Acoust Aco ustic ic emission emission-ba based sed dia diagn - PowerPoint PPT Presentation

Acoust Aco ustic ic emission emission-ba based sed dia diagn gnost ostics and ics and pr prog ogno nostics stics of of slo slow w rot otating ting bea bearing rings s using using Bayesian Bay esian tec techn hnique iques Depart rtmen ment of M Mechani hanical al and Aeronautic nautical al Engi gine neeri ring, ng, Unive versit ity y of Pretoria ria, South h Africa ca 1

 Developm velopmen ent of diagnos gnosti tics cs of slow ow rotat atin ing g bear arin ings gs which hich are robus bust under der chan hangi ging g operati erating g cond onditio itions  Developm velopmen ent of a systemat ematic ic algo gorit ithm hm capabl pable e of selecti lecting the e most st characteristic haracteristic features tures for CM of mach achine nery  Explo loratio ration and d developme velopment nt of an opti tima mal l prognos gnostic ic model odel for the predic ictio ion of RUL UL of slow ow rotatin ating g beari rings ngs 2

 Diagnostic nostics s of slow w rotat ating ing bearings ngs based on a developed ed novel el degradation tion assess ssment ment index (DAI)  Prognos gnostic ics usin ing g va vario ious us approac roaches es  Prognos gnostic ics based ed on n an integr egrat ated ed Gaus aussia sian proc oces ess s regress essio ion model odel 3

delete Schemat matics cs for develop eloping ng DAI for diagno nost stics cs of slow rotati tating ng beari arings ngs Key: AE – Acoustic emission; PKPCA – Polynomial kernel principal component analysis; NLL – Negative log likelihoods; GMM – Gaussian mixture models; EWMA – Exponentially weighted mean average; DAI – Degradation assessment index 4

5

NLL and DAI AI for the whole ole lifes ifespan n of Beari ring ng 1 NLL and DAI AI for the whole ole lifes ifespan n of Beari ring ng 2 NLL and DAI AI for the whole ole lifes ifespan n of Beari ring ng 3 6

7

Fg  Conditi ition on-mon onito itori ring ng indexes for Bearing ring 1 (a) DAI (b) PKPC PCA-GMM GMM (c) GMM- EWMA (d) PKPC PCA-EWMA EWMA-T 2 and (e) PKPC PCA-EWMA EWMA-SP SPE 8

Conditi ition on-mon onito itori ring ng indexes for Bearing ring 2 (a) DAI (b) PKPC PCA-GMM GMM (c) GMM- EWMA (d) PKPC PCA-EWMA EWMA-T 2 and (e) PKPC PCA-EWMA EWMA-SP SPE 9

Conditio ion-mo monit itorin ing indexes s for Bearin ing 3 ( (a) DAI (b) PKPCA CA-GM GMM (c) GMM-EWMA MA (d) PKPCA-EWMA-T 2 and (e) PKPCA-EWMA-SPE SPE 10

11

Framew ework k for DAI integrated approach ch to b bearin ing prognostic ics Key DAI- degradation assessment index MLP- multi-layer perceptron RBF- Radial basis function BLR- Bayesian linear regression GMR- Gaussian mixture regression GPR- Gaussain process regression RUL- Remaining useful life 12

where is the actual value of the DAI for the ith observation which is in this case the time point, is the predicted value of DAI, n is the number of observations 13

RMSE and MAPE for r MLP, RBF, BLR, GMR and GPR models s for r Bearin aring 1 based d on the depe pende ndent sam sample ples RMSE and MAPE for r MLP, RBF, BLR, GMR and GPR models s for r Bearin aring 2 based d on the depe pende ndent sam sample ples RMSE and MAPE for r MLP, RBF, BLR, GMR and GPR models s for r Bearin aring 2 based d on the depe pende ndent sam sample ples 14

Predi edict ction for the whole bear aring ng life fe for Bearing ng 1 using diff fferent erent metho hodologies es based sed on depe ependen ndent samp mples es Predi edict ction for the whole bear aring ng life fe for Bearing ng 2 using diff fferent erent metho hodologies es based sed on depe ependen ndent samp mples es Predi edict ction for the whole bear aring ng life fe for Bearing ng 3 3 using diff fferent erent metho hodologies es based sed on depe ependen ndent samp mples es  15

RMSE and MAPE for r MLP, RBF, BLR, GMR and GPR models s for r Bearin aring 1 based d on the indepe pende ndent sample les RMSE and MAPE for r MLP, RBF, BLR, GMR and GPR models s for r Bearin aring 2 based d on the indepe pende ndent sample les RMSE and MAPE for r MLP, RBF, BLR, GMR and GPR models s for r Bearin aring 3 based d on the indepe pende ndent sample les 16

Predict dictio ion n for r the whole bearin aring life fe using Bearin aring 2 and 3 as trainin ning set and Bearing ring 1 as test set based on diff ffere rent nt methodo dolo logie ies s and indepe pende ndent sample ples Predict dictio ion n for r the whole bearin aring life fe using Bearin aring 1 and 3 as trainin ning set and Bearing ring 2 as test set based on diff ffere rent nt methodo dolo logie ies s and indepe pende ndent sample ples Predict dictio ion n for r the whole bearin aring life fe using Bearin aring 1 and 2 as trainin ning set and Bearing ring 3 as test set based on diffe fferent rent methodo dolo logies ies and indepe ependen ndent samples ples 17

Rankin king g of mode dels ls base sed d on depend ndent nt and indepe pend ndent nt samp mples for Bearing ring 1 Rankin king g of mode dels ls base sed d on depend ndent nt and indepe pend ndent nt samp mples for Bearing ring 2 Rankin king g of mode dels ls base sed d on depend ndent nt and indepe pend ndent nt samp mples for Bearing ring 3 Table le 3.3: 3:  18

KJ   Framewo work rk for r integrat rated d GPR modelli lling and predict diction of remaini ining ng usefu ful l life 19

KJK Averag age RMSE and MAPE for r GPR with zero, , constan ant t and linear ar mean function tions for Bearin ing g 1 Averag age RMSE and MAPE for r GPR with zero, , constan ant t and linear ar mean function tions for Bearin ing g 2 Averag age RMSE and MAPE for r GPR with zero, , constan ant t and linear ar mean function tions for Bearin ing g 3 20

Ranking ng RMSE and MAPE from GPR with constant nt mean and 9 covari arianc ance funct ctio ions s for Beari aring ng 1 Rankin ing RMSE and MAPE E from GPR with constan ant mean and 9 covarian ariance ce functions ns for r Bearin aring 2 Rank nking ing RMSE and MAPE from GPR with constant nt mean and 9 covari arianc ance funct ctio ions s for Beari aring ng 3 21

Ranking ng RMSE and MAPE from GPR with linear mean and 9 c covarian iance ce function ons s for Bearin ing 1 Rankin ing RMSE and MAPE from GPR with linear mean and 9 c covaria iance ce functio ions ns for Bearing 2 Ranking ng RMSE and MAPE from GPR with linear mean and 9 c covarian iance ce function ons s for Bearin ing 3 22

23

Affin ine GPR predic iction tion of RUL with 95% confide idence interval rval and the actu tual al RUL for r Bearin ing g 1 based d on depende dent sample les Affin ine GPR predic iction tion of RUL with 95% confide idence interval rval and the actu tual al RUL for r Bearin ing g 2 based d on depende dent sample les Affin ine GPR predic iction tion of RUL with 95% confide idence interva val l and the actu tual al RUL for r Bearin ing g 3 3 based d on depende dent sample ples 24

25

Affi fine e GPR pred edict ction n of RUL with h 95% confi nfiden ence ce interv erval al and d the actual al RUL for Bear aring 1 b based sed on indep depen endent dent sampl mples es Affi fine ne GPR predi edict ction of f RUL L with h 95% conf nfidenc dence e inter erval al and the actual al RUL for Bearing ng 2 based sed on indepe ndepend nden ent sampl mples es Affi fine e GPR pred edict ction n of RUL with h 95% confi nfiden ence ce interv erval al and d the actual al RUL for Bear aring 3 b based sed on indep depen endent dent sampl mples es 26

 Developm velopmen ent of a nov ovel el DAI I for diagno agnostics stics  Prognos gnostic ics usin ing g a novel el approa roach ch by integ tegrat atin ing g a newly ewly developed veloped DAI with th several veral models odels  Prognos gnostic ics based ed on n an integr egrat ated ed GPR R model odel 27

 The novel l DAI w was applied to a a sp speci cifi fic c bearin ing. g. Hence nce, it is is expec pecte ted that furthe her studies ies would be carried ied out on other r bearing ng types  The DAI and optimised GPR RUL prognos ostics ics need to be furthe her applied in seve veral al applicati tions ons in engi gine neerin ing  The e st standar ndard GM GMR, BLR, RBF and MLP models could also so be modified and optimised for furthe her r testi ting ng of t their effec fecti tivene veness in R RUL predicti ction on  This study y is t the first investi tiga gati tive ve step p of a a si single le applicati tion on of t the metho hod; its s effect fectivene iveness has s to be proved ed with h furthe her r investiga estigations ions 28

 The Almighty God  Prof PS Heyns  The Centre for Asset Integrity and Management (C-AIM)  All staff & colleagues of Mechanical and Aeronautical Engineering department & its affiliates  My wife, children, family & friends 29

Thank you for your attention 30