Mechine Fault Diagnosis Method Based On Variational Bayesian Hidden Markov Models

This dissertation was supported by the National Natural Science Foundation ofChina(No.51075372,50775208),Science and Technology Projects of EducationDepartment of Jiangxi Province, China(No. GJJ12405) and The Open Fund of KeyLaboratory of Machanical Equipment Health of Hunan Province (201204). Based on thedeficiencies of the traditional and statical Independent Component Analysis in the signalblind source separation of mechanical sources,i.e. it needs the hybrid system remain thesame in the process of observation data acquisition,statistical characteristics and thenumber of independent sources remain stable，so it can’t effectively separate thenon-stationary and time-varying signals. This article is based on the excellent separationperformance of Variational Bayesian Independent Component Analysis(VbICA) tostationary signal.Here introducing Hidden Markov Models(HMM) to VbICA sourcemodel,using Variational Bayesian Hidden Markov Models(VbHMM) which is adynamic theory of independent component analysis,a new machinery fault sourceseparation methodbased on VbHMM is proposed.And its main contents include thefollowing aspects:In the first chapter, the subject and the meaning of this thesis is set forth.Thischapter summarizes the development and application of statical IndependentComponent Analysis in mechanical fault diagnosis up to now,and reviews the researchsituation of the dynamic ICA and its application in various fields especially in faultdiagnosis field.On this basis, then states the main points and the innovation of thisthesis.The second chapter states the theory and the basic algorithm of Hidden MarkovModels(HMM) which is suitable for the dynamic process of time series modeling andhas strong ability of sequential pattern classification.Variable Bayesian theory systemare discussed in this chaper, mainly including bayesian inference and variationalapproximation algorithm.The content of this chapter is the basis of the dissertation.In the third chapter,aiming at the deficiency of the traditional and staticalIndependent Component Analysis,a new separation method of machine fault sourcesbased on Variational Bayesian independent component analysis (VbICA) is proposedwhich is a kind of dynamic ICA method based on Hidden Markov Models.The methodis developed based on VbICA,it inherits the advantages of VbICA method with good separation performance.And making up the blind source separation of VbICA todynamic time-series signal with Hidden Markov Models introduced to VbICA.Thecharacteristics of the VbHMM method is that HMM can pick up the dynamic process ofstate in the underlying data generation, and capture temporal information from dynamicand nolinear source signals so as to improve the accuracy of blind source separation.The simulation result shows that the proposed method is superior to the traditionalmethod; the separation error is greatly reduced. The experiment results also verified thevalidity of the proposed method.In the fourth chapter,a new estimation method of mechanical sources based onVbHMM and automatic relevance determination(ARD) is proposed. The basic idea ofthe method is combining hidden Markov models with bayesian inference, the modelsare compared by bayesian inference,then get the optimal the number of hidden sourcesby ARD.The simulation and experiment results also verified the validity of the proposedmethod,and it has good robustness in the noise environment.In the fifth chapter,combining Local Mean Decomposition(LMD) and VbHMM,anblind source separation method of mechanical fault diagnosis based on LMD-VbHMMis proposed.In this method,the dynamic mixing signals are decomposed into a series ofPF firstly,then all the PFs components and the original signal are combined into a newobservation signal.Finally using VbHMM method for blind source separation withconsidering noise factors, and estimated source signals are acquired.For strongnonstationarity and dynamic time-varying mixing observation signal,the simulation andexperiment results also verified the validity of the proposed method.In the last chapter，the conclusions of the dissertation are summarized. Futureresearch of kernel methods is prospected.