On-Lne Fault Detecton n Wnd Turbne Transmsson System usng Adaptve Flter and Robust Statstcal Features Ruoyu L Remote Dagnostcs Center SKF USA Inc. 3443 N. Sam Houston Pkwy., Houston TX 77086 Emal: ruoyu.l@skf.com
Outlne Introducton Theoretcal Bass Data Collecton Analyss Results Concluson
Introducton Wnd Power s the world s fastest growng renewable energy source. Reducng the cost of generatng the wnd energy becomes a crtcal ssue. 70% Mechancal Related Faults [] Rbrant, J., and Bertlng, L.M., "Survey of Falures n Wnd Power Systems Wth Focus on Swedsh Wnd Power Plants Durng 997 2005," Energy Converson, IEEE Transactons on, vol.22, no., pp. 67,73, 2007 30% 2
Introducton Onlne condton montorng of the wnd turbne mechancal transmsson system provdes the followng benefts: Greatly reduce the mantenance cost, Avod catastrophc falure, Improve the relablty of the whole system.
Introducton Challenges: Uncertan knematcs nformaton Speed and loadng varatons Low fault sgnal-to-nose rato Large volume of sgnal wth dfferent types
Introducton Uncertan knematcs nformaton Rght Knematcs Informaton: Bearng Issue Wrong Knematcs Informaton: Healthy
Introducton Speed and loadng varatons Operatng Condton Operatng Condton 2
Introducton Low fault sgnal-to-nose rato Envelope Spectrum Sgnal Processng Envelope Spectrum
Introducton Large volume of sgnal wth dfferent types When gatng condton met the requrements, a wnd farm wth sze of 000 wnd turbnes, 72000 to 230400 etracted trend values daly 24000 to 96000 tme waveforms and spectrums daly
Introducton In real applcatons, effectve fault detecton algorthms are an essental part of the condton montorng system, especally for the onlne contnuous montorng systems, lke wnd turbne condton montorng systems.
Theoretcal Bass Schema of the Proposed Methodology Sgnals Sgnal Processng Feature Etracton Feature Compresson Decson Makng Self-adaptve nose cancellaton Statstcal Features Lnear Dscrmnate Analyss Bayesan Analyss
Theoretcal Bass Sgnal Processng Smulaton Eample Real Vbraton Sgnals y y 2 sn( w* p * t) A t * e N* t *sn( w * p * t) Vbraton sgnal wth bearng nner race defect Ampltude (v) Ampltude (v)
Theoretcal Bass Kurtoss Crest Factor RMS Impulse Factor Skewness ( ) ( ) 2 2 4 n n KT n n n CR n 2 ma n rms n 2 n IF n ma ( ) ( ) 3 2 3 n n KT n n Etracted Features One dmensonal feature Lnear Dscrmnant 2 2 2 2 2 ~ ~ ~ ~ ) ( s s w J + µ µ [ ] y n y y y,...,, 2 ( ) 2 µ µ S W w s w y T [ ] s n s s s,...,, 2
Data Collecton SKF IM-W A robust measurement unt desgned for nstallaton n wnd farms on and off-shore. Steen analogue nputs and two dgtal nputs. Smultaneous measurement of all channels. Mult-parameter gatng. Man Bearng Generator Planetary Stage Intermedate Speed Stage Hgh Speed Stage
Data Collecton Recently, a number of SKF IM-W, WnCon unts were nstalled on a fleet of wnd turbnes to perform the onlne condton montorng. Vbraton sgnal of the healthy gearbo Sgnal of the gearbo wth bearng nner race defect
Data Collecton When gatng condtons meet the settngs: Key features value: 0 Mnutes Tme Waveform: 8 hours Spectrum: 8 hours Samplng frequency: 5.2 khz Samplng tme length:.6 seconds
Analyss Results Vbraton sgnal of the healthy wnd turbne Vbraton sgnal of the damaged wnd turbne Bearng Inner Race Defect Frequency
Analyss Results Table The calculated values of the processed sgnal of healthy gearbo and the gearbo wth defect Healthy Gearbo Gearbo wth Defect Mean STD Mean STD Kurtoss 2.943 0.2080 Kurtoss 6.7379.837 Crest Factor 8.3849.7355 Crest Factor 2.8863.622 RMS 2.0668.0749 RMS.6640 0.6494 Impulse Factor 0.50 2.233 Impulse Factor 7.6538 2.7323 Skewness - 0.08 0.0892 Skewness - 0.0545 0.636 Feature Compresson Fgure The hstogram of the transformed feature of the healthy machne and the machne wth damaged components
Analyss Results 0.9 boundary: The system s deployed to a newly created wnd farm The least fault senstve boundary 98.750% fault detecton accuracy
Analyss Results 0.5 boundary: The moderate fault senstve boundary 99.375% fault detecton accuracy
Analyss Results 0. boundary: The most fault senstve boundary 96.875% fault detecton accuracy
Analyss Results Fgure. The transformed feature and the decson boundares wth dfferent pror statstcs (blue lne: P(Healthy)0., green lne: P(Healthy)0.5,pnk lne: P(Healthy)0.9)
Analyss Results Table The confuson matr of the dfferent boundares Calculated Healthy Calculated Fault 0. Boundary 0.5 Boundary 0.9 Boundary True Healthy 75 5 True Fault 0 80 True Healthy 80 0 True Fault 79 True Healthy 80 0 True Fault 2 78
Concluson An effectve automatc fault detecton methodology has been developed. The methodology conssts: Adaptve flterng technque to mprove the fault SNR LDA to reduce the features dmensons. Wnd turbne vbraton sgnals obtaned n real operaton were used to demonstrate the effectveness of the presented methodology.
Concluson In future work, t wll be nterestng to study the effectveness of the developed method on a varety of rotatng equpments n dfferent ndustres and applcatons, such as rangng arms n the mnng ndustres, pumps n the paper mll, and motors n the food ndustres, and so on.
Questons?