Recognition And Fault Prediction Based On The Degenerate State HMM

Automation and intelligent level of modern mechanical equipment is moreand more advanced,its development has a profound impact on industry andeconomy.With the development of condition based on maintenance (CBM) andPrediction and Health Management (PHM) and other maintenance theory andtechnology,in recent years, research on the state real-time monitoring,statusinformation and processing technology,fault diagnosis and prediction technologyhas become a study hotspot.And in the field of condition monitoring and faultdiagnosis of mechanical equipment, Its running status from nomal to fault stateusually passes through a series of different degraded state,and how to correctlyidentify the current state of equipment,further predicte equipment developmenttrend,provide the basis for maintence of judgment is an urgent need to solve theproblem.Based on the above problems,this paper made the following research:(1) Preprocessing of non-stationary signalIn the mechanical equipment condition monitoring cases,due to thevibration signal acquisition was easier than the other signals,and more sensitiveto the fault.It could provide a wealth of information to equipment operation,sothe vibration signals were regarded as feature of equipment degradated staterecognition.Vibration signal was a typical non-stationary signals,and due tooutside interference,its preprocessing was the key to the later research.Thisarticle introduced the spectral minus denoising and wavelet packet energythreshold denoising method,then discussed the selection of wavelet basis toanalyse and simulate,at last,the applicated environment of two denoising methodwere given through the experiment.Wavelet packet energy threshold denoisingwas suitable for low input SNR of signal,spectral minus denoising was suitablefor high SNR signal,so the two methods could be combined with the vibrationsignal denoising processing.(2) EMD energy entropy feature extractionTraditional Fourier transform couldn’t balance the landscape of the signalin the time-frequency domain and the localization features,and wavelet analysis was not self-adaptive,aiming at the shortcoming of both,this paper presentedempirical mode decomposition based (EMD) method for feature extraction;information entropy was a description of the degree of uncertainty andcomplexity of equipment condition,when sourse contained information unstableand complicated,the entropy value was larger. This paper used EMDdecomposing denoising signals into a set of intrinsic mode function (IMF)components,and then extracted and calculated the IMF energy and energyentropy, as a characteristic parameters describing degradated state.(3) Degraded state identification and fault prediction based on HMMFor hidden Markov model (HMM) algorithm parameter setting, and thetraining algorithm falling into local optimal problems, this paper further studedthe improved algorithm of HMM;for the defect of a single fault predictionmethod,this paper combined HMM with exponential smoothing methods tostudy,which could integrate the both advantages;Finally,it used hydrauliccomponents as the research object to certify the above methods,and comparedwith BPNN,SVM in recognition effect.Numerical results showed that themethod had a good robustness,high distinguish rate and high sensitivity tofailure prediction.