Study Of Rolling Bearing Fault Diagnosis Using Modified VMD,ELM And VPMCD

The manual and intelligent fault diagnosis methods of high-speed train's rolling bearings are investigated,respectively.The manual fault diagnosis was performed based on the combination of the modified variational mode decomposition(VMD)and Hilbert envelope spectrumn analysis.The intelligent fault diagnosis was performed based on the modified extreme learning machine(ELM)and modified variable predictive model based class discrimination(VPMCD).First,the overall theoretical framework of VMD was explained.Then,several groups of simulation signals were analyzed by VMD and empirical mode decomposition(EMD)to illustrate the superiorities including mode mixing resistance,illusive components avoidance and noise robustness of VMD over EMD.However,VMD has two drawbacks: the lack of adaptive selection of decomposition parameters and the end point effect.To sovle these two problems,adaptive VMD(AVMD)based on the kurtosis criterion and the end point extension scheme based on distance correlation coefficients were proposed.The effectiveness of these two proposals was verified by simulation signals.Application to real world vibration signals of rolling bearings showed that the proposed AVMD effectively enhancs the quality of the envelope spectrum and the diagnosis performance of bearing faults.The extreme learning machine(ELM)has extremely fast learning speed,good generalization capability and universal approximation ability.However,ELM has two shortcomings namely the instability of prediction results and the complex procedures needed to decide the best value of the number of hidden layer neurons.In order to overcome these shortcomings,the article introduced an algorithm which does not need to choose the best value of hidden layer neurons,namely the error minimized extreme learning machine(EM-ELM),upon which an ensemble error minimized extreme learning machine(EEM-ELM)algorithm was proposed.This algorithm does not need to determine the best value of hidden layer neurons and is able to enhance the stability of the prediction results of ELM.Due to the resonable weighting method applied,the testing accuracy of this algorithm can be higher than the member EM-ELM classifiers.On application to experimental data,the fault diagnosis model which utilized EEM-ELM as classifier using the normalized energy and normalized permutation entropy of the intrinsic mode functions(IMF)obtained by application of VMD to fault signals as characteristic vector reached a testing accuracy over 95% over a wide range of algorithm parameters' value.This fact indicates that this fault diagnosis model can successfully recognize complex faults of bearings.To exempt from the tediousness caused by searching the best parameters of classification algorithms,the variable predictive model based class discrimination(VPMCD)algorithm without any need of parameter tuning,was introduced.However,VPMCD suffers from several difficultites like the choice of regression models,the inaccuracy of the regression models,the overfitting of the regression models and the multiple colinearity among the predictive variables of the regression models and its square terms and interaction terms.To cope with these difficulties,a modification of VPMCD(VPELMCD)using ELM neural network as its regression models was first proposed by the author.Experiments revealed that compared to linear model,linear interaction model,quadratic model and quadratic interaction model,use of ELM as regression models leads to a significantly higher clssification accuracy of VPELMCD than VPMCD.Next,aiming at the difficulties of the selection of regression models' variables,the inaccuracy of regression models and the overfitting of regression models,the author proposed a variable selection scheme based on distance correlation coefficients.Experiments verified that this variable selection scheme effectively enhanced the accuracy of VPELMCD.