Research On Fault Diagnosis Method Of Rolling Bearing Based On Manifold Learning And Optimized Extreme Learning Machine

As an important component of rotating machinery,the running state of rolling bearings will directly affect its safety and stability.Therefore,it is particularly important to locate and eliminate the fault,especially the early fault of rolling bearing.The early fault signal of the rolling bearing is a typical nonlinear and non-stationary signal,which is not obvious in characteristics and is easily submerged by noise.Therefore,the signal needs to be denoised to improve the accuracy of fault identification.Therefore,this thesis"Research on Fault Diagnosis Method of Rolling Bearing Based on Manifold Learning and Optimized Extreme Learning Machine" come into being.In this thesis,the fault signals of inner ring,outer ring and rolling element of rolling bearing are researched.Based on manifold learning,variational mode decomposition and optimized limit learning machine,the research on signal denoising,fault feature extraction and pattern recognition is carried out.The main research contents are as follows:The research background,significance and development history of rolling bearing fault diagnosis are expounded.The research status of rolling bearing fault diagnosis is introduced.The causes and vibration mechanism of rolling bearing fault are analyzed.The fault characteristic frequency of rolling bearing is calculated,and the current commonly used fault feature extraction method and pattern recognition method.The manifold learning theory is studied.Several commonly used manifold learning algorithms such as equidistant mapping,local tangent space permutation algorithm,local preservation projection algorithm,Laplacian operator mapping algorithm and local linear embedding algorithm are introduced.The typical nonlinear dimensionality reduction algorithms such as LTSA,LEE,ISOMAP are selected.And linear dimensionality reduction algorithm such as the principal component analysis algorithm,denoises the fault signal,and uses the root mean square value and approximate entropy as the quantitative evaluation index.The noise results are quantitatively evaluated.The experimental results show that LTSA has great advantages in the denoising of rolling bearing fault signals.Then the influence of parameters on the variational mode decomposition is analyzed and simulated.Three different types of simulation signals are selected.The VMD and empirical mode decomposition,the collective empirical mode decomposition and the complete overall empirical mode decomposition are compared.The decomposition effects under noise conditions,different frequency combinations and similar characteristic frequencies,and modal aliasing,false components,over-decomposition and so on are analyzed.Aiming at the influence of parameters on the manifold learning dimension reduction algorithm,an adaptive manifold learning algorithm based on the grid search method is proposed to reduce the noise of the bearing signal.The VMD decomposition of the denoised signal is carried out.The modal component selection and reconstruction method based on kurtosis,correlation and envelope entropy are proposed.The optimal modal reconstruction component is obtained.Then,the envelope spectrum analysis is performed on the optimal modal component,and the fault characteristic frequency is extracted.Perform singular value decomposition on the matrix composed of modal components is proposed,and the signal singular eigenvalues is obtained as the data source of the pattern recognition algorithm training and test sample dataset.The influence of the number of hidden layer nodes and the activation function on the recognition accuracy of the extreme learning machine pattern are analyzed.The optimal parameter combination is selected as the parameter input of the ELM algorithm by using the leapfrog algorithm.Furthermore,the accuracy of fault pattern recognition under different fault size data sets,different pattern recognition methods and different fault feature extraction methods are studied,and the accuracy of pattern recognition when the number of test samples and the number of training samples are simulated.