Mechanical Fault Diagnosis Method Based On Wavelet Packet And Variational Bayesian Independent Component Analysis

This thesis is completed under the auspices of the National Natural ScienceFoundation of China (51075372,50775208), the Jiangxi Provincial Department ofEducation Science and Technology Project (No. GJJ12405) and Hunan machinery andequipment health maintenance Key Laboratory (201204) Development Fund. It aimedat the shortcomings of variational Bayesian independent component analysis theory(Variational Bayesian Independent Component Analysis, referred to as VBICA), thatseparation results of VBICA random initialization algorithm cannot maintainconsistency; the separation of mechanical failure source when it is highly correlated ispoor and traditional VBICA method cannot effectively deal with the underdeterminedblind separation of mechanical failure source, it combine wavelet packet transformand variational Bayesian independent component analysis blind separation methodeffectively, and put forward blind separation method based on wavelet packet thetransform-VBICA mechanical fault sources, and have an in-depth study of theproposed method, and the simulation and experimental results verify the effectivenessof the proposed method. The main contents include the following aspects:The first chapter summarizes research and application of variational Bayesianindependent component analysis algorithm and independent component analysisalgorithm (Independent Component Analysis, ICA), and describes the advantages anddisadvantages of VBICA algorithm. Aimed at the the inadequacies existing VBICAalgorithm, this paper discusses the significance of this topic and renders the maincontent and innovation.The second chapter discusses the theory of the independent component analysis,Bayesian modeling, variational approximation algorithms and variational Bayesianindependent component analysis, pointed out the shortage of traditional ICA methodand the advantages of VBICA methods, and comparative analysis the traditional ICAalgorithm and VBICA of methods blind source separation in noisy mixed ability bysimulation studies. The simulation results demonstrate the noise mixed separation ofthe VBICA algorithm is better than ICA method. The content of this chapter is thetheoretical basis of the entire paper. Chapter III, the model parameter initialization selection is very important inVBICA method, it directly determines the model of the final learning result. TheVBICA initialization method uses random initialization algorithm. This algorithm isinitialized randomly, therefore, the results may be different that through the differentlearning and cannot guarantee the consistency of the results. This paper presents arandom initialization method based on principal component analysis (PrincipalComponent Analysis PCA) to overcome the defects. It does a comparative analysis ofthe proposed methods and traditional VBICA algorithm. The simulation results showthat the proposed method is superior to the traditional method of VBICA, and beenreflected by the separation performance indicators, and it overcome the problem of theinstability of the separation results based on the random initialization VBICA method.Finally, the proposed method is successfully applied to the helicopter helicoptertransmission gearbox fault diagnosis, and achieved satisfactory results.Chapter IV, the experiment proved VBICA processing related blind sourceseparation results are very disappointing. This paper introduces the theory of waveletpacket and combines wavelet packet and variational Bayesian independent componentanalysis theory to solve the shortage of VBICA, and put forward blind sourceseparation method based on the wavelet packet-VBICA. This chapter renders thedetailed implementation steps of the method and the study of simulation andexperimental. Simulation results show that the proposed method is effective and madea very good separation performance. Finally, the proposed method is applied to therotor system does not rubbing fault Mans separation, and the experimental resultsconfirm the effectiveness of the proposed method.Chapter5, aimed at the existing machinery fault diagnosis based on VBICAblind separation can only deal with the case that the number of observed signals isgreater than the number of fault source, and cannot handle the case ofunderdetermined blind separation, the paper combines the advantages of waveletpacket analysis and VBICA, put forward the mechanical failure source was estimatedbased on wavelet packet-VBICA and Mount separation method. Compared with thetraditional blind source separation methods, the proposed method has the followingfeatures:(1) the proposed method can not only process overdetermined blindseparation problems, but also deal with the problem of underdetermined blindseparation;(2) the proposed method can effectively estimate the number of the source of the fault, it use Bayesian inference model specific comparison function andcombined autocorrelation measurement (ARD) to estimate a number of source signals;(3) The proposed method maked full use of wavelet packet analysis can effectivelydeal with non-stationary signals, the blind separation of non-stationary signals.Simulation results show that the proposed method achieved satisfactory separation.Finally, the proposed method applied to the bearing inner ring fault blind sourceseparation, and succeeded in isolating the failure of the bearing inner ring and theouter ring fault feature.Chapter VI summarizes the main research content and innovation, and looksforward the further research prospects.