Research On Feature Dimension Reduction And State Recognition Method Of Rolling Bearing Vibration Signal

As one of the indispensable components in many rotating machinery,rolling bearings are widely used in industrial production and have the title of"machine joints".The normal operation of the rolling bearings have a serious impact on the running state of the entire mechanical system.If the rolling bearings fail,they will cause serious economic losses and even cause casualties.If it is possible to effectively identify whether the running state of the rolling bearing is abnormal and the location of the abnormal point,it can be repaired according to the situation and proactively maintained,which can reduce economic losses and casualties to a considerable extent.In this paper,rolling bearings are established as the research object,and feature extraction,feature screening and dimension reduction are carried out for their vibration signals at the multi-domain level,and then the relevant classification algorithms in machine learning are used to realize the state recognition of rolling bearings,and particle swarm algorithm is applied to the classification algorithm for further optimization.The method is verified by using the rolling bearing public dataset of Case Western Reserve University and the real fault dataset of rolling bearing of Xi’an Jiaotong University to demonstrate the effectiveness of the proposed method.The main work of the thesis includes:（1）Research on the fault vibration mechanism of rolling bearing and the method of signal feature extraction.The characteristic frequencies of faults in different parts of the rolling bearing are calculated,and the vibration signal of the rolling bearing is analyzed by using the envelope spectrum.The single-domain feature extraction is performed on the rolling bearing vibration signals in the time domain and frequency domain respectively,16 for the former and 13 for the latter,and the distance evaluation technology is used to screen and evaluate 29 features.Then,the time-frequency domain features are extracted based on the CEEMD multi-scale dispersion entropy,and the time-frequency domain features and the screened time and frequency domain features are assembled into a feature matrix.Finally,a160×8 feature matrix T is obtained,and its expression isT（28）[α₁α₂β₁β₂β₃β₄φ₁φ₂].（2）Multi-domain feature dimensionality reduction and cluster analysis of rolling bearing vibration signals.Use PCA,KPCA,and LDA to reduce the dimension of the feature matrix to avoid redundant data.Then,fuzzy mean clustering is used to cluster the dimensionality-reduced feature matrix,and the clustering effect is evaluated in terms of clustering accuracy,membership matrix and iteration times.The results show that the clustering accuracy rates of the three dimensionality reduction algorithms are 92.13%,43.13%,and 96.78%,respectively.Among the three membership curves,LDA is the most stable.Convergence on 11th and 6th.From three aspects,the effect of LDA is the best among the three.（3）Research on the state recognition method of rolling bearings based on machine learning.SVM,K-nearest neighbor and ELM are used to identify the state of rolling bearings according to the feature matrix T after dimension reduction.The accuracy rates of the three algorithms are 91.88%,94.38%and 95.63%,respectively.Then the PSO algorithm is used to optimize the method with the best recognition effect among the three,namely ELM.The accuracy rates before and after optimization are 95.63%and 98.75%,respectively.Further,the method is verified by using the rolling bearing dataset of Case Western Reserve University and the real fault dataset of rolling bearing of Xi’an Jiaotong University.For the former,when the load is 0,1,2 and 3 HP,the recognition accuracy of PSO-ELM is 96.00%,97.00%,96.00%and 94.50%respectively.For the latter,the recognition accuracy of PSO-ELM under working conditions 1,2 and 3 is 94.17%,96.67%and 95.83%respectively.