Research On Fault Feature Extraction Of Rolling Bearing Based On T-SNE And Sparse Representation

With the rapid development of science and technology,the update speed of modern industrial machinery and equipment is accelerating,and it presents diversity.The inspection and maintenance of large mechanical equipment is facing greater challenges in such a development background.Rolling bearing is widely used in modern mechanical equipment,and its operation quality directly affects the normal operation of the machine.Therefore,it is one of the important tasks to find the fault location and extract the fault features accurately.This thesis starts with the measured fault signals of rolling bearing,mainly including the signals of inner ring,outer ring and rolling element.The fault diagnosis of rolling bearing is realized by signal denoising,fault feature extraction and fault feature clustering.Therefore,this thesis “Research on Fault Feature Extraction of Rolling Bearing Based on T-SNE and Sparse Representation” is put forward.The main contents of this thesis are as follows:This thesis expounds the significance and purpose of the research on fault feature extraction of rolling bearing.The development process of fault feature extraction of rolling bearing is introduced.The research status of fault feature extraction technology of rolling bearing is analyzed systematically,and the mechanism principle and basic fault frequency of rolling bearing are explained.The feature extraction method and fault clustering method of rolling bearing fault signal are expounded theoretically and systematically.In view of the preprocessing of the original time series signal,the phase space reconstruction technology is studied theoretically.The influence of reconstruction parameters on the reconstruction phase space is analyzed.The two parameters,time delay and embedding dimension,are determined for phase space reconstruction by combining C-C algorithm and Cao algorithm.Fuzzy entropy and root mean square(RMS)are introduced as the quantitative indexes to evaluate the reconstruction effect of the proposed method when it is used alone with C-C algorithm and Cao algorithm.The results show that the method of phase space reconstruction parameter selection proposed is more suitable for the fault feature extraction of rolling bearing.In the part of signal denoising,the manifold learning algorithms are analyzed systematically.This thesis introduces the current manifold learning algorithms,and typical nonlinear dimensionality reduction algorithms is selected,inluding local tangent space algorithm(LTSA),local linear embedding(LLE),t-distributed stochastic neighbor embedding(t-SNE),and principal component analysis(PCA)in linear dimensionality reduction algorithm to denoise the fault signal respectively.The correlation coefficient,envelope entropy and kurtosis are used to evaluate the noise reduction results.The results show that t-SNE algorithm has more advantages in denoising rolling bearing fault signals.At the same time,the adaptability and visualization effect of multi-dimensional scaling(MDS)algorithm and t-SNE algorithm in nonlinear data are compared and analyzed.In order to obtain the optimal sparse representation of denoised signal and enhance the effect of fault feature extraction,the construction of learning dictionary,solution algorithms and dictionary learning algorithms are systematically studied.Compared with wavelet dictionary and discrete dosine transform(DCT)dictionary,it is found that DCT dictionary has richer atomic structure and higher computational efficiency.In this thesis,two sparse algorithms,matching pursuit(MP)and orthogonal matching pursuit(OMP),are deeply studied.Through the comparison and analysis of the two algorithms,the result shows that OMP algorithm has the advantage of faster convergence speed in rolling bearing fault feature extraction.The computational complexity of different dictionary learning algorithms of MOD,K-SVD and alternative direction method of multiplier(ADMM)is compared and analyzed.It is found that the learning speed of ADMM algorithm is faster.and the reconstruction error obtained by ADMM algorithm is significantly reduced compared with that obtained by no dictionary learning method.Therefore,this thesis chooses DCT dictionary,ADMM algorithm and OMP algorithm to decompose and reconstruct the fault signal sparsely,so as to establish the optimal sparse representation of the denoised singal.Finally,the fault frequency of rolling bearing is extracted by envelope demodulation analysis,and the results are compared and verified.In order to realize the visualization effect of fault feature extraction of rolling bearing,K-means clustering is used to cluster the fault signals of rolling bearing.The clustering results are compared with different fault feature extraction methods by davidson baoding index(DBI),and the results of rolling bearing fault clustering are compared from different fault size sample data sets and different length sample data.The results verify the superiority of the fault feature extraction method in clustering analysis.