Research On Rolling Bearing Fault Diagnosis Method Based On Peak Envelope Spectrum Fourier Decomposition

With the gradual development of industrial equipment in the direction of intelligence,the condition monitoring of machines and equipment has been widely emphasized.Rolling bearings as the key basic components of rotating machinery,its operating condition directly affects the normal operation of machinery and equipment.It is important to carry out the research of rolling bearing fault diagnosis method and realize the intelligent monitoring of its operation status to guarantee the safe operation of machines and equipments.Using vibration signals to identify whether a rolling bearing fault has occurred is currently the main diagnostic method,but many noise interference in vibration signal will reduce the accuracy of fault diagnosis.To improve the accuracy of rolling bearing fault diagnosis,the Peak Envelope Spectrum Fourier Decomposition Method(PESFDM)is proposed in this paper,which is used to remove the noise interference in the rolling bearing vibration signal and extract its fault characteristics.Support Vector Machine(SVM)and 2Dimensional-Convolutional Neural Network(2D-CNN)are used to achieve the fault mode recognition of rolling bearings respectively.In view of the problem that the rolling bearing vibration signal is disturbed by noise,the noise reduction pre-processing method based on the Fourier decomposition of the peak envelope spectrum is proposed.The peak envelope is used to smooth the spectrum,and the spectrum segmentation boundary is determined adaptively using the improved "Locmaxmin" spectrum segmentation method,which solves the over-decomposition of Fourier decomposition method and the boundary concentration problem of empirical wavelet transform.The proposed noise reduction method is used to decompose the bearing fault simulation signal and the actual fault signal,and the effectiveness of the noise reduction performance is verified by comparing it with the three time-frequency analysis methods of empirical modal decomposition,empirical wavelet transform and Fourier decomposition.A component evaluation index combining time domain kurtosis,envelope spectral kurtosis,and correlation coefficient is proposed to select components of PESFDM for reconstruction to improve signal to noise ratio.A rolling bearing fault diagnosis method based on the combination of PESFDM and SVM is proposed for the problems of low accuracy and poor noise resistance of shallow machine learning algorithm fault diagnosis.PESFDM is used to decompose the signal and extract three eigenvalues of energy proportion of each component,singular value of matrix,and fine composite multiscale scattering entropy value of reconstructed signal.SVM is used to evaluate the fault diagnosis capability of the three eigenvalues and their different combinations,and the combination of the eigenvalues with the highest accuracy is used as the fault feature vector.The four machine learning classification methods including SVM are used to conduct fault diagnosis experiments respectively,and the effectiveness of the fault diagnosis method is verified by comparing with the two fault feature vectors of wavelet package subband energy ratio and time domain index.The superiority of the noise immunity performance is verified under bearing data sets with different levels of added noise.For the problem of poor noise immunity of 2D-CNN based fault diagnosis method,a rolling bearing fault diagnosis method based on the envelope spectrum grayscale image of PESFDM combined with 2D-CNN is proposed.The bearing signal is decomposed using PESFDM,and a feature-enhanced envelope spectrum grayscale image is designed according to the amplitude distribution characteristics of the component envelope spectrum.Using 2D-CNN to iteratively train the proposed envelope spectrum grayscale images and achieve fault diagnosis.A comprehensive comparative analysis is conducted between the proposed method and the methods using time domain grayscale images,envelope spectral grayscale images,time-frequency images,etc.as 2D-CNN inputs,and 1D-CNN method to verify the effectiveness of the proposed method.