction considering the the fluctuation of performance degradation indicator.Firstly,non-repeating samples in the degradation interval are randomly extracted,and the samples are smoothed with the bathtub curve model of Weibull distribution,while the parameters in the model are identified with the PSO algorithm.Secondly,the fitted curves are treated as training input samples,while another group of non-repeating samples in the degradation interval is randomly extracted again as validation input samples.The corresponding outputs for both of them are the lifetime percentage.They are used to respectively train the LSSVM and RVM.In the following,the four kinds of errors as mean relative error,root mean square error,mean absolute error and normalized mean square error are calculated.Next,the weighting coefficients of LSSVM and RVM for RUL prediction are calculated respectively,and the final RUL prediction is obtained through the ensemble learning.The proposed method is verified by the life-cycle data of rolling bearings.The results show that the smoothed performance degradation indicator could efficiently reduce the impact of the bearing degradation fluctuation on the prediction accuracy;the long range prediction of RUL can be realized through the similar measurement between the training sample and the test samples by machine learning algorithm;the ensemble learning algorithm can make up for the shortcomings of a sin gle machine learning algorithm,and with the increase of machine learning algorithms,the predicted results will be closer to the real values.Therefore,the prediction results of the proposed method are robust.