Research On Fault Diagnosis Of Bearings Under Missing Data

Bearing failure,as a common fault in the operation of a motor,is difficult to be found in time at the beginning.Long-term operation of bearing failure will bring hidden dangers to the normal operation of the motor system,even causing overall damage or scrap of a motor in serious cases,which will threaten human life and property.Therefore,accurate and timely bearing fault diagnosis has important practical significance.Traditional bearing fault diagnosis methods rely on experts’ advanced experience to design diagnostic schemes.Complex background noise seriously interferes with the extraction of fault features.As a result,it is difficult for traditional methods to establish accurate mathematical models for bearing fault diagnosis.However,bearing fault diagnosis technology based on deep learning(DL)can still accurately achieve bearing fault diagnosis using the collected data without an accurate mathematical model,which makes it more and more popular with researchers.However,in the actual data collection process,due to sensor failure and information conversion failure,data loss occurs from time to time.The accuracy and validity of fault diagnosis will be significantly reduced by using the missing data to train and detect in the DL model.The bearing fault diagnosis with missing data needs to be studied urgently.At present,the method of filling up missing data and then using the DL model for fault diagnosis can achieve better diagnostic results with lower data missing rate.However,there are still the following problems:(1)With a large data missing rate,the existing data filling methods have a large loss of performance,and it is difficult to accurately achieve bearing fault diagnosis;(2)Frequency domain analysis has the strong anti-noise ability,so it is difficult for existing methods to recover the features of missing data in the frequency domain and to diagnose faults.Because of the above problems,the main research contents of this paper are as follows:A bearing fault diagnosis method combining singular value thresholding(SVT)and one-dimensional convolutional neural network(1D convolutional neural network,1DCNN)is presented to overcome the problem that traditional methods to fill missing data cannot effectively recover data characteristics and realize fault diagnosis under DL fault diagnosis model.This method takes into account the high feature reduction property of SVT in restoring missing low-rank matrices and uses it to restore low-rank missing bearing vibration signal data,which improves the accuracy of extracting missing data features in the DL model to some extent.However,1DCNN is convenient and efficient,it can take bearing data collected in the industrial field as input without additional pre-processing,and has fewer parameters and higher structure efficiency when building network models,which makes it more advantageous in bearing fault diagnosis than other DL models.Experiments were performed to compare the accuracy of data recovered by SVT filling,mean filling,median filling,local maximum filling,and KNN neighbor filling under the 1DCNN bearing fault diagnosis model.The experimental results show that the SVT recovered data has a higher diagnostic accuracy under the 1DCNN model with a higher ratio of missing data.A bearing fault diagnosis method based on sparse Bayesian learning(SBL)and 1DCNN is presented to solve the problem that traditional signal processing methods cannot recover the missing data in the frequency domain and diagnose the fault.By utilizing the sparsity of the fault feature frequency component in the envelope spectrum,this method converts the problem of fault feature frequency estimation into a sparse frequency recovery problem and directly restores the envelope spectrum of the fault pulse signal.Considering that the amplitude of the harmonic component of the fault feature decreases sharply with the increase of frequency in the spectrum,resulting in the frequency of the fault feature concentrating on a specific frequency band.The proposed SBL method uses a uniform grid to cover a specific truncation frequency domain to reduce the computational complexity of the SBL.In addition,for the mismatch between the fixed grid and the true fault feature frequency,the method proposed an off-line model to bridge the model error,improve the accuracy of envelope spectrum data recovery,and finally,diagnose the fault using the recovered envelope spectrum data as the input of 1DCNN.In the experiment,the fast Fourier transform(FFT),envelope method(ENV),and SBL method mentioned in this chapter are used to process bearing vibration signal data with low and high missing rates respectively,and the processed data is used as input of 1DCNN for fault diagnosis by the classifier.The simulation experiments and results show that the envelope spectrum data recovered by the proposed method contains more information about the fault characteristics and is more convenient for accurate bearing fault diagnosis under the DL model.