Vibration Analysis PDF
Vibration Analysis PDF
Vibration Analysis PDF
Technical Report
NREL/TP-5000-54530
July 2012
This report was prepared as an account of work sponsored by an agency of the United States government.
Neither the United States government nor any agency thereof, nor any of their employees, makes any warranty,
express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of
any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately
owned rights. Reference herein to any specific commercial product, process, or service by trade name,
trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation,
or favoring by the United States government or any agency thereof. The views and opinions of authors
expressed herein do not necessarily state or reflect those of the United States government or any agency thereof.
Cover Photos: (left to right) PIX 16416, PIX 17423, PIX 16560, PIX 17613, PIX 17436, PIX 17721
Printed on paper containing at least 50% wastepaper, including 10% post consumer waste.
Acknowledgements
The National Renewable Energy Laboratorys (NREL) contributions to this report were funded
by the Wind and Water Power Program, Office of Energy Efficiency and Renewable Energy of
the U.S. Department of Energy, under contract No. DE-AC02-05CH11231. The authors are
solely responsible for any omission or errors contained herein. NREL wishes to acknowledge
and thank the Office of Energy Efficiency and Renewable Energy and its staff who have
supported this work from its inception. Specifically, NREL would like to thank Mark Higgins
and Michael Derby for their support and guidance.
NREL deeply appreciates the voluntarily support from all sixteen partners of the Wind Turbine
Gearbox Condition Monitoring Round Robin project. One project partner estimates that the
value of the voluntary support from all partners would be worth $2 to $3 million.
i
Nomenclature
Acronym Definition
A/D analog to digital
AM amplitude modulation
BPFI ball-passing frequency inner race
BPFO ball-passing frequency outer race
BSF ball-spinning frequency
CF crest factor
CI condition indicator
CM condition monitoring
COE cost of energy
CRB cylindrical roller bearing
DAS data acquisition system
DOD Department of Defense
DOE Department Of Energy
DSTO Defence Science and Technology Organisation (Australia)
DTF Dynamometer Test Facility
EO energy operator
fcCRB full complement cylindrical roller bearing
FFT Fast Fourier Transform
FTF fundamental train frequency
FM frequency modulation
GE General Electric
GL Germanischer Lloyd
GRC Gearbox Reliability Collaborative
GMF gear meshing frequency
HS High speed
HSGM high-speed gear meshing stage
HSIS high-speed intermediate shaft
HSS high-speed shaft
ISGM intermediate speed gear meshing stage
IMS intermediate-speed shaft
IMS Intelligent Maintenance Systems
INT intermediate
JTFA joint time frequency analysis
ii
Acronym Definition
LS low speed
LSIS low-speed intermediate shaft
LSS low-speed shaft
MS main shaft
NASA National Aeronautics and Space Administration
NREL National Renewable Energy Laboratory
NWTC National Wind Technology Center
O&M operation and maintenance
PGSF spin frequency of the planetary gear
PLC planet carrier
PLTGM planetary gear meshing stage
RMS root mean square
RPM revolutions per minute
SER sideband energy ratio
SF severity factor
SH shaft
SK spectral kurtosis
SO shaft order
TRB tapered roller bearing
TSA time synchronous averaging
iii
Executive Summary
Utility-scale wind turbines have historically experienced premature component failures, which
subsequently increase the cost of energy. The majority of these failures are caused by faults in
the drivetrain, led by the main gearbox. To understand the possible causes for gearbox failures
and recommend practices for improvement, the National Wind Technology Center (NWTC), at
the National Renewable Energy Laboratory (NREL), started a project called the Gearbox
Reliability Collaborative (GRC). Condition Monitoring (CM) is one research area under the
GRC. It is a method to assess a systems health, which enables proactive maintenance planning,
reduces downtime and operations and maintenance costs, and, to some extent, increases safety.
To understand the dynamic responses of wind turbine gearboxes under different loading
conditions, the GRC tested two identical gearboxes. One was tested on the NWTCs 2.5 MW
dynamometer and the other was first tested in the dynamometer, and then field tested in a turbine
in a nearby wind plant. In the field, the test gearbox experienced two oil loss events that resulted
in damage to its internal bearings and gears. Since the damage was not catastrophic, the test
gearbox was removed from the field and retested in the NWTCs dynamometer before it was
disassembled. During the dynamometer retest, various condition monitoring systems, e.g.,
vibration and oil debris, collected data along with testing condition information. The vibration-
based condition monitoring system and the test condition data enabled NREL to launch a Wind
Turbine Gearbox Condition Monitoring Round Robin project, as described in this report. The
main objective of this project is to evaluate different vibration analysis algorithms used in wind
turbine CM and determine whether typical practices are effective. With the involvement of both
academic researchers and industrial partners, the Round Robin provides cutting edge research
results to industry stakeholders.
Under this project, the collected vibration and testing condition data, along with the test gearbox
configuration information, were shared with partners who signed memoranda of understanding
documents with NREL. The partners were given a time window of two months to analyze the
shared data using whichever algorithms they had or could develop. Partners did not have prior
knowledge of the extent or location of the damage in the test gearbox. After their diagnostic
results were submitted to NREL, the actual damage information on the test gearbox was
disclosed to them so they could further fine tune their results. The project had sixteen partners,
including seven universities and nine from the private sector. The main body of this report
discusses detailed analysis algorithms and diagnostic results from eight of the sixteen partners.
In Chapter 1, the background and objectives of this Round Robin project are presented,
along with the summary of blind study stage diagnostics results from all sixteen project
partners.
In Chapter 2, the test gearbox, dynamometer test facility, one customized vibration data
acquisition system, test conditions, and actual damage found on the test gearbox through
its disassembly are presented.
iv
In Chapter 3, to overcome the inaccuracy incurred by speed variations, a synchronous
sampling technique is introduced. To accommodate high gear ratios in wind turbine
gearboxes, a digital domain synchronous re-sampling technique is presented.
In Chapter 4, an evaluation is conducted of analysis algorithms originally developed for
Department of Defense applications, including the results of these algorithms when
applied to the GRC wind turbine dataset. The algorithms consist of sensor validation,
bearing fault detection/isolation, and gear fault detection/isolation modules. A joint time
frequency analysis is also discussed.
In Chapter 5, a validation is presented of analysis algorithms that are used in the
aerospace community for the wind industry. The focus is on two methodologies:
synchronous analysis of shaft/gear components and non-synchronous analysis of
bearings.
In Chapter 6, various vibration signal processing and feature extraction algorithms are
evaluated. It details the evaluated methods including frequency domain, cepstrum,
bearing envelope analysis, spectral kurtosis filtering, time synchronous averaging, and a
planet separation method.
In Chapter 7, sideband pattern analysis is performed on all gears for gear fault diagnosis.
Data from torque measurements have also been analyzed to facilitate annulus gear
diagnosis. For the bearing diagnosis, a multi-scale enveloping spectra technique is
investigated.
In Chapter 8, analysis of jerk data derived from vibration acceleration data of the test
wind turbine gearbox are discussed. For component failure identification, the correlation
coefficient analysis and clustering analysis are applied to identify the failure stage of the
gearbox in the time domain.
In Chapter 9, the algorithms for bearing diagnostics are presented. They consist of several
different stages to separate and enhance the bearing signals, and then envelope analysis is
applied. For parallel stage gear diagnostics, classic synchronously-averaged signatures
are studied and comparisons are made of spectra and cepstra from the healthy and faulty
conditions. For individual planet and sun gear diagnostics, the premium current method is
investigated.
In Chapter 10, a two stage fault detection framework, with analytical and graphical
analysis for wind turbine gearbox CM, is proposed. The analytical diagnostics and
graphical analysis are performed for fault detection and defect severity level evaluation of
different damage modes based on sideband and kurtosis analyses.
In Chapter 11, some recommended practices for data acquisition and data analysis are
described for use in conducting vibration-based wind turbine drivetrain condition
monitoring.
It is worth noting that the synopses detailed for Chapters 3-10 were based on the analysis
algorithms of the project's partners. Detailed diagnostic results obtained by each partner are
listed in this report.
v
The study has demonstrated that the wind industry can improve vibration analysis algorithms for
drivetrain condition monitoring. Both the presented algorithms and the recommended practices
can be considered by CM equipment suppliers in their future product releases for the benefit of
the wind industry.
vi
Table of Contents
Acknowledgements ..................................................................................................................................... i
Nomenclature .............................................................................................................................................. ii
Executive Summary ................................................................................................................................... iv
1 Introduction ............................................................................................................................................ 1
2 Tests and Actual Gearbox Damage ..................................................................................................... 3
2.1 Test Article ............................................................................................................................... 3
2.2 Dynamometer Test Facility ...................................................................................................... 6
2.3 One Customized Vibration Data Acquisition System ............................................................... 7
2.4 Test Conditions ........................................................................................................................ 9
2.5 Actual Gearbox Damage ........................................................................................................ 10
3 Analysis Algorithms and Diagnostics Results from General Electric ........................................... 12
3.1 Fundamentals......................................................................................................................... 12
3.2 Round Robin Analysis Results ............................................................................................... 19
3.3 Discussions ............................................................................................................................ 32
4 Combining Novel Approaches with Proven Algorithms for Robust Wind Turbine Gearbox Fault
Detection .................................................................................................................................... 34
4.1 Introduction............................................................................................................................. 34
4.2 Algorithm Overview ................................................................................................................ 34
4.3 Results Summary ................................................................................................................... 38
4.4 Lessons Learned and Conclusions ........................................................................................ 48
5 Analysis Algorithms and Diagnostics Results from NRG Systems ............................................... 49
5.1 Introduction............................................................................................................................. 49
5.2 Analysis Algorithms ................................................................................................................ 49
5.3 Analysis Results ..................................................................................................................... 52
5.4 Discussion .............................................................................................................................. 56
6 Review and Application of Methods and Algorithms in Wind Turbine Gearbox Fault Detection 58
6.1 Introduction............................................................................................................................. 58
6.2 Signal Processing and Feature Extraction Methods .............................................................. 58
6.3 Summary of Results ............................................................................................................... 80
6.4 Conclusions and Future Work ................................................................................................ 82
7 Defect Diagnosis in Wind Turbine Gearbox basd on Sideband Energy and Enveloping Spectral
Analysis ...................................................................................................................................... 83
7.1 Introduction............................................................................................................................. 83
7.2 Algorithms .............................................................................................................................. 83
7.3 Results ................................................................................................................................... 84
7.4 Lessons Learned .................................................................................................................... 90
8 Fault Analysis of a Wind Turbine Gearbox: A Data Driven Approach ........................................... 91
8.1 Introduction............................................................................................................................. 91
8.2 Methodologies ........................................................................................................................ 91
8.3 Results ................................................................................................................................... 93
8.4 Conclusion and Discussion .................................................................................................... 97
9 Techniques for Separation and Enhancement of Various Components in the Analysis of Wind
Turbine Vibration Signals ......................................................................................................... 98
9.1 Introduction............................................................................................................................. 98
9.2 Algorithms .............................................................................................................................. 99
9.3 Results ................................................................................................................................. 105
9.4 Discussion, Conclusions, Lessons Learned ........................................................................ 122
10 A Two Stage Fault Detection Framework for Wind Turbine Gearbox Condition Monitoring ..... 124
10.1 Introduction........................................................................................................................... 124
10.2 Vibration Based Condition Monitoring Framework............................................................... 124
10.3 Analytical Diagnostics .......................................................................................................... 126
10.4 Graphical Diagnostics .......................................................................................................... 129
10.5 Results ................................................................................................................................. 130
vii
11 Recommended Practices .................................................................................................................. 132
11.1 Data Acquisition ................................................................................................................... 132
11.2 Data Analysis ....................................................................................................................... 133
Appendix A Project Partners .............................................................................................................. 135
References ............................................................................................................................................... 136
List of Figures
Figure 1.1. Blind study stage diagnostics results comparison ........................................................ 2
Figure 2.1. GRC test turbine ........................................................................................................... 3
Figure 2.2. GRC gearbox internal components view ...................................................................... 4
Figure 2.3. GRC gearbox internal nomenclature and abbreviations ............................................... 4
Figure 2.4. GRC gearbox layout and bearing nomenclature .......................................................... 5
Figure 2.5. Diagram of NREL 2.5 MW dynamometer test facility ................................................ 7
Figure 2.6. NREL dynamometer test stand with the test article installed....................................... 7
Figure 2.7. Vibration data acquisition system sensor locations ...................................................... 8
Figure 2.8. Physical sensor installation........................................................................................... 9
Figure 2.9. Test gearbox high speed stage gear damage............................................................... 11
Figure 3.1. Synchronous sampling analog approach ................................................................. 13
Figure 3.2. Synchronous sampling digital approach ................................................................... 14
Figure 3.3. Time synchronous averaging ...................................................................................... 14
Figure 3.4. Synthesized tachometer generation from speed function ........................................... 16
Figure 3.5. Analog devices-based approach ................................................................................. 18
Figure 3.6. Digital processing-based approach ............................................................................. 19
Figure 3.7. Gearbox power flow ................................................................................................... 19
Figure 3.8. HSGM (22) modulated by HSS (1) ............................................................................ 23
Figure 3.9. HSGM X2 (44) modulated by HSS (1) ...................................................................... 24
Figure 3.10. HSGM X3 (66) modulated by HSS (1) .................................................................... 24
Figure 3.11. HSGM (22) modulated by HSIS (0.25) .................................................................... 25
Figure 3.12. HSGM X2 (44) modulated by HSIS (0.25) .............................................................. 25
Figure 3.13. HSGM X2 (66) modulated by HSIS (0.25) .............................................................. 26
Figure 3.14. ISGM and higher order harmonics ........................................................................... 26
Figure 3.15. Planetary gear stage meshing order and harmonics from AN4 ................................ 27
Figure 3.16. PLTGM X3 modulated by planet passing order (0.037) .......................................... 28
Figure 3.17. Planetary gear stage meshing order and harmonics from AN3 ................................ 28
Figure 3.18. PLTGM X3 and X4 modulated by planet passing order (0.037) ............................. 29
Figure 3.19. Sensor AN3 acceleration enveloping order spectrum .............................................. 29
Figure 3.20. Modulation by LSIS in PLTGM is very small ......................................................... 30
Figure 3.21. Envelope spectrum of AN7 ...................................................................................... 31
Figure 3.22. Zoomed envelope spectrum of AN7......................................................................... 31
Figure 3.23. Possible bearing D BPFI .......................................................................................... 32
Figure 4.1. False alarm caused by faulty sensor ........................................................................... 35
Figure 4.2. ImpactEnergy overview.............................................................................................. 36
Figure 4.3. GearMod overview ..................................................................................................... 37
Figure 4.4. Example JTFA approach (short time Fourier transform) ........................................... 38
Figure 4.5. High speed gear fault evidence, blind results (AN6) ................................................. 40
viii
Figure 4.6. High speed gear fault evidence, blind results (AN7) ................................................. 40
Figure 4.7. Intermediate speed downwind bearing fault evidence, blind results (AN6,
Data 2b)............................................................................................................................. 41
Figure 4.8. Intermediate speed downwind bearing fault evidence, blind results.......................... 42
Figure 4.9. Sun pinion gear fault evidence, blind results.............................................................. 43
Figure 4.10. JTFA speed gear fault evidence, blind results .......................................................... 44
Figure 4.11. Intermediate speed upwind bearing damage evidence ............................................. 45
Figure 4.12. Intermediate speed upwind bearing fault evidence, post-inspection results ............ 46
Figure 4.13. High speed downwind bearing fault evidence, revisited (Data 2a) .......................... 47
Figure 4.14. High speed downwind bearing fault evidence, revisited (Data 2c) .......................... 47
Figure 5.1. Generation of the TSA and selected CIs .................................................................... 51
Figure 5.2. Process for generating gear CIs .................................................................................. 52
Figure 5.3. Synthetic tachometer .................................................................................................. 53
Figure 5.4a. HSS TSA/spectrum................................................................................................... 53
Figure 5.4b. HSS gear analysis ..................................................................................................... 53
Figure 5.5a. TSA intermediate shaft ............................................................................................. 54
Figure 5.5b. Intermediate speed pinion, where the units for the Energy Operator, Narrowband
and Amplitude Modulation analysis are in Gs, and the Frequency Modulation analysis
is in radians. ...................................................................................................................... 54
Figure 5.6a. Sun gear .................................................................................................................... 55
Figure 5.6b. Planet gears............................................................................................................... 55
Figure 5.7. Ring gear .................................................................................................................... 55
Figure 5.8a. High speed shaft, downwind side ............................................................................. 56
Figure 5.8b. Intermediate speed shaft downwind side .................................................................. 56
Figure 5.9. Low speed shaft downwind side................................................................................. 56
Figure 6.1. Vibration spectrum - Case C: top plot - AN7 baseline; bottom plot - AN7 degraded
gearbox.............................................................................................................................. 61
Figure 6.2. Sideband ratio gear features - Case C: (a) Low speed shaft pinion; (b) High speed
shaft gear; (c) High speed shaft pinion ............................................................................. 63
Figure 6.3. Real cepstrum - Case C: top plot - AN7 baseline; bottom plot - AN7 degraded
gearbox.............................................................................................................................. 64
Figure 6.4. Cepstrum peak features from Case C: blue baseline; red - degraded gearbox ........ 64
Figure 6.5. Cepstrum health Indicator for Case C calculated for high speed shaft gear and
pinion ................................................................................................................................ 65
Figure 6.6. Bearing envelope analysis flow chart ......................................................................... 66
Figure 6.7. Envelope spectrum - Case C: (a) AN6 - peaks at BPFI for ISS upwind bearing and
HSS downwind bearing; (b) AN7 - BPFI peak for HSS downwind bearing.................... 67
Figure 6.8. Envelope spectrum accelerometer AN10 - Case C: (a) band-pass filter from 9500
Hz - 10,500 Hz, peaks at BPFO and 2X BPFO for ISS downwind bearing; (b) band pass
filter from 4000 Hz - 6000 Hz, peak at BPFO for planet carrier upwind bearing and also
peak at 2X BPFO for ISS downwind bearing ................................................................... 67
Figure 6.9. (a) Wiener filter based on spectral kurtosis; (b) raw and filtered AN4 accelerometer
signal Case A ................................................................................................................. 69
Figure 6.10. Filtered AN4 signal showing the periodic repetition based on 2 revolutions of the
carrier Case A ................................................................................................................ 69
Figure 6.11. Kurtosis of filtered signal - shown for all 3 cases .................................................... 70
ix
Figure 6.12. TSA signal and residual signal from accelerometer AN7 - Case C: top plot - TSA
signal for high speed shaft pinion; bottom plot - residual signal for high speed shaft
pinion ................................................................................................................................ 72
Figure 6.13. TSA vibration spectrum for accelerometer AN7 and high speed shaft - Case C ..... 72
Figure 6.14. TSA signal and residual signal from accelerometer AN3 - Case C: top plot - TSA
signal for ring gear; bottom plot - residual signal for ring gear ........................................ 73
Figure 6.15. High speed pinion amplitude and phase modulation signal from accelerometer
AN7 - Case C: top plot - Time Synchronous Average; middle plot - amplitude
modulation signal; bottom plot - phase modulation signal .............................................. 74
Figure 6.16. Ring gear amplitude and phase modulation signal from accelerometer AN3 - Case
C: top plot - TSA; middle plot - amplitude modulation signal; bottom plot - phase
modulation signal .............................................................................................................. 75
Figure 6.17. Flow chart for planet separation algorithm .............................................................. 76
Figure 6.18. Narrow band amplitude modulation signal for determining planet passing
Case C ............................................................................................................................... 77
Figure 6.19. Example Tukey window used for planet separation algorithm - in this study, Nv
was set to 3 to include 3 mesh periods .............................................................................. 77
Figure 6.20. Top - TSA signal for Planet 2; bottom - residual signal for Planet 2 Case C ....... 78
Figure 6.21. TSA vibration spectrum for Planet 2 Case C ........................................................ 78
Figure 6.22. Top - TSA signal for sun gear; bottom - residual signal for sun gear Case C....... 79
Figure 6.23. Planet Gear 2 amplitude and phase modulation signal from accelerometer AN3
Case C: top plot TSA; middle plot - amplitude modulation signal; bottom plot -
phase modulation signal.................................................................................................... 80
Figure 7.1. Locations of defective components in the gearbox assembly .................................... 84
Figure 7.2. Comparison analysis between test data and reference data for HS_Pinion and
INT_Pinion ....................................................................................................................... 85
Figure 7.3. Comparison analysis between test data and reference data for Annulus_Gear and
Sun_Gear........................................................................................................................... 86
Figure 7.4. Time series of torque data under 1200 rpm ................................................................ 87
Figure 7.5. The envelope spectrum of torque data under 1200 rpm ............................................. 88
Figure 7.6. Wavelet enveloping spectrum of sensor AN3 at 1,800 rpm ....................................... 88
Figure 7.7. Wavelet enveloping spectrum of sensor AN6 at 1,800 rpm ....................................... 89
Figure 8.1. Run chart of maximum rate of speed ......................................................................... 94
Figure 8.2. Bar char of R .............................................................................................................. 94
Figure 8.3. RMS across 12 sensors - Case 2b ............................................................................... 96
Figure 8.4. Crest factor across 12 sensors - Case 2b .................................................................... 97
Figure 8.5. Kurtosis across 12 sensors - Case 2b.......................................................................... 97
Figure 9.1. Signal processing approach pre release of inspection report.................................... 100
Figure 9.2. Reference (speed) signal extraction stages: (a) identifying a separable band; (b)
extracting the band into a new buffer; (c) inversing the transform signal b into the time
domain [60] ..................................................................................................................... 101
Figure 9.3. HSS estimates: Top - data 2a:5; Bottom - data 2c:5 ................................................ 102
Figure 9.4. Synchronously averaged signals from sensor 3, data 2a:5 ....................................... 103
Figure 9.5. Synchronously averaged signals from sensor 3, Data 2c:5 ...................................... 104
Figure 9.6. Schematic diagram of the cepstral method for removing selected families of
harmonics and/or sidebands from time signals [57] ....................................................... 105
x
Figure 9.7. Squared envelope spectrum for data 2_a_10 sensor 7.............................................. 106
Figure 9.8. Squared envelope spectrum for data 2_c_10 sensor 7.............................................. 106
Figure 9.9. Zoom-in around the BPFI. Harmonic cursors for the ISS ........................................ 107
Figure 9.10. Squared envelope spectrum for data 2_a_5 sensor 8 showing the BPFI of bearing
NU2220 ........................................................................................................................... 108
Figure 9.11. Squared envelope spectrum for data 2_a_5 sensor 8 showing the FTF harmonics
of bearing NU2220 ......................................................................................................... 108
Figure 9.12. Squared envelope spectrum of data 2c_5 sensor 6 showing the shaft speed (30.06
Hz), what appears as 2BPFO for bearing 32032X and the BPFI for bearing NU2220 109
Figure 9.13. Power spectrum density comparison of the high speed data through sensor 5 ...... 110
Figure 9.14. Residual of signal 2a_5 sensor 5 ............................................................................ 111
Figure 9.15. Squared envelope spectrum of the residual signal shown in Figure 9.14 .............. 111
Figure 9.16. Spectrum comparison using the data from sensor 3 ............................................... 113
Figure 9.17. Cepstrum comparison using the data from sensor 3 ............................................... 113
Figure 9.18. Spectrum comparison using the data from sensor 5 ............................................... 114
Figure 9.19. Cepstrum comparison using the data from sensor 5 ............................................... 114
Figure 9.20. Spectrum comparison using the data from sensor 6 ............................................... 115
Figure 9.21. Spectrum comparison using the data from sensor 6 ............................................... 115
Figure 9.22. Cepstrum comparison using the data from sensor 6 ............................................... 116
Figure 9.23. Spectrum comparison using the data from sensor 7 ............................................... 117
Figure 9.24. Time records from the averaged ISS signals: (a) original, including four rotations of
the HSS; (b) Residual after removal of the HSS average. .............................................. 118
Figure 9.25. Spectra of signals of Figure 9.24: (a) original including four rotations of the HSS;
(b) residual after removal of the HSS average. ............................................................... 118
Figure 9.26. Whitened synchronously averaged signals corresponding to the periods of all major
gear components in the gearbox, which enhance local faults ......................................... 119
Figure 9.27. Phase shifts for separated sun gear averages .......................................................... 120
Figure 9.28. Sun Gear residual of DSTO average data set 2c (high speed, high load) ........... 120
Figure 9.29. Sun Gear residual of DSTO average data set 2a (low speed, low load) .............. 121
Figure 9.30. Residual signals for the three planet gears ............................................................. 121
Figure 9.31. Average for the annulus gear .................................................................................. 122
Figure 10.1. Vibration based condition monitoring approach .................................................... 125
Figure 10.2. HSS downwind bearing BPFO graphical analysis .............................................. 130
List of Tables
Table 2.1. Gear element dimensions and detail .............................................................................. 5
Table 2.2. Bearing type, number, and location ............................................................................... 6
Table 2.3. Sensor notations and descriptions .................................................................................. 8
Table 2.4. Test conditions ............................................................................................................. 10
Table 2.5. Actual damage on the test gearbox .............................................................................. 10
Table 3.1. Gear damage features................................................................................................... 20
Table 3.2. Bearing damage features.............................................................................................. 21
Table 3.2. Bearing damage features (continued) .......................................................................... 22
Table 4.1. Select gear diagnostic feature definitions .................................................................... 37
Table 4.2. Initial blind results summary ....................................................................................... 39
xi
Table 4.3. Post-inspection results summary ................................................................................. 45
Table 6.1. Summary of evaluated methods advantages and disadvantages ............................... 60
Table 6.2. Frequency domain gear features .................................................................................. 62
Table 6.3. Summary of results for each algorithm with the following notation: L-low
confidence, M-medium confidence, H-high confidence, NA not applicable or
evaluated; black - indicates a method that was evaluated before the failure report,
blue - indicates a method that was evaluated after the failure report .......................... 81
Table 7.1. Sideband energy comparison between new gearbox and gearbox at the end of service
life ............................................................................................................................... 87
Table 7.2. Comparison between the analysis result and the actual damage of a tested gearbox .. 89
Table 8.1. Correlation coefficient analysis of the mean of jerk data: Case 2b ............................. 95
Table 8.2. Correlation coefficient analysis of the mean of jerk data: Case 2c ............................. 95
Table 8.3. Clustering based classification..................................................................................... 96
Table 10.1. Procedure for vibration data preprocessing ............................................................. 125
Table 10.2. Sensor and component relationship ......................................................................... 126
Table 10.3. Procedure for analytical diagnostics ........................................................................ 126
Table 10.4. Sideband-based severity definition .......................................................................... 127
Table 10.5. Severity factor analysis of sensor AN 6 for Case 2a ............................................... 128
Table 10.6. Vibration-based condition monitoring results ......................................................... 130
xii
1 Introduction
Wind energy is currently the fastest growing type of renewable energy resource in the world [1].
However, the industry still experiences premature component failures, which increase operation
and maintenance (O&M) costs, and subsequently, the cost of energy (COE) for wind power. As
turbines increase in size and are installed offshore, these failures will become even more costly.
To make wind power more competitive, there is a need for the industry to improve turbine
reliability and availability.
Given that the gearbox is the most costly drivetrain component to maintain throughout the
expected 20-year design life of a wind turbine, the National Renewable Energy Laboratory
(NREL) organized a consortium called the Gearbox Reliability Collaborative (GRC) to address
the gearbox reliability and availability challenges. The GRC engages key representatives in the
wind turbine gearbox supply chain, including turbine owners, operators, gearbox manufacturers,
bearing manufacturers, lubricant companies, and wind turbine manufacturers. The GRC's goals
are to conduct research that improves gearbox reliability and increases turbine availability. The
GRC engages a multi-track approach, which includes modeling and analysis, dynamometer
testing, field testing, condition monitoring (CM), and developing a gearbox failure database [2].
CM is a method to assess a systems health, which enables proactive maintenance planning,
reduces downtime and operations and maintenance costs, and, to some extent, increases safety
[3]. It will be the main focus of this report.
The GRC uses two identical test gearboxes: one was tested on NRELs 2.5 MW dynamometer;
the other was first tested in the dynamometer, and then field tested in a turbine in a nearby wind
plant. In the field, the test gearbox experienced two oil loss events that resulted in damage to its
internal bearings and gears. Additional field tests of this gearbox were terminated to prevent
further damage to the gearbox. From the CM point of view, however, it provided a unique
opportunity to evaluate different monitoring techniques by retesting the gearbox in NRELs
dynamometer under controlled testing conditions. The gearbox was removed from the field and
retested in the NRELs 2.5 MW dynamometer before it was disassembled. During the
dynamometer retest, various condition monitoring systems data were collected, e.g., vibration
and oil debris, along with testing condition information. The vibration-based condition
monitoring system and the test condition data enabled NREL to launch the Wind Turbine
Gearbox Condition Monitoring Round Robin (Round Robin) project that involves the analysis of
the collected vibration data by several independent research partners and then draws
conclusions from the comparison of their analysis results.
The main objective of the CM Round Robin project was to evaluate different vibration analysis
algorithms used in wind turbine CM and to determine whether typical practices are effective.
Another project objective was to assess the capability of vibration-based CM and to establish a
baseline from which future improvements can be measured. With the involvement of both
academic researchers and industrial partners, the Round Robin provides cutting edge research
results to industry stakeholders.
In the project, the collected vibration and testing condition data, along with the test gearbox
configuration information, were shared with partners who signed memorandum of understanding
1
documents with NREL. The partners were given a time window of two months to analyze the
shared data using whichever algorithms they had or could develop. Partners did not have prior
knowledge of the actual damage within the test gearbox. After their diagnostics results were
submitted to NREL, the actual damage information within the test gearbox was disclosed to
them, so they could further fine tune their results. The project had sixteen partners, including
seven universities and nine from the private sector. The main body of this report discusses
detailed analysis algorithms and diagnostics results from eight of the sixteen partners. (For a list
of partners, see Appendix A.)
The project is unique since the initial diagnostic results from the partners were obtained during a
blind study. Also, the test gearbox did not begin with seeded faults, as have been investigated in
many other studies. Based on the particulars of the actual damage found after the test gearbox
was disassembled [4], all of the partners agreed that seven damage instances could be detected
through vibration analysis. These damage instances were chosen as the reference for the partners'
diagnostics performance evaluation. The evaluation criteria included successful identifications,
false alarms, and missed detections. A comparison of the results, without the partners' names is
illustrated in Figure 1.1. The chart depicts the highest ratio of successful identification as five of
the seven damage instances. Most partners had more missed detections than false alarms. Thus,
there is room for the industry to improve vibration-based diagnostic algorithms. Most of the
Round Robin study partners agreed that this project was a valuable effort.
The next chapter of this report describes the test gearbox configuration, its customized data
acquisition system, test conditions, and actual damage information obtained during the test
gearbox disassembly. The main body of the report contains a detailed presentation of the analysis
algorithms and diagnostic results from eight out of the sixteen research partners. Finally, the
report concludes with some recommended practices for conducting vibration-based wind turbine
drivetrain CM.
2
2 Tests and Actual Gearbox Damage
The test gearbox, dynamometer test facility, one customized vibration data acquisition system,
test conditions, and actual damage found on the test gearbox through its disassembly are
presented in this chapter.
The gearbox under test was one of two units, which were originally taken from the field and
redesigned, rebuilt and instrumented with more than 125 sensors. The gearbox first finished its
run-in in the NREL dynamometer test facility and later was sent to a wind plant located near to
NREL for field tests, where two oil loss events occurred while the turbine was being tested. The
gearbox has an overall ratio of 1:81.491. It is composed of one low speed (LS) planetary stage
and two parallel stages, as shown in the expanded view in Figure 2.2. Nomenclature for the
internal elements is described in Figure 2.3, and the gear dimensions, teeth number, and helix
angles are listed in Table 2.1.
Generator Generator
Brake Shaft
Gearbox
Main Main
Hub Bearing Shaft
Bed Plate
3
Figure 2.2. GRC gearbox internal components view
Pinion
Intermediate Speed
Annulus Shaft
IMS-SH
Gear Pinion
Planet
PLC
Sun
Gear
4
Table 2.1. Gear element dimensions and detail
Gear Element No. of Mate Root Helix Face Ratio
Teeth teeth diameter angle width
(mm) (mm)
Ring gear 99 39 1047 7.5L 230
Planet gear 39 99 372 7.5L 227.5
Sun pinion 21 39 186 7.5R 220 5.71
Several bearing types are employed in the gearbox, according to the loading conditions and
gearbox life requirements. The planet carrier is supported by two full-complement cylindrical
roller bearings (fcCRB) and each planet gear is supported by two identical cylindrical roller
bearings (CRB). Each parallel shaft in the gearbox is supported by a CRB on the upwind side of
the assembly, and by two back-to-back mounted, tapered roller bearings (TRB) on the downwind
side. Table 2.2 provides the location and bearing manufacturer part number of all bearings in the
gearbox. Location and shaft designations are as noted in Figure 2.4. The letter following the
abbreviation indicates the position of the bearing according to the component from upwind (A)
to downwind (B, C). Lubrication oil is another important component in the test gearbox,
although it is not shown in either Table 2.2 or Figure 2.4.
Pinion
Gear Pinion
Planet
PL-A PL-B
LSS-A
LSS-B LSS-C
Figure 2.4. GRC gearbox layout and bearing nomenclature
5
Table 2.2. Bearing type, number, and location
Location
Location Type Provider Part Number
Designation
PLC-A fcCRB INA SL 18 1892 E
Planet carrier
PLC-B fcCRB INA SL 18 1880 E
PL-A CRB FAG NJ 2232 E.M1.C3
Planet
PL-B CRB FAG NJ 2232 E.M1.C3
LS-SH-A fcCRB INA SL 18 1856E
Low Speed Shaft LS-SH-B TRB SKF 32948
LS-SH-C TRB SKF 32948
IMS-SH-A CRB FAG NU 2220 E.M1.C3
Intermediate Speed Shaft IMS-SH-B TRB SKF 32032 X
IMS-SH-C TRB SKF 32032 X
HS-SH-A CRB FAG NU 2220 E.M1.C3
High Speed Shaft HS-SH-B TRB SKF 32222 J2
HS-SH-C TRB SKF 32222 J2
The operating gear mesh and bearing characteristic frequencies can be determined by the project
partners, based on the data shown in Tables 2.1 and 2.2, along with catalogue information from
bearing suppliers.
6
Figure 2.5. Diagram of NREL 2.5 MW dynamometer test facility
Figure 2.6. NREL dynamometer test stand with the test article installed. NREL/PIX #16913.
7
simplicity of implementation, data was collected at 40 kHz per channel using a National
Instruments PXI -4472B high speed DAS.
The accelerometers mounting locations are illustrated in Figure 2.7, with sensor labels and
descriptions given in Table 2.3. The mounting locations were chosen to reflect typical sensor
placement practices seen in commercial wind turbine drivetrain vibration-based condition
monitoring systems. The physical installation of these accelerometers is shown in Figure 2.8.
8
a) AN1, AN2, AN3, and AN4 (From left to right, NREL/PIX #19589, 19590, 19588, 19587)
b) AN5, AN6, AN7, and AN8 (From left to right, NREL/PIX #19591, 19592, 19594, 19593)
c) AN9, AN10, AN11, and AN12 (From left to right, NREL/PIX #19595, 19598, 19597, 19596)
9
Table 2.4. Test conditions
Test Case LSS Speed Nominal HSS Speed Electric Power Duration
(rpm) (rpm) (% of rated) (min)
CM_2a 14.72 1200 25% 10
CM_2b 22.09 1800 25% 10
CM_2c 22.09 1800 50% 10
10
Figure 2.9. Test gearbox high speed stage gear damage. NREL/PIX #19599.
The root cause of the faults was assembly damage and oil starvation resulting from the two oil
loss events in the field test. Among the 12 damaged items listed in Table 2.5, the consensus
reached among the sixteen Round Robin partners was that the first seven should be detectable by
vibration analysis. Damage 12 could potentially be detected by vibration analysis as well, but
most partners considered it to be an operational condition and not damage. Therefore, the first
seven damage instances were used as references for performance evaluations of the partner's
diagnostic results, as discussed in the introduction chapter of this report. The task for the partners
was to diagnose possible internal component damage of the test gearbox based on the shared
vibration, rpm, and torque data. As the project progressed, each of the partners recognized that
some baseline data collected from a healthy test gearbox would be beneficial. Therefore,
vibration spectrum data collected by several accelerometers mounted on the test gearbox, when it
was considered healthy, were later shared with the project partners. The following report sections
were submitted by eight out of the sixteen partners who took part in the CM study.
11
3 Analysis Algorithms and Diagnostics Results from General
Electric
Huageng Luo*, Charles Hatch, Matthew Kalb, Jesse Hanna, Adam Weiss
General Electric Energy
*
Corresponding Author Email: [email protected]
3.1 Fundamentals
Solutions provided by the General Electric (GE) Energy Team are mainly based on order
analysis to accommodate the constant speed variations in a wind farm. For gear damage
detection, the sideband distributions were used to estimate the gear meshing condition and a
sideband energy ratio was used to qualitatively evaluate the gear damage. For early bearing
damage detection, the acceleration enveloping detection technique was used.
In the wind farm, the wind speed is not predictable, thus many wind turbines are operated at
variable speed. As a result, the gearbox operational speed is constantly changing. Due to this
kind of speed variation, a direct application of the conventional Fast Fourier Transform (FFT)
will not result in accurate gearbox condition features, especially those features extracted from
high frequency response, such as acceleration enveloping analysis techniques [7,8]. On the other
hand, to improve the energy extraction efficiency, the wind turbine rotor speed has to be geared
up about two orders of magnitude before being used to drive the generator shaft. Because of the
high gear ratio, a very high-count encoder is needed. For example, in a 1.5 MW wind turbine,
the ratio between the high-speed gear meshing frequency and the rotor frequency can easily be
greater than 1500.
To overcome these difficulties, the GE team utilized a series of signal processing techniques,
such as synchronous sampling, synchronous analysis, digital domain encoder synthesizing,
acceleration enveloping analysis, and sideband energy ratio (SER), in the data processing and
damage feature extractions.
O RPM
f = (1)
60
If the rotating speed is fixed, a regular FFT analysis can have the desired results. However, if the
rotor speed changes within the time window of data acquisition, the variation of the rotor speed
will cause the fundamental order and harmonics in the frequency domain to be smeared into
multiple frequency bins. On the other hand, the frequency of interest may not always be right at
a bin, depending on the resolution of the frequency analysis. The signal energy has to be split
into neighboring bins, in such a case. That is why in rotor dynamics, order analysis is preferred
over the frequency spectrum analysis. In the order domain, the values of the fundamental order
12
and the harmonics remain constant with respect to the shaft speed; the first order is always at the
shaft speed and the second order is always twice the shaft speed, and so on.
To achieve order analysis in rotating machinery applications with variable running speeds,
instead of sampling at equal increments of time, a different sampling technique has to be used.
Sampling is conducted at equal increments of shaft rotation position, thus reducing the effect of
the shaft speed variations. This is called synchronous sampling. The synchronous sampling
technique is very useful for rotating machinery-related data processing, especially in instances of
varying shaft speed.
Generally, there are two approaches to achieving synchronous sampling - analog and digital
approaches. One of the analog approaches uses an Analog to Digital (A/D) sampling clock to
achieve the synchronous sampling. The key to this approach is generating an appropriate
sampling clock based on shaft rotation conditions. As shown in Figure 3.1, the sampling clock is
derived from the shaft encoder by an analog ratio generator to meet the desired order analysis
requirements (such as order resolution and maximum order). In cases where only lower order
components are of interest, the encoder output can be used as a sampling clock directly.
In the digital approach, or synthesized synchronous sampling [9], both vibration and encoder
signals are discretized simultaneously, preferably at high speed. Different signal processing
techniques can be used to resample the data and convert time domain data into shaft cycle
domain data, with the help of an encoder signal from the shaft (refer to Figure 3.2). In this
approach, both the encoder and key phaser can be used as a shaft speed reference. Regardless,
the availability of the shaft encoder/key phaser is crucial to both analog and digital synchronous
sampling approaches.
13
Figure 3.2. Synchronous sampling digital approach
With the synchronously sampled data, a common way to enhance the signal components of
interest is through time synchronous averaging [10]. With a shaft encoder/key phaser, the
vibration signal detected contains three major components: the synchronous coherent signal
component, the synchronous, non-coherent component, and random noise. Conventional time
synchronous averaging can only enhance the synchronous coherent signal component. The
synchronous, non-coherent component and the random noise will be averaged out with a
sufficient number of averages.
Figure 3.3 (a) shows a simulation result with a combination of the shaft response, a sin (2f 0t ) ; its
second harmonic, b sin (4f 0t ) ; a nonsynchronous coherent signal, c sin (2 1.3 f 0t ) ; and a uniform
random noise. After 250 times of synchronous averaging, the results are shown in Figure 3.3 (b);
the random noise and the nonsynchronous coherent component have been successfully removed.
14
When applying the synchronous sampling technique to diagnose bearing/gear damage, there are
two issues that need to be resolved. The first is that the bearing damage signatures are usually
nonsynchronous to the shaft order. In complicated gear sets, such as in the planetary gear, not all
gear meshing frequencies are integer multipliers of a shaft frequency. So, if the regular time
domain synchronous averaging technique is applied to the vibration response, the bearing
damage signatures will be averaged out. Thus, instead of averaging in the time domain, order
domain averaging should be used in gearbox health feature extractions. The second problem is
unique to a wind turbine gearbox. For example, in a typical 1.5 MW wind turbine, the order
span between the main shaft and the third stage gear meshing (the high speed shaft gear meshing
frequency) can be above 1500. Care must be taken on the order resolution and on the maximum
order applied during the data acquisition.
In the event that the direct shaft tachometer signal is not available, traditional synchronized
sampling becomes difficult, if not infeasible. For example, in this Round Robin project, only the
speed profile is provided. If a synthesized tachometer signal can be generated from the speed
signal, then the equal circumferential space sampling (synchronous sampling) can be carried out
following well-established routines.
The following describes a procedure to synthesize a tachometer from the shaft speed profiles. In
Figure 3.4, assume we have determined that the synthesized tachometer generated a pulse at time
t i . We need to find out the location of the next pulse timing t i +1 . The time elapsed, from t i to
t i +1 , i.e., t1 = t i +1 t i , is the instantaneous shaft rotation period. On the other hand, if we have
determined the t i +1 , since we know the shaft speed as a function of time, the average speed, n ,
between t i and t i +1 can be calculated numerically. Therefore, the shaft instantaneous period can
also be approximated by the averaged instantaneous shaft speed, i.e., t 2 = 60 / n . In theory, by
equalizing t1 and t 2 , we can determine t i +1 ; thus, calculating the next pulse location. In
practice, due to time resolution and speed accuracy, an approximation procedure is used instead
of solving for an exact solution. The following steps further explain the procedure of the
synthesized synchrophaser:
1 ti+1
n(ti+1 ) = ShaftSpeed (t )dt (2)
ti+1 ti ti
15
t1 = t i +1 t i (3)
and the time elapsed by one instantaneous rotation
t 2 = 60 / n (4)
The t i +1 then is the approximate location of the (i + 1)th synchrophaser pulse. The tachometer
can be generated from the synchrophaser, say, by equal spacing between the consecutive
synchrophaser pulses.
With this method, one of the major error sources is the discretization resolution. The maximum
error in the shaft period is T , where T the sampling period is. Fortunately, for bearing and gear
2
dynamic response analysis, especially acceleration enveloping analysis, the frequency of interest
is usually much higher than the shaft speed. In other words, the digitization rate is usually
several orders of magnitude higher than the shaft speed. Thus, the synthesizing error from the
digitization error is expected to be negligibly small.
16
accurately calculate SER. Once the spectrum is generated the SER algorithm sums the
amplitudes of the first six sideband peaks on each side of the center mesh frequency and divides
by the amplitude of the center mesh frequency.
6
i=1 Sideband Amplitude
SER= (5)
Center mesh frequency amplitude
SER is sensitive to the sideband amplitudes relative to the center mesh frequency. In a healthy
gear mesh, any sidebands have small amplitude compared to the center mesh frequency, or they
may be missing altogether resulting in a low SER. SER is typically less than one for a healthy
gear mesh. As damage develops on a gear tooth that passes through the gear mesh, the sidebands
increase in amplitude, as well as in number, and SER will go up. In GE Bently Nevadas wind
turbine condition monitoring system, ADAPT.wind, SER is calculated for the fundamental mesh
frequency and the first two harmonics of each gear mesh.
In theory, any vibration sensor can achieve bearing/gear damage detection through the
enveloping or demodulation processes, as long as the sensor has the frequency range required.
Since the bearing/gear damage excited response is known to have high frequency content, the
accelerometer has an advantage over velocity and displacement sensors.
In the early days, this enveloping detection process was performed using several analog devices.
As shown in Figure 3.5, the conditioned vibration sensor signal is first passed through an analog
filter to isolate the impulse response excited by the bearing damage. The filtered response is
then passed through a rectifier to flip the negative half of the oscillation signal to the positive
side. The rectified signal is fed into an envelope detector to identify the envelope of the signal.
The envelope signal is then used to identify bearing damage signature through a signal analyzer.
If necessary, a low-pass filter can be added before the analyzer.
17
The process shown in Figure 3.5 works well, if all the analog devices are appropriately designed
for a particular application. However, a different application may require different parameter
settings of the analog devices. For example, the bearing support system may have a different
resonant structure; thus, it requires a different cut-off frequency design for the band-pass filter to
isolate the damage impulse response. For different structural damping, the envelope detector
needs a different time constant design to match the impulse response decay rate, so that the
bearing damage-related high frequency and low amplitude vibration signals can be maximized.
More importantly, the bearing damage detection is usually conducted in a harsh environment.
An increased number of electronic components involved in the bearing defect detection process
will usually decrease the overall system reliability.
18
Figure 3.6. Digital processing-based approach
19
Table 3.1. Gear damage features
20
Table 3.2. Bearing damage features
21
Table 3.2. Bearing damage features (continued)
22
3.2.2 Tool Development
A MATLAB-based tool was developed for easy data processing operations. The tool
incorporated regular FFT spectrum analysis, acceleration enveloping analysis, synthesized
synchronous sampling, and order analysis.
3.2.3.1 HSGM
The sensor AN7 was used to evaluate the HSGM stage health condition. From the order
spectrum, it can be clearly seen that the fundamental gear meshing order is heavily modulated by
the HSS speed, as seen in Figure 3.8.The sideband energy ratio (SER) is over three, which
indicates severe gear damage in the HSGM pinion. The sideband contents in the higher order
harmonics of HSGM also indicate the pinion damage, as seen in Figure 3.9 and Figure 3.10.
Post-test examination indicated that the HSP had severe scuffing [4].
23
Figure 3.9. HSGM X2 (44) modulated by HSS (1)
Similar order analyses also indicated that the HSGM order was modulated by the HSIS shaft
(Figure 3.11). This kind of modulation is visible in the HSGMX2 and HSGMX3 as well (refer
to Figure 3.12 and Figure 3.13 respectively). Post-test examination indicated that the HSG also
had severe scuffing [4].
24
Figure 3.11. HSGM (22) modulated by HSIS (0.25)
25
Figure 3.13. HSGM X2 (66) modulated by HSIS (0.25)
3.2.3.2 ISGM
Damage signatures related to the ISGM were not shown in the vibration analysis. As seen in
Figure 3.14 though, the ISGM and higher order harmonics do exist. However, the shaft
modulations are too small to draw any conclusion regarding the gear damage. The post-test
examinations revealed fretting corrosion, polishing wear, and scuffing damage in the ISGM
pinion; however, the damage was imprinted on all teeth because the gear-set had a hunting tooth
gear ratio, which has potentially smoothed out the damages and reduced vibration responses
incurred by the damage.
26
3.2.3.3 PLTGM
For planetary gear stage damage detection, sensors AN3 and AN4 were used (refer to Table 2.3
for more information on sensors). The planetary gear set arrangement in this gearbox is a
simultaneous mesh; thus, there was no cancellation/enhancement around the gear mesh
frequency and harmonics, as seen in the sequential mesh design. Detailed order spectrum
analysis indicated that the planetary gear mesh order and its harmonics are present Figure 3.15
for sensor AN4). Sidebands of planet passing are visible around the PLTGM and its harmonics.
The modulation is significant at the third harmonic of the PLTGM (Figure 3.16). Similar
information can be extracted from the sensor AN3 and the corresponding order spectra are
shown in Figure 3.17 and Figure 3.18, respectively.
The acceleration enveloping analysis of the sensor AN3 (Figure 3.19) indicated that the damage
is likely associated with the ring gear (0.037 Order), instead of planet gear (0.087 Order) or sun
gear (0.174 Order). Post-test examination confirmed the ring gear scuffing and polishing
damage [4].
The post-test also indicated severe fretting corrosion on the sun pinion. Though the sun/plant
gear meshing is evident in the data analysis, as shown in Figure 3.15 and Figure 3.17, the LSIS
shaft modulation (0.07 Order) is very small (Figure 3.20); thus, the sun pinion damage is
inconclusive using current vibration-based analysis.
Figure 3.15. Planetary gear stage meshing order and harmonics from AN4
27
Figure 3.16. PLTGM X3 modulated by planet passing order (0.037)
Figure 3.17. Planetary gear stage meshing order and harmonics from AN3
28
Figure 3.18. PLTGM X3 and X4 modulated by planet passing order (0.037)
29
Figure 3.20. Modulation by LSIS in PLTGM is very small
The post-test examination revealed that the HSS-A1 bearing had mild overheating, which created
straw-yellow temper colors near each end of the IR raceway [4]. However, it is not likely the
overheating caused the BPFI response in the envelope spectrum. Acceleration enveloping
analysis can be very sensitive to bearing mechanical damage. It is likely that the damage
response was caused by the minor scratches on the race edge or the small indentations in the
middle of the raceway.
An acceleration enveloping analysis on the high speed shaft sensor AN7 also indicated possible
HSS-A1 cage damage. In the envelope spectrum (Figure 3.22), there is a minor but clear tone at
0.4261 Order and its harmonics at 0.8522 and 1.278 Order. By referring to Table 3.2, we can see
that this fundamental tone of 0.4261 Order is very close to HSS-A1 cage damage tone of 0.425
Order. However, HSS-A1 cage damage was not reported in the post-test inspection.
30
Figure 3.21. Envelope spectrum of AN7
By careful re-examination of the order spectrum Figure 3.23, we saw that there was a small peak
at 2.43 Order, which is very close to bearing D BPFI 2.457 Order; however, it was also the
second harmonic of the planet gear mesh between the ring gear and the planets. Therefore, it
31
was very difficult to draw a conclusion about the bearing D BPFI damage signature at the current
stage.
3.3 Discussions
This Round Robin project provided a platform for commercial wind turbine condition
monitoring system suppliers, as well as, academic institutions to exercise different
methodologies. The GE Bently Nevada team provided a unique set of solutions based on the
techniques developed and implemented into its commercially available wind turbine drivetrain
condition monitoring system, ADAPT.wind.
Compared to the condition monitoring systems for other rotating machinery, condition
monitoring systems for wind turbines face a few unique challenges including constantly variable
operating speeds and a high gear-up ratio from the rotor to the high-speed shaft. To overcome
the inaccuracy incurred by speed variations, synchronous sampling is the preferred data
acquisition technique. Subsequently, a synchronous analysis technique can be used to extract the
bearing and gear damage signatures. To accommodate high gear ratios in wind turbine
gearboxes, a digital domain synchronous re-sampling is very useful for signature extraction. The
digital synchronous re-sampling utilizes a low count mechanically- or electrically- based
32
encoder, or a key phaser, and interpolates the data between the pulses, with linear or more
sophisticated speed variation assumptions.
The data provided by NREL for the Round Robin analysis included accelerometer data and
speed profile data. To carry out the digital synchronous sampling, without shaft encoder or key
phaser data, the GE team developed and implemented a so-called synthesized synchronous
sampling technique. Although the phase information cannot be recovered, the technique
successfully carried out the synchronous analysis using the shaft speed profile.
The vibration sensing-based method is believed to be the cost effective approach for wind
turbine condition monitoring. However, any gearbox anomalies that do not incur additional
vibrations, such as overheating, minor fretting, and smoothed polishing will not be effectively
detected by vibration sensors.
Generally speaking, vibration-based wind turbine condition monitoring systems can detect
damage in the high speed side with higher confidence than that from the low speed side,
especially planetary gear set-related component damage. This is because accelerometers are
inherently more sensitive to high frequency vibrations. In addition, the mechanical
transmissibility from planetary gear components is usually low. Damage detection and condition
monitoring related to the gearbox low speed side is an area that needs more research attention in
the future.
33
4 Combining Novel Approaches with Proven Algorithms for
Robust Wind Turbine Gearbox Fault Detection
Jeremy Sheldon*, Matthew Watson, Genna Mott and Hyungdae Lee
Impact Technologies, A Sikorsky Innovations Company
*
Corresponding Author Email: [email protected]
4.1 Introduction
Impact Technologies analysis efforts focused on applying a number of novel vibration
diagnostic algorithms to the GRC data set. They have been developed and matured by the team
in Department of Defense (DOD) applications for more than 10 years. The algorithms and results
are summarized herein. Generally, the methods employed by the team worked well, once the
challenges and peculiarities of the data set were realized. In particular, the absence of raw, time-
domain data from a healthy system (only FFT plots were provided) made it difficult to baseline
the system for comparison purposes. Regardless, the results of the automated algorithms were
corroborated with visual spectral analysis and are provided herein.
For example, Figure 4.1 shows the result of the authors previous analysis of a gear pinion failure
that occurred on the test stand of a high-speed (thousands of RPMs), high-power (tens of
thousands of horsepower) military fighter aircraft drive train. As seen, several vibration features
react simultaneously, indicating that a potential fault is present in the system. Information
gathered solely from this sensor would confidently indicate a fault. However, upon further
investigation of the raw sensor data (shown in the top plot of the figure), one can see that this
reaction was caused by faulty (intermittent) data and, therefore, should not be trusted.
34
Figure 4.1. False alarm caused by faulty sensor
In addition, because many diagnostic feature algorithms are based on higher order statistics and
energy measures; it is possible for a corrupt signal to generate feature values that are within an
acceptable range despite the signal containing no periodic frequency content. Rigorous and
automated analysis of the integrity of accelerometer data is, therefore, critical to providing
accurate health assessments. To address this potential source of false alarms, the validity of the
high frequency vibration sensors is first evaluated as an initial step in the analysis, using an
approach termed FirstCheck. This module looks at a number of signal characteristics (including
the range, bias, and other proprietary characteristics) to verify the integrity of the vibration signal
before it is analyzed by the other algorithms.
The ImpactEnergy module (Figure 4.2) integrates traditional spectral analysis techniques with
high-frequency demodulation and advanced feature extraction algorithms, providing a more
effective solution. The advantages of using the high frequency response to identify and track
35
bearing damage is well documented [14,15] and proven to be an effective method.
Demodulation, or enveloping, allows the broadband energy caused by failure effects to be
differentiated from the energy due to normal system noise. This approach provides the ability to
detect defect impulse events much easier than traditional analysis techniques allow. A key
consideration is selecting the band-pass filter that is centered on the expected carrier frequencies.
Through proprietary knowledge and field-application experience, the authors have developed a
process to identify key carrier frequencies.
For complete characterization of bearing health from incipient fault to failure, the ImpactEnergy
module includes algorithms to extract an extensive set of time and frequency domain features
from both the raw, unprocessed, and demodulated vibration signals. Some time domain features
include traditional statistical measures, such as RMS, kurtosis, and Crest Factor. Frequency
domain features include the power levels of specific bearing defect frequencies compared against
known, health baseline thresholds, which can be very useful in diagnosing a fault [16].
4.2.3 GearMod and GearMod-Shaft: Gear and Shaft Fault Detection and Isolation
The GearMod module (Figure 4.3) is used to extract diagnostic features that are used for gear
fault detection and isolation. This module contains a broad range of statistical methods based on
the time synchronous averaged (TSA) signal and other processed signals. The time synchronous
averaging technique is a useful technique to reduce the random noise level, as well as
disturbances from events unrelated to the gear of interest, and it has been extensively used to pre-
process gear vibration signals [17,18]. The fundamental principle of the TSA is that the vibration
components related to a shaft rotation and the gears on that shaft repeat periodically with the
shaft rotation. By dividing the vibration signal into contiguous segments, of exactly one shaft
rotation, and averaging a sufficiently large number of segments, the vibration components that
are synchronous to the shaft rotation, are reinforced. Non-synchronous vibrations are cancelled
out because they are out of phase in consecutive rotations.
36
Figure 4.3. GearMod overview
GearMod calculates time-domain features, such as RMS, skewness, kurtosis, and Crest Factor, as
well as features from the spectrum of the averaged signal, including FM0 (the peak-to-peak
amplitude compared to summation of GMF & harmonic magnitudes), Sideband Index (the
average spectral magnitude from sidebands on the 1st GMF), and Sideband Level Factor (the
spectral magnitude of sidebands on the 1st GMF, normalized by TSA RMS). The equations for
some of these features are included in Table 4.1 [19-21].
PPA
FM 0 = n PPA = peak-to-peak amplitude
FM0
A( fi )
i =1
A(fi) = GMF & harmonic amplitudes
RMC1sb, ndo min ant 1 ( s ) + RMC1sb, ndo min ant +1 ( s ) RMC1sb, ndo min ant 1 and RMC1sb, ndo min ant +1
Sideband Index SI = = 1st order sideband amplitudes about
2 fundamental GMF
Additional features are also calculated using proprietary methods. In addition, GearMod contains
built-in functionality to extract the TSA signal without a tachometer signal, making it appropriate
for situations when an accurate or useable tachometer signal does not exist.
37
Spectral smearing, in which energy from an evolving characteristic frequency (i.e., shaft
frequency, bearing fault frequency, gear mesh frequency) is spread across multiple frequency
bins, can reduce the efficacy of traditional frequency domain analysis, including Fourier
transforms. Typically, this is avoided by defining steady-state operating conditions in which to
perform the analysis. Although this may be acceptable for some systems, most wind turbines
have constantly varying shaft speeds and loads. In addition, certain component faults and their
progressions can also lead to non-stationary signals that could be missed by traditional
techniques. As a result, the authors have developed a novel vibration diagnostics methodology
that is applicable during non-steady operation through application of joint time-frequency
analysis (JTFA)[22]. These methods use various techniques to transform the two dimensional
time domain signal into a three dimensional, time-frequency domain signal to increase feature
extraction accuracy. An example is shown in Figure 4.4. Various features are then extracted from
the three dimensional signals for fault detection.
Segment 1
Extract
Spectrum
Segment 2
Extract
Spectrum
Extract
Segment n
Spectrum
Mag
Mag
Time
38
Table 4.2. Initial blind results summary
39
Figure 4.5. High speed gear fault evidence, blind results (AN6)
Figure 4.6. High speed gear fault evidence, blind results (AN7)
40
4.3.1.2 Intermediate Shaft Bearings
Impacts ImpactEnergy analysis of the bearings on the intermediate shaft is summarized by the
spectral plots in Figure 4.7 and Figure 4.8. For the downwind bearings (IMS-SH B&C, which are
indistinguishable since they are the same model bearing running at the same speed), the clear
peaks in the conventional, non-demodulated spectrum (Figure 4.7) at the outer raceway fault
frequency indicated substantial localize defects. Notice, too, the clear spectral peak at the IMS-
SH-A inner raceway defect frequency, in both the faulted and baseline (supposed healthy) FFT.
Because the bearing fault frequency was visible in the baseline FFT data that was provided,
Impact was not confident in the evidence enough to call out the fault. As such, this fault was not
called out in the blind analysis (see Section 4.3.2.1 for more details). .
Figure 4.7. Intermediate speed downwind bearing fault evidence, blind results (AN6, Data 2b)
Damage in the IMS-SH B component was also evident by the elevated levels of multiple
harmonics of the inner raceway fault frequency seen in the demodulated spectrum (Figure 4.8).
41
Figure 4.8. Intermediate speed downwind bearing fault evidence, blind results
42
Figure 4.9. Sun pinion gear fault evidence, blind results
43
Figure 4.10. JTFA speed gear fault evidence, blind results
44
Table 4.3. Post-inspection results summary
45
With this information and the results exemplified by Figure 4.12, Impact could correctly identify
the fault as a progressed or distributed inner raceway fault because there was little modulation
and clear inner raceway frequency vibration.
Figure 4.12. Intermediate speed upwind bearing fault evidence, post-inspection results
Using the correct fault frequencies, Impacts analysis clearly showed small, early stage inner
raceway defects, as evident by the multiple BPFI harmonics (1x-4x) that are dominant in the
demodulated FFT (for data sets 2a and 2c). Example demodulated spectra and fault frequency
peaks are shown in Figures 4.13 and 4.14. Although Impact did detect a defect in the bearing, the
diagnosed fault signature seems unlikely to have been caused by only the cited overheating
event. Instead, Impact believes the evidence points to some additional defect, which may have
resulted from the overheating.
46
Figure 4.13. High speed downwind bearing fault evidence, revisited (Data 2a)
Figure 4.14. High speed downwind bearing fault evidence, revisited (Data 2c)
47
4.4 Lessons Learned and Conclusions
Impacts blind analysis successfully detected three of the seven damaged components that were
present in the gearbox, with a fourth component detected but not called out due to the relative
immaturity of the approach that detected the damage. Of the remaining three, two were detected
during the secondary analysis that was performed after the inspection report was provided. These
two were originally missed due to: 1) the use of incorrect bearing design information that was
provided in the original analysis, and 2) the presence of the damage during assembly, which
caused the baseline results that were used for comparison to be higher. It is Impacts opinion that
the seventh damaged component is not detectable with vibration analysis since the damage
appears on the outside bore of the outer raceway, not in the contact zone of the bearing.
In general, this analysis was confounded by the number and severity of the defects, especially the
obfuscation of bearing and other faults by the widespread gear damage. Impact believes that in
actual practice the expected diagnostic performance will be better since a smaller and less severe
set of faults will be present during the early stages of fault evolution. To clarify, this data set
contains multiple progressed faults that would be better detected as they individually occur over
time versus diagnosing each one once they have all occurred. Additionally, assessing the fault
severity was more difficult due to the lack of time-domain baseline data. Although this was
overcome by using our experience of previous analysis of different machinery, higher fidelity
results may be produced by comparing like data to like data and allowing the algorithms to be
baselined against typical, healthy vibration levels and trended over time. Regardless of these
minor issues, this analysis and effort should enable a good assessment on the state of wind
turbine diagnostics, as well as an indication of what work remains.
48
5 Analysis Algorithms and Diagnostics Results from NRG
Systems
Eric Bechhoefer
NRG Systems
Email: [email protected]
5.1 Introduction
NRG Systems, as a participant of this Round Robin study, wished to validate the effectiveness of
analysis algorithms used in the aerospace community, for the wind industry. The aerospace
industry, specifically vertical flight, developed an extensive toolset for gearbox fault detection as
a result of helicopter gearbox failures. These algorithms were specifically developed for shaft,
gear, and bearing fault detection. This is counter to most installed condition monitoring systems,
which have matured out of industrial monitoring of large turbo-machinery (turbo-machinery
have no gearbox, focusing extensively on shaft misalignment, out of balance, or rub conditions).
NRG implemented two analysis methodologies: synchronous analysis of shaft/gear components
and non-synchronous analysis of bearings. Synchronous methods were based on the work of
McFadden [23], while the bearing analysis was based on the work of Randall [24].
The data set consisted of 10 samples each under three operating conditions, for a total of 30 files.
The samples were separated by a short time interval. Because these industrial gearboxes are
designed to run for years versus minutes, we assumed that the analysis would consist of taking a
snap shot of the current gearbox condition. No attempt was made to trend component condition
indicators or look at statistical differences between early and late files. It was assumed that over
the period of the test, there was no appreciable degradation of the gearbox. The analysis
consisted of viewing the output of various analysis algorithms for each component, and based on
some nominal experience with aerospace gearboxes, defining the wind turbine gearbox as good,
bad, or indifferent.
Gear analysis was based on operations of the time synchronous average [23]. Time synchronous
averaging (TSA) is a signal processing technique that extracts periodic waveforms from noisy
data. The TSA is well suited for gearbox analysis, where it allows the vibration signature of the
gear under analysis to be separated from other gears and noise sources in the gearbox that are not
synchronous with that gear. Additionally, variations in shaft speed can be corrected, which
would otherwise result in spreading of spectral energy into adjacent gear mesh bins. To do this, a
signal is phased-locked with the angular position of a shaft under analysis.
49
This phase information can be provided through an n per revolution tachometer signal (such as a
Hall sensor or an optical encoder, where the time at which the tachometer signal crosses from
low to high is called the zero crossing) or though demodulation of gear mesh signatures [25]. In
the case of the GRC data set, the phase information was extracted from the generator signal itself
(e.g., voltage signal capacitively coupled onto the RPM signal).
The model for vibration in a shaft in a gear box was given in [23] as:
The mesh frequency is a function of the shaft rotational speed: fm = Nf, where N is the number of
teeth on the gear and f is the shaft speed, with no reduction in the analysis performance. This
vibration model assumes that f is constant. In most systems, there is some wander in the shaft
speed due to changes in load or feedback delay in the control system. This change in speed will
result in smearing of amplitude energy in the frequency domain. The smearing effect, and non-
synchronous noise, is reduced by re-sampling the time domain signal into the angular domain:
mx() = E[x()] = mx(+). The variable is the period of the cycle that the gearbox operation
is periodic, and E[] is the expectation (e.g., ensemble mean). This results in the assumption that
mx() is stationary and ergodic. If this assumption is true, then non-synchronous noise is reduced
by 1/sqrt(rev), where rev is the number of cycles measured for the TSA.
50
Figure 5.1. Generation of the TSA and selected CIs
Residual Analysis. Shaft order 1, 2, and 3 frequencies, and the gear mesh harmonics, of
the TSA are removed. Faults such as a soft/broken tooth generate a 1 per rev impact in
the TSA. In the frequency domain of the TSA, impacts are expressed as multiple
harmonics of the 1 per rev. The residual analysis removes the shaft order 1, 2, and 3
frequencies and gear mesh harmonics in the frequency domain, and then the inverse FFT
is performed. This allows the impact signature to become prominent in the time domain.
CIs are statistics of this waveform (RMS, peak to peak, Crest Factor, and kurtosis).
Energy Operator (EO), which is a type of residual of the autocorrelation function. For a
nominal gear, the predominant vibration is gear mesh. Surface disturbances and scuffing
generate small higher frequency values, which are not removed by autocorrelation.
Formally, the EO is: TSA2:n-1 x TSA2:n-1 TSA1:n-2 x TSA3:n . The bold indicates a vector
of TSA values. The CIs of the EO are the standard statistics of the EO vector.
Narrowband Analysis operates the TSA by filtering out all tones except that of the gear
mesh and with a given bandwidth. It is calculated by zeroing bins in the Fourier
transform of the TSA, except the gear mesh. The bandwidth is typically 10% of the
number of teeth on the gear under analysis. For example, a 23-tooth gear analysis would
retain bins 21, 22, 23, 24, and 25, and their conjugates in the Fourier domain. Then, the
inverse FFT is taken, and statistics of the waveform are taken. Narrowband analysis can
capture sideband modulation of the gear mesh tone due to misalignment, or detect a
cracked/broken tooth.
51
Amplitude Modulation (AM) analysis is the absolute value of the Hilbert transform of the
Narrowband signal. For a gear with minimum transmission error, the AM analysis feature
should be a constant value. Faults will greatly increase the kurtosis of the signal.
Frequency Modulation (FM) analysis is the derivative of the angle of the Hilbert
transform of the narrowband signal. It is a powerful tool capable of detecting changes of
phase due to uneven tooth loading, a characteristic of a number of fault types.
For a more complete description of these analyses, see [25] or [26]. Figure 5.2 is an example of
the processing to generate the gear CIs for a spiral bevel gear with surface pitting and scuffing.
The cepstrum was also evaluated, although it is difficult to implement in an automated system.
For bearing analysis, the envelope analysis was used.
52
Figure 5.3. Synthetic tachometer
53
5.3.3 Intermediate Speed Shaft/High Speed Gear and Intermediate Speed Pinion
The ISS/HSG/ISP showed a large shaft order 4 in the TSA (Figures 5.5a and 5.5b). Since there is
a 4:1 relationship with the HSS, this further confirms the eccentricity of the HSP. Both the HSG
and ISP showed large (e.g., .05 Gs) energy operator, indicative of sever pitting/scuffing. The
AM and FM signals did not indicate any missing teeth/soft teeth, but did reflect a high variation
in loading due to the HSP (Figure 5.5b).
Figure 5.5a. TSA intermediate shaft Figure 5.5b. Intermediate speed pinion, where
the units for the Energy Operator,
Narrowband and Amplitude Modulation
analysis are in Gs, and the Frequency
Modulation analysis is in radians.
54
Figure 5.6a. Sun gear Figure 5.6b. Planet gears
55
x 10AN6 ISSRadial Bearing: Int Speed Shaft: Downwind
-3
8
Cage: 2
Ball: 50
7 Outer: 85
Gear: 5 Hz Inner: 70
6
Gs
5
2
0 50 100 150
Freq (Hz)
Figure 5.8a. High speed shaft, downwind side Figure 5.8b. Intermediate speed shaft
downwind side
It was found (Figures 5.8a, 5.8b, and 5.9) that most bearings had some level of damage. The
analysis could not find direct evidence of a fault on the planet bearing. While some attempt was
made in optimizing the window frequency using spectral kurtosis, its likely that the lack of
performance was due to poor window selection in the envelope analysis.
5.4 Discussion
The lack of fleet data forces one into an analysis of individual algorithm waveforms. It is
desirable to have vibration analysis off of a fleet of gearboxes to compare the test gearbox
against. Additionally, it is desirable to have at least six months of condition indicator data on a
gearbox to observe (or capture) degradation. Because of the lack of fleet data or any appreciable
history, analysis was based on a by eye analysis. It is likely that better performance could be
gained with more experience on this gearbox, or in comparison of waveforms with a known,
good gearbox. The algorithms, based on the TSA, appear to find faults that were consistent with
[26]. In fact, most analysis did show a response indicative of a wear/fault. Because there seemed
to be no broken/soft tooth, the residual analysis was nominal on all gears. It was noted that
56
cepstrum analysis showed numerous harmonics (indicative of fault), but because the cepstrum is
not a synchronous analysis, it was difficult to assign a particular frequency with a component.
In general, the analysis methodology seemed appropriate for wind turbines, and has the
advantage of being relatively simple to implement in an autonomous manner, e.g., the generation
of statistics from the analysis waveforms, which could be trended or a threshold set to indicate
when a maintenance action needs to be performed. This lends credence that aerospace gearbox
analysis techniques are appropriate for wind turbine gearbox analysis.
57
6 Review and Application of Methods and Algorithms in Wind
Turbine Gearbox Fault Detection
David Siegel*, Wenyu Zhao, Edzel Lapira, Mohammed AbuAli, and Jay Lee
Center for Intelligent Maintenance Systems, University of Cincinnati
*
Corresponding Author Email: [email protected]
6.1 Introduction
This chapter contains a description of the method and algorithms used by the research team at the
University of Cincinnati Center for Intelligent Maintenance Systems (IMS) for the Round
Robin study. The outline of this chapter is as follows: Section 6.2 provides an overview of the
signal processing and feature extraction methods evaluated in the study by the IMS research
team, followed by more specific details for each method. Section 6.3 provides a summary of the
results for each evaluated method along with some additional discussion. The final summary
table in Section 6.3 also provides an indication of which analysis methods were considered prior
to and after the failure report was released. Lastly, conclusions and suggestions for future work
are provided in Section 6.4.
For bearing condition monitoring, the most established method in the literature is bearing
envelope analysis, also called the high frequency resonance technique [24]. The general concept
is that a spall or damage on the bearing race or rolling element causes a series of impacts that
excite the structural resonances of the mechanical system; this causes an amplitude modulation
effect in which the carrier frequency is the resonance frequency and the bearing fault frequency
is the modulation frequency. By filtering around the excited resonance and performing the
demodulation, the envelope spectrum, along with the calculated fault frequencies, can be used to
diagnose the bearing condition. A more detailed description of bearing faults, the bearing
envelope analysis method, and methods for selecting the band-pass filter frequency range are
provided in [24]. Despite bearing envelope analysis being a very effective technique; the method
usually requires a high sampling rate since the excited resonance can occur at frequencies above
10 KHz for many applications. The selection of the band-pass filter is also a crucial aspect in the
method and a current area of research [24]. The use of spectral kurtosis filtering can be used to
select the filter band for the bearing envelope analysis method, as well as for calculating
indicators for the overall health condition of the monitored gearbox [28].
58
Time synchronous averaging represents one of the most established signal processing techniques
for gear condition monitoring. The method is ideally suited for the processing of gear vibration,
since the synchronous averaging method enhances and separates the periodic gear vibration from
the cyclostationary vibration of rolling element bearings. Additional processing methods can be
performed on the time synchronous average signal, including the gear residual signal and the
amplitude and phase modulation functions [18]. For planetary gearboxes, due to the relative
motion of the planet gears and the multiple contact points between each planet gear meshing
with the sun and ring gear, the traditional synchronous averaging algorithm is not able to isolate
the individual vibration for each planet gear or the sun gear. Specific algorithms for performing
synchronous averaging for planetary gears are also evaluated in this work; the method suggested
by McFadden [29] is used in this study. This specific algorithm for planetary gearboxes does
have some potential drawbacks; in particular, the long data acquisition period needed to perform
the calculation procedure is a major challenge for implementing this method.
59
Table 6.1. Summary of evaluated methods advantages and disadvantages
60
rotational speed; this initial example provides some evidence that the high speed pinion is
degraded due to the sidebands observed in the vibration spectrum.
To further quantify these observations from the vibration spectrum; a set of gear wheel vibration
features was extracted using the baseline spectrum and the spectrum from the degraded gearbox.
To quantify the magnitude of the sidebands, the sideband level was calculated using Equation
(6). In this calculation, SBLa stands for the sideband level, Sba1 is the magnitude of the lower
sideband and SBa2 is the magnitude of the upper sideband. In addition, a sideband ratio was also
calculated using Equation (7); this normalizes the sideband ratio by the gear mesh frequency
peak. Prior work by Combet et al. [31] has shown this sideband ratio feature to be an effective
metric to quantify gear health since it is less susceptible to load fluctuations due to the sideband
magnitude being divided by the gear mesh frequency peak [31]. Table 6.2 provides a listing of
the frequency domain gear features; a total of 16 were calculated. For each respective gear, four
features were calculated. The sideband ratio and sideband level were calculated for the gear
mesh frequency and the first harmonic of the gear mesh frequency. Also, the frequency domain
gear features were only calculated for the four gears on the parallel gearbox stage; the analysis of
the sideband patterns for the planetary gearbox is quite complicated and more advanced
techniques were evaluated for the planetary gearbox.
61
Table 6.2. Frequency domain gear features
Signal Feature Name # of Features
Side Band Ratio and Sideband level for Intermediate
AN5 4
Speed Shaft Gear
Side Band Ratio and Sideband level for Intermediate
AN6 8
Speed Shaft Pinion and High Speed Shaft Gear
Side Band Ratio and Sideband level for High Speed Shaft
AN7 4
Pinion
62
Figure 6.2. Sideband ratio gear features Case C: (a) Low speed shaft pinion;
(b) High speed shaft gear; (c) High speed shaft pinion
63
Figure 6.3. Real cepstrum - Case C: top plot - AN7 baseline; bottom plot - AN7 degraded gearbox
Figure 6.4. Cepstrum peak features from Case C: blue baseline; red - degraded gearbox
64
Additional vibration features were extracted from the cepstrum at the corresponding shafts, using
data from both the baseline gearbox and the degraded gearbox. Example results from the
cepstrum features are provided in Figure 6.4, in which several peaks in the cepstrum are
noticeably larger in magnitude when comparing the degraded gearbox to the baseline gearbox.
The cepstrum peak related to the high speed shaft pinion is highlighted, since this feature was
dramatically larger in magnitude for the degraded gearbox. This provides an additional set of
evidence that the high speed shaft pinion is damaged. To further quantify the difference in the
cepstrum features from the baseline state, and diagnose which gear in a meshing pair had the
most severe condition, a cepstrum based health indicator using Equation (9) was calculated [32].
In this calculation, d(t) is the cepstrum health indicator, Ap(t) is the cepstrum peak for the input
gear at time t, Ar(t) is the cepstrum peak for the output gear at time t, Ap(0) is the cepstrum
baseline peak for the input gear, and Ar(0) is the cepstrum baseline peak for the output gear. For
a monitored system, this health indicator would be zero in the baseline condition, close to -1, if
the output gear is degraded, and close to 1 if the input gear is degraded. Figure 6.5 provides a
result from this health indicator calculation, for the high speed shaft gear and pinion meshing
pair. The health indicator is near -1 for all 10 data samples, which indicates that the high speed
pinion is the gear with the more severe level of damage, according to this metric. This agrees
with the previous result from the sideband features in the vibration spectrum, in which there was
evidence of damage on the high speed shaft pinion, but little evidence of damage on the high
speed shaft gear.
0.4
Cepstrum Peak / Ratio
0.2
-0.2
-0.4
-0.6
-0.8
-1
1 2 3 4 5 6 7 8 9 10
Sample #
Figure 6.5. Cepstrum health Indicator for Case C calculated for high speed shaft gear and pinion
65
6.2.3 Bearing Envelope Analysis
Although the area of bearing condition monitoring has been an area of research for quite some
time, with new algorithms and methods proposed each year; the bearing envelope analysis
method remains one of the more effective techniques for bearing condition monitoring [28]. A
more detailed description of bearing envelope analysis can be found in prior work [28, 33],
therefore, this study simply provides a brief review of the method and highlights the results for
this study. A flow chart that shows the steps in this algorithm is provided in Figure 6.6, in which
the initial step is to band pass filter around an excited natural frequency. The selection of the
band-pass filter center frequency and bandwidth is an important step and also a current area of
research [28]. In this study, the filter parameters were selected by inspection of the frequency
domain spectrum for the respective accelerometers, but alternative methods could also be
considered. After filtering the signal, the Hilbert Transform is used to extract the envelope
signal, which is further analyzed in the frequency domain. For a bearing with damage on the
rolling element or bearing races, the bearing fault frequency peaks are usually much easier to
distinguish in the envelope spectrum when compared to the frequency spectrum.
Vibration Signal
66
bearing. The envelope spectrum in Figure 6.8 (b) shows a peak at the BPFO for the planet
carrier upwind bearing; this bearing also had outer race damage according to the failure report.
Using a filter from 9,500 Hz-10,500 Hz would have resulted in a missed detection for the planet
carrier upwind bearing. Using the frequency band from 4000 Hz - 6000 Hz resulted in a
detection of an outer race fault on the planet carrier bearing; however, it provided a less clear
detection for the ISS downwind bearing, in which only the first harmonic of the BPFO could be
identified.
Figure 6.7. Envelope spectrum - Case C: (a) AN6 - peaks at BPFI for ISS upwind bearing and HSS
downwind bearing; (b) AN7 - BPFI peak for HSS downwind bearing
Figure 6.8. Envelope spectrum accelerometer AN10 - Case C: (a) band-pass filter from 9500 Hz -
10,500 Hz, peaks at BPFO and 2X BPFO for ISS downwind bearing; (b) band pass filter from 4000
Hz - 6000 Hz, peak at BPFO for planet carrier upwind bearing and also peak at 2X BPFO for ISS
downwind bearing
67
6.2.4 Spectral Kurtosis Filtering
For condition monitoring of mechanical systems, the vibration signals for damaged gear and
bearing components typically display an impulsive signature. Detecting that impulsive signature
is not a trivial task since the signature could be masked by other sources of vibration.
Techniques and filtering methods based on spectral kurtosis are aimed at finding the optimal
frequency band for recovering the impulsive fault signature that could be hidden in the raw
vibration waveform. A brief review of the calculation procedure and the results of this study are
provided, and the interested reader is referred to the work by Antoni et al [30] and Combet et al
[34] for a more detailed discussion on the use of spectral kurtosis for filtering vibration signals.
The initial step in this algorithm is to calculate the short time Fourier Transform of the vibration
signal, denoted by H(t,f). Equation (10) indicates that the average value of the fourth power of
H(t,f) is divided by the mean square value of H(t,f), which provides a kurtosis value as a function
of frequency. The Wiener filter is constructed using the kurtosis values for each frequency bin,
as shown in Equation (11); the frequency bin is only included if the kurtosis value is above a
statistical threshold at a given confidence level [30]. The Wiener filter is then multiplied by the
frequency domain representation of the original signal, X(f), and the result is transformed back to
the time domain as indicated in Equation (12). The advantage of this method is that the signal is
filtered without any a priori knowledge of which frequency band to filter in, and instead is based
on which frequency band is most impulsive.
H 4 (t , f ) (10)
Kr ( f ) = 2
2
H 2 (t , f )
K ( f ) for K r ( f ) > s (11)
W ( f ) = r
0 Otherwise
{
y (t ) = 1 W ( f ) X ( f ) } (12)
In this study, the filtering algorithm was used to process data for all 12 accelerometers using a
block size of 256 data samples and an overlap of 80% when performing the short time Fourier
Transform calculation. When applying this processing method, only high kurtosis values were
observed for accelerometers AN3 and AN4; thus, the example results do not include the other
accelerometers. An example result from the filtering method is provided in Figure 6.9 from
accelerometer AN4, in which one can observe that the Wiener filter is focused on the high
frequency content of the signal from approximately 8 KHz - 18 KHz. This implies that although
the rotational frequencies of the carrier are quite low; structural resonances at a high frequency
appear to be excited by defects and damage from the internal components within the planetary
gearbox. Figure 6.9 illustrates how the impulsive signature is masked in the raw time signal, but
is quite clear in the filtered signal; the raw signal has a kurtosis value of only 3.39 compared to a
kurtosis value of 169 for the filtered signal. Further examination of the filtered signal shows a
pattern that repeats for every 2 revolutions of the carrier. This periodic pattern in the filtered
signal from accelerometer AN4 is shown in Figure 6.10. The high kurtosis value of the filtered
signal, along with the periodic pattern that is related to the carrier rotation, and the location of the
accelerometer each point to a problem with the internal components in the planetary gearbox.
However, it was difficult to determine which gear or bearing component was the cause of this
problem from the filtered signal and the envelope spectrum; a potential reason is that multiple
68
faults occurred in the planetary gearbox. The results from the failure report indicate that the ring
gear and sun pinion both had scuffing and corrosion damage and the planet carrier's upwind
bearing had damage on the outer race.
Figure 6.9. (a) Wiener filter based on spectral kurtosis; (b) raw and filtered AN4 accelerometer
signal Case A
-100
0 1 2 3 4 5 6 7 8
Time (s)
Filtered AN4 Signal (Third and Fourth Rotations)
100
Vibration (m/s 2)
-100
0 1 2 3 4 5 6 7 8
Time (s)
Filtered AN4 Signal (Fifth and Sixth Rotations)
100
Vibration (m/s 2)
-100
0 1 2 3 4 5 6 7 8
Time (s)
Figure 6.10. Filtered AN4 signal showing the periodic repetition based on 2 revolutions of the
carrier Case A
69
The results in Figures 6.9 and 6.10 and were from an example data file from Case A. The same
filtering method was also applied to the remaining data files and the kurtosis feature from the
filtered signal was stored. The results of the filtered kurtosis value are provided in Figure 6.11
and provide an interesting point for discussion. The kurtosis values were very large for files
from Case A, with values above 50. It is also worth noting that the kurtosis values show a
decreasing trend for the data files in Case A, in which each file was collected in sequential order.
For Case B and Case C, the kurtosis values are quite low and in a normal range between 2 and 4.
This implies that the fault signature was not present for accelerometer AN3 and AN4 for the
operating conditions in Case B and Case C. This could imply that the operating conditions for
Case A are more conducive for detecting this type of problem in the planetary gearbox.
However, it is also worth noting that the data was collected in sequential order from Case A to
Case C; this could imply that the vibration signature became less impulsive with the running
time of the gearbox. Despite these discussion points for potential reasons as to why the signature
was not present in all three operating conditions, the spectral kurtosis filtering method provided a
very clear detection of a problem in the planetary gearbox for Case A, based on the very high
kurtosis values. This provided enough confidence and evidence to believe that the internal
components in the planetary gearbox stage were degraded.
150
100
50
0
0 5 10 15 20 25 30
Sample #
70
The additional processing methods for the residual signal (section 6.2.5.1) and the amplitude and
phase modulation signals (section 6.2.5.2), first require the extraction of the time synchronous
average signal for each shaft. The time synchronous averaging algorithm requires a reference
pulse train for aligning the data with respect to a given shaft, and ensemble averaging the signal
over several rotations. In this study, a tachometer signal was not provided, but an alternative
method was used for extracting a reference signal. The provided speed signal had a clear tone at
the generator shaft speed (20 Hz or 30 Hz); band pass filtering in a range between 15 Hz 35 Hz
provided a way of extracting a pulse train from the speed signal. The filtered speed signal was
used as a surrogate for the tachometer signal; alternative methods that use the gear mesh
frequency peak for estimating a synthetic tachometer signal could also have been used [35].
With the necessary reference signal, the vibration signals could be aligned and ensemble
averaged with respect to the carrier, low speed shaft, intermediate speed shaft, and the high speed
shaft, using the established synchronous averaging methods. More specific details on the
synchronous averaging method including the different interpolation methods; the frequency
domain implementation can be found in [36]. In this study, a time domain interpolation method
was used. From the provided vibration signals, accelerometer AN3 was used for calculating the
time synchronous average signal for the carrier shaft, AN5 was used for the low speed shaft,
AN6 was used for the intermediate speed shaft, and AN7 was used for the high speed shaft.
The time synchronous average signal and the residual signal are shown in Figure 6.12 for the
high speed shaft pinion. The kurtosis value of the residual signal for this gear is quite low (2.34),
and there appear to be no abnormalities that can be seen in the time averaged signal or the
residual signal. However, previous results from the cepstrum and frequency domain methods
indicated large sidebands and a significant problem with this high speed shaft pinion; these were
also confirmed in the failure report study. Large sidebands are noticed in the frequency domain
representation of the time synchronous average signal for the high speed shaft, which is provided
in Figure 6.13. The gear mesh frequency peak (order 22) is lower in magnitude then a peak at
one of the sidebands (order 23). The inability for the residual signal to detect this fault on the
high speed shaft pinion highlights the importance of extracting multiple gear vibration features to
have better coverage for the different failure modes.
71
Time Synchronous Signal HSS Pinion
4
Vibration (m/s 2)
2
-2
-4
0 50 100 150 200 250 300 350 400
High Speed Shaft Angle (degree)
Residual Signal HSS Pinion
4
Kurtosis Value of 2.3433
Vibration (m/s 2)
-2
-4
0 50 100 150 200 250 300 350 400
High Speed Shaft Angle (degree)
Figure 6.12. TSA signal and residual signal from accelerometer AN7 - Case C: top plot - TSA signal
for high speed shaft pinion; bottom plot - residual signal for high speed shaft pinion
1.2
0.8
Vibration (m/s 2)
0.6
X= 23
Y= 0.35919
0.4
0.2
0
0 50 100 150
Orders of High Speed Shaft
Figure 6.13. TSA vibration spectrum for accelerometer AN7 and high speed shaft Case C
Another example residual signal is shown for the ring gear in Figure 6.14. The kurtosis value of
the ring gear is also in a normal range (3.36), and there appears to be no abnormal patterns or an
72
indication of a fault in the time synchronous average signal or the residual signal for the ring
gear. This is in sharp contrast to the results from the phase modulation function provided in the
subsequent section, in which there is a clear indication of damage on the ring gear. The residual
signal for the other parallel shaft gears also offers no indication of damage. This highlights that
the residual signal was not the most appropriate algorithm for detecting the failure modes that
were occurring on the parallel shaft gears and the ring gear.
-0.05
-0.1
-0.15
0 50 100 150 200 250 300 350 400
Carrier Angle (degree)
Residual Signal Ring Gear
0.1
Kurtosis Value of 3.3604
Vibration (m/s 2)
0.05
-0.05
-0.1
0 50 100 150 200 250 300 350 400
Carrier Angle (degree)
Figure 6.14. TSA signal and residual signal from accelerometer AN3 - Case C: top plot - TSA signal
for ring gear; bottom plot - residual signal for ring gear
73
the high speed shaft pinion. As one can observe, there are significant jumps in the phase
modulation function for this gear and a high kurtosis value at 12.8. This would be an indication
of a damage on the high speed shaft pinion. In addition, the amplitude modulation function is
close to zero when these significant changes in phase occur. The phase modulation function for
the high speed shaft gear had a moderate indication of a gear problem, with a kurtosis value of
5.2. However, there was no indication of a problem for the intermediate speed shaft gear or
pinion. Another example is provided in Figure 6.16, in which the amplitude and phase
modulation for the ring gear is provided. There appears to be a clear indication of a problem
with the ring gear through visual observation of the amplitude and phase modulation signals.
The phase modulation function, in particular, has a high kurtosis value and two noticeable shifts
in phase, which indicate a damaged gear. In summary, the amplitude and phase modulation
functions provide a strong indication of a problem with the ring gear and high speed shaft pinion,
and a moderate indication for the high speed shaft gear; however, there was no indication of a
problem for the other gear wheels.
-5
0 50 100 150 200 250 300 350 400
High Speed Shaft Angle (degree)
AM and PM Modulation Signal HSS Pinion
1.5
Envleope (AM) (m/s2)
0.5
0
0 50 100 150 200 250 300 350 400
High Speed Shaft Angle (degree)
Phase Modulation (Rad)
10
-10
-20
0 50 100 150 200 250 300 350 400
High Speed Shaft Angle (degree)
Figure 6.15. High speed pinion amplitude and phase modulation signal from accelerometer AN7 -
Case C: top plot - Time Synchronous Average; middle plot - amplitude modulation signal;
bottom plot - phase modulation signal
74
Time Synchronous Signal Ring Gear
0.1
Vibration (m/s 2)
Kurtosis Value of 3.3138
0
-0.1
0 50 100 150 200 250 300 350 400
Carrier Angle (degree)
AM and PM Modulation Signal Ring Gear
Envleope (AM) (m/s2)
0.01
0
0 50 100 150 200 250 300 350 400
Carrier Angle (degree)
Phase Modulation (Rad)
5
Kurtosis Value of 33.0163
0
-5
-10
0 50 100 150 200 250 300 350 400
Carrier Angle (degree)
Figure 6.16. Ring gear amplitude and phase modulation signal from accelerometer AN3 - Case C:
top plot TSA; middle plot - amplitude modulation signal; bottom plot - phase modulation signal
75
of rotations is equal to the reset time for the planet or sun gear. The result is an assembled
vibration signal for each gear tooth. This process is repeated until several assembled signals can
be constructed. Lastly, the constructed waveforms are ensemble averaged and this completes the
process for extracting the time synchronous average
1. Vibration Signal
76
Figure 6.18. Narrow band amplitude modulation signal for determining planet passing - Case C
0.9
0.8
0.7
Window Magnitude
0.6
0.5
0.4 Tv=NvTm
0.3
0.2
0.1
0
0 0.5 1 1.5 2 2.5 3 3.5 4
Tooth Mesh Period
Figure 6.19. Example Tukey window used for planet separation algorithm - in this study,
Nv was set to 3 to include 3 mesh periods
Sample results from the synchronous averaging signal and the residual signal are provided in
Figure 6.20, in which the result is shown for one of the planet gears. In this example, the time
synchronous average signal and the residual signal show no abnormal behavior for the planet
gear; the failure report found no defects on any of the planet gears. The spectrum for the time
77
synchronous average signal for this planet gear is shown in Figure 6.21, and one can observe a
clear gear mesh frequency peak at the 39th order. The sidebands are not large in magnitude in the
spectrum and the planet gear appears to be in a health state from these data processing results.
The other two planet gears were also considered to be in a normal condition based on similar
results that were observed in their time synchronous average signal and residual signal.
0.1
-0.1
-0.2
0 5 10 15 20 25 30 35 40
Planet Tooth
Residual Signal Planet 2
0.1
Kurtosis Value of 3.287
Vibration (m/s 2)
0.05
-0.05
-0.1
0 5 10 15 20 25 30 35 40
Planet Tooth
Figure 6.20. Top - TSA signal for Planet 2; bottom - residual signal for Planet 2 Case C
X= 39
Y= 0.028945
0.025
0.02
Vibration (m/s 2)
0.015
0.01
0.005
0
0 50 100 150
Orders for Planet 2
78
The time synchronous average signal and the residual signal analysis were also performed for the
sun gear, with the results provided in Figure 6.22. In this example, the residual signal is
providing a moderate indication of damage on the sun gear, with a kurtosis value of 5.36. The
failure report confirms scuffing and fretting corrosion on the sun pinion. Additional narrowband
amplitude modulation and phase modulation analysis was also performed on the time
synchronous average signals for the planet gears and the sun gear. Although the residual signal
provided an indication of damage on the sun gear, the amplitude and phase modulation signals
did not provide any indication of damage on the sun gear. A sample result from one of the planet
gears is provided in Figure 6.23. Both the amplitude and phase modulation signals do not
indicate damage on this particular planet gear. The other two planet gears also did not have any
indication of damage from the narrowband amplitude and phase modulation analysis; this is
encouraging since the failure report did not find any damage on any of the three planet gears.
For the sun gear, the residual signal provided a moderate indication of damage, but no damage
was indicated from the amplitude and phase modulation signals.
-0.05
-0.1
-0.15
0 2 4 6 8 10 12 14 16 18 20
Planet Tooth
Residual Signal Sun Gear
0.1
Kurtosis Value of 5.3562
Vibration (m/s 2)
0.05
-0.05
-0.1
0 2 4 6 8 10 12 14 16 18 20
Planet Tooth
Figure 6.22. Top - TSA signal for sun gear; bottom - residual signal for sun gear Case C
79
Time Synchronous Signal Planet Gear
0.2
Vibration (m/s 2)
Kurtosis Value of 3.3138
-0.2
0 5 10 15 20 25 30 35 40
Planet 2 Tooth
AM and PM Modulation Signal Planet2 Gear
Envleope (AM) (m/s2)
0.04
0.03
-2
-3
Kurtosis Value of 2.1986
-4
0 5 10 15 20 25 30 35 40
Planet2 Tooth
Figure 6.23. Planet Gear 2 amplitude and phase modulation signal from accelerometer AN3
Case C: top plot TSA; middle plot - amplitude modulation signal;
bottom plot - phase modulation signal
From the tabular results in Table 6.3, one can observe that the high speed pinion had high
indications of damage from several techniques, including the vibration spectrum, cepstrum
processing, and narrowband analysis from the phase modulation signal. Only the residual signal
80
did not provide an indication of a fault for the high speed pinion. The residual signal in general
did not detect damage on the gear wheels for this study, with only a moderate indication of
damage on the sun gear. The intermediate speed shaft pinion had a moderate indication of
damage from both frequency domain analysis and the cepstrum features; however, there was no
indication of damage by the narrow band analysis. Bearing envelope analysis provided high
confident indications for three of the bearing failures that were on the intermediate and high
speed shafts. The bearing damage on the planet carrier upwind bearing had only a moderate
detection; the peak was less noticeable and a different band pass filter range had to be used to
detect this fault. For the planetary gearbox, the ring gear appeared to be the easiest to detect,
with a high confidence indication of damage from the narrowband phase modulation signal. It
should be noted that none of the planet gears had damage according to the failure report; this
agrees with the results from the data processing, in which none of the algorithms detected any
abnormality or damage with the planet gears. In summary, many of the failed components could
be detected using the evaluated algorithms. However, for gear components, in particular,
multiple algorithms appear necessary since many of the algorithms are only tuned to one failure
mode.
Table 6.3. Summary of results for each algorithm with the following notation: L-low confidence, M-medium
confidence, H-high confidence, NA not applicable or evaluated; black - indicates a method that was
evaluated before the failure report, blue - indicates a method that was evaluated after the failure report
TSA
Bearing TSA
Frequency Spectral Amplitude /
Failed Component Cepstrum Envelope Residua
Domain Kurtosis Phase
Analysis l Signal
Modulation
HSS Pinion H H NA NA L H
HSS Gear L L NA NA L M
ISS Pinion M M NA NA L L
ISS Gear L M NA NA L L
Ring Gear NA NA H - stage NA L H
Sun Pinion NA NA H - stage NA M L
ISS Upwind
NA NA NA H NA NA
Bearing
ISS Downwind
NA NA NA H NA NA
Bearings
HSS Downwind
NA NA NA H NA NA
Bearings
Planet Carrier
NA NA NA M NA NA
Upwind Bearing
81
6.4 Conclusions and Future Work
This chapter provides an evaluation of vibration signal processing and feature extraction
algorithms used by the research team at the Center for Intelligent Maintenance Systems (IMS)
for the Condition Monitoring Round Robin study. As one can observe from the summary results
table, most of the failed components could be detected by one or more of the processing
methods. This is encouraging and highlights that vibration-based condition monitoring can be
used to assess and diagnose which components are in a failed condition. For gear wheels in
particular, the use of multiple algorithms appears necessary, considering the different number of
gear failure modes that are possible. The residual signal, in particular, did not seem suited for
the failure modes exhibited by the damaged gear wheels on the parallel stage shaft, while the
other algorithms provided more confident detections. The detection results for each algorithm do
not account for the additional hardware or on-board processing requirements. The planet
separation algorithm, in particular, is difficult to implement given the time period required to
accumulate enough rotations of the carrier shaft.
Although the results from this study were encouraging, there are some aspects that could have
aided the study or should be considered for future work. This study could have been aided by
time domain waveforms instead of frequency spectrums provided for the baseline data. Many of
the more advanced algorithms require the raw time waveform and they could not be evaluated
based on the baseline data set. In addition, alternative methods were used for acquiring a
reference signal for performing synchronous averaging. Further experimental studies should
acquire and save a tachometer pulse train to avoid this issue. Considering that the gearbox was
already in a severely damaged condition, the algorithms were evaluated on the basis of their
ability to detect the health state of the various bearing and gear components. Unfortunately, this
does not allow one to evaluate the algorithms ability to provide an early detection of a problem
or whether the extracted vibration features are monotonic with the damage level. Both early
detection and severity estimation are additional aspects worth evaluating for vibration-based
condition monitoring techniques for wind turbine drivetrains. Continuous monitoring of a wind
turbine drivetrain from a baseline condition until failure could provide a way to further evaluate
the merits of the vibration-based condition monitoring algorithms.
82
7 Defect Diagnosis in Wind Turbine Gearbox based on Sideband
Energy and Enveloping Spectral Analysis
Robert X. Gao*1, Jinjiang Wang1, and Ruqiang Yan2
1
Department of Mechanical Engineering, University of Connecticut, Storrs, CT, USA
2
School of Instrument Science and Engineering, Southeast University, Nanjing, China
*
Corresponding Author Email: [email protected]
7.1 Introduction
To improve cost-effectiveness of wind energy, wind turbines must operate in a highly reliable
fashion, given the significant cost associated with system repair, maintenance, and unexpected
failure. Accordingly, techniques for the condition monitoring and fault diagnosis of wind turbine
structures and components have been gaining increasing attention [41].
Of the various components in a wind turbine, the gearbox is a major component that is costly and
vulnerable to failure. Accordingly, signal processing for gearbox defect identification and
diagnosis has been an active research area. There are two major components in a gearbox: gears
and bearings. For gear diagnosis, sideband analysis, amplitude, and phase modulation [42],
wavelet transform [43], and spectral kurtosis [44] have been investigated. Typically, diagnosis is
achieved through comparisons between a defective gear and a healthy gear. As for the bearing
diagnosis, a band pass resonant signal processing technique has been reported. Choosing an
appropriate bandwidth remains an important issue, given its effect on the diagnosis result.
As part of the CM Round Robin study, three sets of vibration data measured by NREL on a wind
turbine gearbox (of unknown damaged condition) were analyzed at three different operating
speeds. For gear fault diagnosis, sideband pattern analysis was performed on all gears. Data from
torque measurements were also analyzed to facilitate annulus gear diagnosis; whereas for bearing
diagnosis, the multi-scale enveloping spectra technique [45] has been investigated. The result of
the analysis was compared with that of a spectral analysis of a healthy gearbox that was provided
by NREL as a reference base.
7.2 Algorithms
Vibration of a gearbox can be caused by various sources, such as gear meshing, interaction
between the rolling elements and raceways in bearings, and shaft rotation. Structural defects on
gears surface also constitute a source of vibrations that are carried by gear meshing frequencies.
Research reported in the literature has shown that the energy associated with frequency
components at the sidebands around the gear meshing frequency will increase as the health
condition of the gear deteriorates, in comparison to a healthy gear. Accordingly, sideband
analysis has been performed for gear analysis in this study.
For bearing defect diagnosis, the multi-scale enveloping spectra technique has been investigated
[45], which makes use of the time, scale, and frequency information contained in the bearing
vibration signal. The algorithm first decomposes the bearing vibration signal into a series of
wavelet basis functions, through variations of the scales and time shifts of the wavelet function.
The envelope of each decomposed wavelet function is then extracted from the modulus of the
wavelet coefficients. Next, spectral analysis is performed repeatedly on the envelope signal,
resulting in an envelope spectrum of the original signal at the various scales. The integration of
83
the wavelet transform, using post-spectral analysis, reveals the defect characteristic more clearly,
enhancing its effectiveness in bearing defect diagnosis.
7.3 Results
Vibration data measured by NREL on a wind turbine gearbox was analyzed for gear and bearing
defect diagnosis, respectively. Figure 7.1 illustrates the physical system analyzed for this study,
with the locations of the specific bearings, shaft, and gears identified.
84
AN7_New
AN7 AN7_Defective
AN7__HSSRadial
0.07 0.4
(a) (b)
0.06 0.35
Amplitude (mW)
0.04
fINT_Mesh 0.2
0.03
0.15
fHS_Mesh
0.02
0.1
0.01 0.05
0 0
0 200 400 600 800 1000 1200 1400 1600 1800 2000 0 200 400 600 800 1000 1200 1400 1600 1800 2000
Frequency (Hz) Frequency (Hz)
AN7_New
AN7 AN7_Defective
AN7__HSSRadial
0.02 0.14
(c) (d)
0.018
fINT_Mesh fINT_Mesh+f2
0.016
fINT_Mesh 0.12
0.1
0.014
Amplitude (mW)
Amplitude (mW)
0.012
0.08
0.01
fINT_Mesh+f2 0.06
0.008
0 0
120 130 140 150 160 170 180 190 200 210 220 120 130 140 150 160 170 180 190 200 210 220
Frequency (Hz) Frequency (Hz)
AN7_New
AN7 AN7_Defective
AN7__HSSRadial
0.06 0.08
(e) fHS_Mesh (f) fHS_Mesh
fHS_Mesh 0.07
fHS_Mesh-f1
0.05
fHS_Mesh+f2
0.06
fHS_Mesh-f2
0.04
0.05 fHS_Mesh+f1
Amplitude (mW)
Amplitude (mW)
0.03
fHS_Mesh+f2
0.04
fHS_Mesh-f2
fHS_Mesh+f2 0.03
0.02
0 0
600 650 700 750 600 650 700 750
Frequency (Hz) Frequency (Hz)
Figure 7.2. Comparison analysis between test data and reference data for HS_Pinion and
INT_Pinion
85
AN7_New
AN7 AN7_Defective
AN7__HSSRadial
0.04 0.25
(a) (b)
0.035
fAnnulus_Mesh
0.025
Amplitude (mW)
Amplitude (mW)
0.15
0.02 2*fAnnulus_Mesh
0.1
0.015
0.01
fSun_Mesh
0.05
0.005
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Frequency (Hz) Frequency (Hz)
AN7_New
AN7 AN7_Defective
AN7__HSSRadial
0.04 0.14
(c) (d)
0.035 fAnnulus_Mesh 0.12 fAnnulus_Mesh
0.03
0.1
0.025
fAnnulus_Mesh-f4
Amplitude (mW)
Amplitude (mW)
0.08
0.02
0.06
0.015
fAnnulus_Mesh+f4
0.04
0.01
0.005 0.02
0 0
15 20 25 30 35 40 45 50 55 15 20 25 30 35 40 45 50 55
Frequency (Hz) Frequency (Hz)
AN7_New
AN7 AN7_Defective
AN7__HSSRadial
0.04 0.04
(e) (f)
0.035 0.035
0.03 0.03
0.025 0.025
Amplitude (mW)
Amplitude (mW)
fSun_Mesh
0.02 0.02 fSun_Mesh-f3
0.015 0.015
fSun_Mesh+f3
0.01 0.01
0.005 0.005
0 0
32 34 36 38 40 42 44 46 48 50 52 32 34 36 38 40 42 44 46 48 50 52
Frequency (Hz) Frequency (Hz)
Figure 7.3. Comparison analysis between test data and reference data for Annulus_Gear and
Sun_Gear
86
Table 7.1. Sideband energy comparison between new gearbox and gearbox at the end of service life
Components Healthy Gearbox Gearbox at End of Service Life
Meshing 1st sideband Sideband Meshing 1st sideband Sideband
energy energy (mw) Energy ratio energy energy (mw) Energy ratio
(mw) (mw)
HS_Pinion 0.051 0.016 31.4% 0.051 0.114 223.5%
INT_Pinion 0.017 0.005 29.4% 0.13 0.14 107.7%
Sun_Gear 0.0002 0.0004 200% 0.008 0.019 237.5%
Annulus_Gear 0.035 0.014 40% 0.135 0.059 43.7%
The increase in the sideband energy ratio related to the annulus gear is not as significant as that
of the other three gears (relative increase is 9.3%). Considering that the annulus gear runs at low
speed under high torque conditions resulting from the gearbox transmission mechanism, the
torque measurement may be more effective for low speed gear diagnosis, due to its sensitivity to
angular vibrations of the gear. Based on this consideration, data obtained from torque
measurements of the gearbox have been analyzed. Figure 7.4 shows the waveform of the torque
data under 1,200 RPM. The interval between the peaks is approximately 4.03s, corresponding to
the rolling-over period of the annulus gear. Figure 7.5 shows the result of envelope spectrum
analysis of torque data. A peak at frequency of 0.248 Hz, which corresponds to the roll-over
period of the annulus gear, is identified. This indicates a structural defect on the annulus gear.
125
1200RPM
120
115
Amplitude
110
105
100
95
0 10 20 30 40 50 60
Time (Sec)
Figure 7.4. Time series of torque data under 1200 rpm
87
Enveloping Spectrum (mW)
1
0.248 Hz
0.5
0.496 Hz
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Frequency (Hz)
Figure 7.5. The envelope spectrum of torque data under 1200 rpm
3000
HSS_Shaft
Bearing H 3*fu
2500 fBPFO
2000
Spectrum (mW)
HSS_Shaft HSS_Shaft
1500 imbalance fu 2*fu
1000
500
0
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hz)
Figure 7.6. Wavelet enveloping spectrum of sensor AN3 at 1,800 rpm
88
Figure 7.7 shows the wavelet enveloping spectrum of data measured by sensor AN6 (adjacent to
bearing D) at the running speed of 1,800 rpm. Defect frequency, fBPFI, of bearing D can be
identified, although its amplitude is not as significant due to masking from the high energy
component, fu, which is related to imbalance of the HSS_Shaft.
HSS_Shaft imbalance fu
Bearing D fBPFI
Table 7.2. Comparison between the analysis result and the actual damage of a tested gearbox
Damage Component Damage mode Severity Rationale
1 HS gear set Scuffing Severe Sideband
2 Intermediate gear set Fretting corrosion and scuffing Severe Sideband
3 Annulus gear Scuffing/polishing Moderate Sideband
4 Sun pinion Fretting corrosion Severe Sideband
5 Bearing H Fretting corrosion Severe Wavelet envelope
6 Bearing D Assembly damage Moderate Wavelet envelope
7 Bearing C1/C2 Assembly damage on spacer Severe Not identified
8 Bearing A1 Overheating Severe Not identified
89
7.4 Lessons Learned
The diagnosis results have demonstrated that sideband analysis provides an effective and
computationally efficient approach to gear defect diagnosis. Considering that the energy content
associated with a structural defect may not be significant at the defect incipient stage,
complementing sideband analysis with other advanced techniques should be considered.
As for the bearing diagnosis, defects in bearing H and D have been identified by means of
wavelet enveloping. Because of the low signal-to-noise ratio and interference caused by gear
meshing frequencies, diagnosis of bearing D was more challenging than that of bearing H. This
indicates the need for separating the signal related to gear meshing from that of the bearing
vibration before performing the bearing diagnosis. Because the nature of the damage to Bearing
A1 is related to overheating, instead of a surface defect, and damage on bearing C1/C2 is not on
the roller raceway, but on the spacer, defects from bearings C1/C2 and A1 could not be identified
by the vibration data analysis conducted herein. This leads to the consideration that, besides
vibration sensing, other sensing techniques (e.g., temperature sensing) should be considered as
well to enable fusion of diverse sensing modalities for improved gearbox diagnosis. Research is
needed to address this issue.
90
8 Fault Analysis of a Wind Turbine Gearbox: A Data Driven
Approach
Zijun Zhang*, Anoop Verma, Andrew Kusiak
Department of Mechanical and Industrial Engineering, The University of Iowa
*
Corresponding Author Email: [email protected]
8.1 Introduction
The wind industry has been affected by failures of wind turbine components, such as main
bearings, gearboxes, and generators. The high cost of replacing failed components impacts the
energy cost. Therefore, research in fault identification and condition monitoring is warranted.
Fault identification is concerned with a fault that has occurred and its labeling. In condition
monitoring, parameters reflecting the component conditions are identified and their changes are
analyzed to detect an emerging failure. In this chapter, the fault identification analysis is studied
in the time domain based on the vibration data of an impaired gearbox tested by NREL, which is
different from the traditional fault analysis from the frequency domain [ 46- 51].
8.2 Methodologies
This section describes the data processing and analysis methods applied to the detection of the
gearbox faults.
RMS is the simplest method for measuring abnormalities in the time domain. The RMS value
can be used to detect unbalanced rotating elements. It is a statistical measure of the magnitude
with varying quantity, and it is expressed in Equation (15).
1 N
(15)
(s )
2
RMS = i
N i =1
The crest factor (CF) is a measure used to detecting changes in the signal pattern due to
impulsive vibration sources, such as tooth breakage. It can be useful in detecting high peaks in
the signal at higher magnitudes of the peak and for smaller numbers of peaks. A small value of
91
RMS and high peak value implies a higher crest factor. A CF with values in the range 2-6
represents normal operations; whereas, a value higher than six represents a defective component.
The crest factor is computed by dividing the peak level of the signal average by the standard
deviation (RMS) of the signal average, as shown in Equation (16).
CF =
Peak level (16)
RMS
Kurtosis is defined as the fourth statistical moment of an array of values about the mean. It
indicates the existence of major peaks. A kurtosis value of less than three represents a component
in a normal health condition; whereas, a value greater than three represents abnormality. The
greater the number of peaks in the signal, the larger is the kurtosis. The kurtosis is expressed in
Equation (17).
N
(17)
N . ( si s )
4
K= i =1
2
N
2
( si s )
i =1
n xy ( x )( y ) (18)
r=
n( x ) ( x )
2 2
n( y ) ( y )
2 2
where r is the correlation coefficient, and x and y are two different parameters.
8.2.4 Clustering
Clustering analysis is an unsupervised method of data analysis. Clustering algorithms group
observations into clusters by evaluating similarities among the observed data. A k-means
algorithm [ 52] is modified in this study to establish clusters. In the original version of the k-
means algorithm, the number of clusters, k, should be arbitrarily set by the analyst. In this study,
a clustering cost function is introduced to evaluate the cluster quality with k. The clustering cost
function is formulated as Equation (19) and used in a 10-fold, cross-validation scheme [ 53, 54].
1 k 2 (19)
d ( k , x, c )
= k x j ci
m
i =1
=i 1
i
x j Ci
where d is the clustering cost, k is the number of clusters, m is the number of observations
(sensors) contained in each cluster, x is a vector of parameters used in this research, c presents
the centroid of each cluster, j is the index of each data point, and Ci represents cluster i.
92
Step 1. Set the initial value of k to 2
Step 3.1. Randomly select nine subsets for training and use the 10th subset for testing.
Step 3.3. Repeat the following two steps until the centroids do not change.
ci = x j /n
Step 3.3.2. Update the values of the centroids by x j Ci
, where n is the total number of
observations.
Step 5. Stop the algorithm if d(k,x,c) d(k 1,x,c) or k = 12; otherwise, go back to Step 1.
8.3 Results
8.3.1 Data Process and Description
Acceleration data are sampled at 40,000 Hz and recorded for 10 minute intervals. The data set is
large. As the sensors are used to recording acceleration, the data is transformed based on (1) to
obtain the jerk values. Acceleration data for all three test cases (2a, 2b, and 2c) at 10 minute
intervals are transformed. Each data set is divided into 40 data subsets of equal size (or equal
length, 15 seconds) for the further investigation.
Since the sampling frequency of the acceleration data is high, 40,000 Hz, viewing and data
analysis of the high frequency data directly in the run-chart form is not feasible. Therefore, four
statistical metrics, the mean, standard deviation, maximum, and minimum, are utilized to
compute the Jerk value for each data subset discussed in Section 8.2.1. The values of the mean,
standard deviation, maximum, and the jerk for all 40 data subsets are used to develop three new
data sets for analysis discussed in Section 8.3.2, 8.3.3, and 8.3.4. The minimum Jerk value is
always zero and, therefore, it is excluded from this research.
93
maximum and minimum of the maximum jerk data of each accelerometer are estimated. Then,
the ratio, R, is estimated based on the maximum and minimum values of the maximal jerk based
on Equation (20).
R = (Max{maximum jerk} Min{maximum jerk}) / Max{maximum jerk} (20)
Figure 8.2 shows the R for each accelerometer. As shown, the Ratio 1 of sensors AN3 and AN4
is much higher than the R of other sensors. Therefore, the location and component that AN3 and
AN4 monitored are considered as the possible location where the fault occurred. From the
specification provided by NREL, the component monitored by AN3 and AN4 is a ring gear and
the location is in the low speed stage (LSS-T) of the gearbox.
600
Maximal rate of speed (rpm/15-s)
500
400
300
200
100
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
Time (15 s intervals)
1
0.9
0.8
0.7
Percentage
0.6
0.5
0.4
0.3
0.2
0.1
0
AN1 AN2 AN3 AN4 AN5 AN6 AN7 AN8 AN9 AN10 AN11 AN12
Index of sensors
Ratio 1
In this section, a correlation coefficient analysis is performed based on a data set containing the
mean jerk described in Section 8.3.1. Table 8.1 and Table 8.2 present the results of correlation
coefficient analysis based on cases 2b and 2c. As shown in Table 8.1 and Table 8.2, sensors,
94
AN5, AN6, AN7, AN8, AN9, AN11, and AN12, are highly correlated. Since the component
monitored by AN1 and AN2 is considered healthy, results in Table 8.1 and Table 8.2 indicate
that significant vibration originates at one of the areas AN5, AN6, AN7, AN8, and AN9
monitored (HS-ST and IS-ST), and it impacts the generator monitored by AN11 and AN12.
Table 8.1. Correlation coefficient analysis of the mean of jerk data: Case 2b
Sensor AN1 AN2 AN3 AN4 AN5 AN6 AN7 AN8 AN9 AN10 AN11 AN12
AN1 1.00 0.98 0.53 0.20 0.02 0.14 0.00 -0.11 0.16 0.16 0.16 0.03
AN2 0.98 1.00 0.59 0.06 0.18 0.30 0.17 0.05 0.32 0.25 0.33 0.19
AN3 0.53 0.59 1.00 0.09 0.31 0.53 0.41 0.19 0.62 0.77 0.48 0.32
AN4 0.20 0.06 0.09 1.00 -0.88 -0.78 -0.85 -0.94 -0.70 -0.11 -0.76 -0.89
AN5 0.02 0.18 0.31 -0.88 1.00 0.94 0.97 0.97 0.90 0.50 0.92 0.97
AN6 0.14 0.30 0.53 -0.78 0.94 1.00 0.98 0.93 0.99 0.60 0.95 0.96
AN7 0.00 0.17 0.41 -0.85 0.97 0.98 1.00 0.97 0.96 0.56 0.94 0.97
AN8 -0.11 0.05 0.19 -0.94 0.97 0.93 0.97 1.00 0.88 0.39 0.90 0.98
AN9 0.16 0.32 0.62 -0.70 0.90 0.99 0.96 0.88 1.00 0.67 0.92 0.93
AN10 0.16 0.25 0.77 -0.11 0.50 0.60 0.56 0.39 0.67 1.00 0.59 0.42
AN11 0.16 0.33 0.48 -0.76 0.92 0.95 0.94 0.90 0.92 0.59 1.00 0.92
AN12 0.03 0.19 0.32 -0.89 0.97 0.96 0.97 0.98 0.93 0.42 0.92 1.00
Table 8.2. Correlation coefficient analysis of the mean of jerk data: Case 2c
Sensor AN1 AN2 AN3 AN4 AN5 AN6 AN7 AN8 AN9 AN10 AN11 AN12
AN1 1.00 0.83 0.32 -0.73 0.50 0.56 0.49 0.67 0.49 0.48 0.64 0.56
AN2 0.83 1.00 0.76 -0.91 0.88 0.90 0.87 0.92 0.87 0.86 0.91 0.91
AN3 0.32 0.76 1.00 -0.67 0.94 0.92 0.95 0.78 0.96 0.95 0.80 0.90
AN4 -0.73 -0.91 -0.67 1.00 -0.85 -0.90 -0.84 -0.97 -0.83 -0.80 -0.94 -0.91
AN5 0.50 0.88 0.94 -0.85 1.00 0.99 0.99 0.92 0.99 0.98 0.92 0.98
AN6 0.56 0.90 0.92 -0.90 0.99 1.00 0.99 0.95 0.98 0.97 0.95 0.99
AN7 0.49 0.87 0.95 -0.84 0.99 0.99 1.00 0.91 0.99 0.98 0.92 0.98
AN8 0.67 0.92 0.78 -0.97 0.92 0.95 0.91 1.00 0.91 0.88 0.96 0.96
AN9 0.49 0.87 0.96 -0.83 0.99 0.98 0.99 0.91 1.00 0.98 0.90 0.98
AN10 0.48 0.86 0.95 -0.80 0.98 0.97 0.98 0.88 0.98 1.00 0.88 0.96
AN11 0.64 0.91 0.80 -0.94 0.92 0.95 0.92 0.96 0.90 0.88 1.00 0.96
AN12 0.56 0.91 0.90 -0.91 0.98 0.99 0.98 0.96 0.98 0.96 0.96 1.00
Besides correlation coefficient analysis, the k-means clustering algorithm is utilized to examine
the relationship among sensors. Since the k-means algorithm groups parameters into clusters by
examining their similarity, it is capable of evaluating the relationship among sensors. Table 8.3
shows the clustering result of both case 2b and 2c. In Table 8.3, AN6, AN7, AN8, AN9 and
AN12 are grouped into one cluster while other sensors are grouped into another cluster. This
95
result points to the same pattern as the correlation coefficient analysis, which indicates the faulty
location of HS-ST and IS-ST.
Case 2b Case 2c
Index of AN Final classification Index of AN Final classification
1 2 1 2
2 2 2 2
3 2 3 2
4 2 4 2
5 2 5 2
6 1 6 1
7 1 7 1
8 1 8 1
9 1 9 1
10 2 10 2
11 2 11 2
12 1 12 1
The results of RMS, CF, and kurtosis also show an agreement with the correlation and clustering
analysis. Figure 8.3 presents the RMS values of 2b, averaged over a 1 minute interval. The case
2b reveals a pattern; whereas, the low RMS values of sensors AN1 and AN2 cause a high crest
factor. The increase in RMS of AN8 may indicate that a fault is in progress in the gearbox. It
could be due to the oil leakage.
Figure 8.4 displays the crest factor of 2b across 12 sensors. In Figure 8.4, it can be observed that
the main bearing is affected in test case 2b. It can be assumed that the significant amount of crest
factor near AN1 is contributed by the vibrations of other components.
96
Figure 8.4. Crest factor across 12 sensors - Case 2b
Figure 8.5 represents the kurtosis of three test cases across 12 sensors. In case of CM_2b,
kurtosis across all 12 sensors shows increasing patterns. This could indicate a gradual wear.
97
9 Techniques for Separation and Enhancement of Various
Components in the Analysis of Wind Turbine Vibration Signals
Nader Sawalhi1, Robert B. Randall2*, and David Forrester3
1
College of Engineering, Prince Mohammad Bin Fahd University, The Kingdom of Saudi
Arabia.
2
School of Mechanical and Manufacturing Engineering, The University of New South Wales,
Australia
3
Defence Science and Technology Organisation, Australia
*
Corresponding Author Email: [email protected]
9.1 Introduction
This chapter provides a summary of the handling and processing of wind turbine data provided
by NREL for the CM Round Robin study. The University of New South Wales took an active
role in this study through Dr. Nader Sawalhi and Professor Bob Randall.
NREL provided three sets of data (taken from a number of accelerometers on the planetary
gearbox) at different speeds and load conditions. The data was for a faulty condition only. The
data was first analyzed blindly, in the sense that there was originally no information given as to
the type and location of faults in the gearbox. Our group concentrated on the search for bearing
faults, because it is our opinion that to be sure of detecting gear faults, it is necessary to make
comparisons with signals from the gears in healthy condition. We did, however, look for
indications of local faults on the gears, as these might show up clearly as local impulsive
responses in the gear signatures.
Later, after receiving the inspection report, and spectra from the gearbox in good condition, we
made a further analysis, in particular of the gear signals, and were able to detect other indicators
of the actual faults. Most of these could have been detected in the original blind analysis, if we
had had the signals for good condition.
Our group had already had a certain amount of experience in analyzing signals from wind
turbine transmissions, and was aware that the main differences, with respect to other similar
gearboxes, were because the load can vary considerably over relatively short periods, at least
with respect to the low speed input sections of the transmission. The vibration signal from gears
is affected greatly by the load, and so some means has to be found to distinguish such variations
from changes in condition. We have considerable experience with the diagnoses of helicopter
gearboxes, which are somewhat similar. However, these operate at perhaps ten times higher
speeds, and it is possible to obtain reasonably long signals with an approximately constant load.
In contrast to gear fault signals, bearing signals are not so sensitive to torque load (although in
gearboxes, radial load depends on torque load), and there is usually a dramatic difference in the
signals in the presence of faults, which often allows them to be diagnosed without necessarily
having access to historical data. This is because of the development of spectral kurtosis (SK)
techniques in recent years by our group and colleagues, in particular Professor Jerome Antoni,
now of INSA Lyon, in France [55]. The techniques we have used for analyzing the signals for
gear and bearing faults are described in more detail below. Those bearing diagnosis are primarily
based on a semi-automated procedure, with several different stages to separate and enhance the
98
bearing signals. Then envelope analysis is applied (spectrum analysis of squared envelope
signals) to diagnose the fault repetition frequencies and their modulations by lower frequencies
[56]. We also applied a recently developed cepstral pre-whitening technique, which can
circumvent some of the stages in the earlier procedure [57]. For gear diagnosis we applied
classic techniques based on obtaining a synchronously averaged signature for each gear, and then
looking for localized impulses characteristic of local faults. This requires the signals to be order
tracked, or re-sampled in the angular domain, with equal numbers of samples in each
revolution. This normally requires a tachometer or shaft encoder signal for synchronization. The
supplied speed signal was not suitable for this so we extracted a pseudo-encoder signal from
the vibration signal to use for order tracking. For obtaining signatures of the individual planet
gears and sun gear in the planetary part of the gearbox, the premium current method is one
patented by DSTO (Defence Science and Technology Organisation), of the Australian Defence
Department. We engaged Dr. David Forrester of DSTO, the inventor of the technique, to obtain
these signatures for us for the latest results presented at the Wind Turbine Condition Monitoring
Workshop in Broomfield, CO, in September 2011. Those results are included here. It should be
mentioned that Dr. Forrester was surprised by the design of the planetary section of the gearbox,
as the choice of tooth numbers was far from a hunting tooth design, normally considered good
practice, and as a result of this, groups of teeth always mesh in the same way and repeat
frequently. The effects of this are discussed below. It is also somewhat unusual that the ratio of
the high speed section was exactly 4:1 (88:22) meaning that the 22 teeth on the pinion always
mesh in exactly the same way with four groups of 22 teeth on the intermediate shaft wheel.
Therefore, a fault on one tooth transfers to individual teeth on the mating gear, and is not
smeared out as it is in a hunting tooth design.
After receiving spectra for the gearbox in good condition, we were also able to make spectrum
comparisons to detect changes in modulation sideband patterns, often indicative of faults, and
also cepstrum analysis to concentrate the information in the sideband patterns.
9.2 Algorithms
Our general approach is to separate the signals, into the components coming from the gears and
bearings, and analyze them separately. The separation is based on the assumption that the gear
signals are deterministic (with respect to rotation angle), and the bearing signals are stochastic,
because of the minor random slip between the components and the random positioning of the
rolling elements in the clearance of the cage. These two effects give an approximately 1-2%
deviation of the mean value of the actual bearing fault frequencies, with the same order of
random variation around the mean from those frequencies and calculated on the basis of no slip
and perfectly uniform spacing. The signals can then be classified as approximately second order
cyclostationary, which allows their separation from the deterministic gear components [58].
There are a number of methods for achieving this separation [28], but the one initially used in
this research was to first isolate and then remove the deterministic components corresponding to
each gear in the system, by synchronous averaging, leaving a residual stochastic signal, which
should be dominated by bearing faults in some frequency bands. The optimum frequency bands
are found using some sort of kurtogram to find the frequency band with maximum SK. In this
case, a wavelet kurtogram [59] was used. An alternative preprocessing technique used in this
case was cepstrum pre-whitening [57]. By this means, the signal spectrum amplitude is set to a
constant value (whitened) and the original phase used to generate a time signal. This
simultaneously nullifies the effect of both discrete frequencies and resonances, so that a
99
frequency band containing an impulsive signal will tend to dominate the time signal. SK can be
used to further isolate the impulsive band.
The first stage involves extracting a pseudo tachometer/encoder (tacho) signal from the measured
vibration signal; it was found that the speed signal provided by NREL was only useful for giving
an arithmetic mean estimate of the speed of the high speed shaft (generator rotor) and could not
be used for re-sampling purposes. The extracted tacho signal was used to resample the signal of
interest and extract the synchronous average for the intermediate shaft. The residual signal was
pre-whitened as a first step. The squared envelope spectrum was then obtained using the Hilbert
transform and scanned for bearing defect frequencies, which had been calculated for each
bearing in the gearbox. Pre-whitening was achieved using the newly proposed approach based on
the cepstrum (Cepstrum-pre-whitening). This squared enveloped signal was extracted and
inspected for bearing faults.
100
filled with zeros up to the sampling frequency - negative frequencies were set to zero - the
inverse transform signal is analytic (complex), and it is the real part that will then represent the
reference signal. Note that zero crossings represent 180 increments in rotation phase, and this is
unaffected by amplitude modulation by any positive modulating function.
1200
Amp
amplitude of the spectrum.
400
Frequency axis scaled in lines
0
200
rather than Hz
0
-200
650
700 Frequency (line no.) 1000
700 750 800 850 900 950 1000 1050 1100
800
800
700
600
100
0
0
0 Frequency (line no.) NFFT
-3
x 10
8
2
(c) Step 3: real part of the
IFFT (Inverse Fast Fourier
0
Transform) of the buffer
-2
obtained in step 2.
-4
-6
-8
0
0 1
Time (sample
2 3 4
no.) 5 6 7
NFFT
x 10
5
8
Figure 9.2. Reference (speed) signal extraction stages: (a) identifying a separable band; (b)
extracting the band into a new buffer; (c) inversing the transform signal b into the time domain
[60]
The signal obtained in 2.c is a sinusoidal-type signal whose periods represent the speed for each
shaft rotation. The speed variation in this signal (a reflection of the speed variation of the shaft
under investigation) can be traced using the zero crossings of the consecutive periods, which can
be achieved by detecting the zero crossings (an interpolation between the samples on either side).
This signal can also be used to order track the signal. Note, however, that this process may have
to be repeated progressively to order track the signal to higher harmonics and achieve better
results. This means that after each stage, a higher harmonic will be made available due to the
reduction of speed fluctuations and the analyst can select bands around the gear mesh harmonics
to improve the quality of order tracking and gain more accuracy.
The approach illustrated in Figure 9.2 was used to extract a tachometer and speed signal for the
High Speed shaft (HSS) of the gearbox. The speed extraction was based on the gear mesh
frequency of the high speed stage (22 HSS). For this purpose, sensor 7 was selected for
extracting the gear mesh signal for the HSS, although this can also be achieved using other
101
sensors in close proximity to the high speed shaft. The zero crossings for the signal in Figure 9.2
(c) were used to estimate the speed of the HSS. Examples of estimates at two speeds based on
this procedure (scaled in rpm) from data set a and data set c are presented in Figure 9.3. It is
noted that the speed is relatively constant.
2c: 5
1804
1803.5
Rotaional speed of high speed shaft (rpm)
1803
1802.5
1802
0 0.5 1 1.5 2 2.5 3 3.5 4
4
x 10
2a: 5
1206.5
1206
1205.5
1205
0 0.5 1 1.5 2 2.5
Gear mesh Periods 4
x 10
Figure 9.3. HSS estimates: Top - data 2a:5; Bottom - data 2c:5
The algorithm works by re-sampling the order-tracked signal to obtain an integer number of
samples per revolution for a specific shaft. The removal of the harmonics of that specific shaft
can be achieved by one of two methods. The first is by finding the synchronous average and
subtracting it (repeated periodically) from the signal. The second is by truncating the signal to an
integer number of revolutions (preferably a power of 2) and setting the lines corresponding to the
harmonics of that shaft (after FFT analysis) to 2 times the mean (complex) value of the adjacent
frequencies (the multiplication by 2 is to make the amplitude statistically the mean of the two
amplitudes). To avoid treating the negative frequency components, it is recommended that they
be set to zero after the FFT step, and double the positive frequency components, then take the
real part of the resulting analytic signal in the time domain. Both methods arrive at the same
result, as was presented in [62].
The extracted tacho signal was used to resample the signal of interest and extract the
synchronous average for the intermediate shaft. As the HSS and the intermediate speed shaft
(ISS) have a ratio of four, the removal of the harmonics of HSS was included at this stage. The
end result of this stage was four synchronously averaged signals for the ISS, Low Speed shaft
102
(LSS), the planet carrier shaft, and the planetary gears (a composite of all the planetary gears;
this was later updated to extract an average for each planetary gear). Typical results for the four
synchronously averaged signals are presented in Figure 9.4 and Figure 9.5 for sensor three
(planetary stage and low speed shaft) for both data set a and data set c.
2a:5 sensor 3
Intermediate Shaft
5
-5
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
Low Speed Shaft
0.5
0
Acceleration (m.s )
2
-0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Planet carrier
1
-1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Planet gear
0.5
-0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Time (s)
Figure 9.4. Synchronously averaged signals from sensor 3, data 2a:5
103
Intermediate Shaft
0.5
-0.5
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
0.5
-0.5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Planet carrier
0.5
-0.5
-1
0 0.5 1 1.5 2 2.5 3
Planet gear
0.5
-0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Time (s)
Figure 9.5. Synchronously averaged signals from sensor 3, Data 2c:5
The whitening stage can, in fact, be used on the raw signals giving an enhancement of the
bearing related signature. The advantage of pre-whitening is that all frequency components in
this signal are equally weighted and, thus, the potential to detect faults is enhanced.
104
Figure 9.6. Schematic diagram of the cepstral method for removing selected families of harmonics
and/or sidebands from time signals [57]
The squared envelope spectra for the pre-whitened residual signals were obtained using Hilbert
Transform techniques (simply by inverse transforming one-sided spectra, shifted to zero
frequency, and taking the modulus of the resulting complex numbers).
9.3 Results
9.3.1 Bearings
9.3.1.1 Initial Bearing Diagnosis Results
Through the inspection of the squared envelope spectra of the residual signals obtained from
section 9.2.4, two defective bearings were identified. These are FAG 3222 and NU 2220 ECM.
The diagnosis indicated an inner race fault on bearing FAG 3222 and an inner race fault and
roller/cage defect on the NU 2220 bearing.
The inner race fault (localized spalling) of the FAG thrust bearing (3222) was mainly detected
through signals from sensors seven and nine. The envelope spectra from these sensors contained
the ball pass frequency of the inner race (BPFI), and its harmonics were clearly modulated by the
high speed shaft speed. An example from the low speed data (set a) and the high speed data (set
c) is shown in Figure 9.7 and Figure 9.8 respectively. In Figure 9.7, the high speed shaft
frequency (20.1 Hz) and its second harmonic are clearly visible. The 230.7 Hz component
(suspected BPFI) and its second harmonic are modulated by the frequency of the high speed
shaft. The calculated BPFI for the FAG bearing (3222) at an inner race shaft speed of 20.1 Hz is
around 218.5 Hz. The difference between the observed BPFI in the figure and the calculated one
is around 5.5%, which can be due to slippage and an incorrect estimation of load angle. This rate
of slippage is also observed for the high speed data (set c) in Figure 9.8, where the BPFI of 345.3
Hz and its second harmonic are observed. The 345.3 Hz component has around a 5% difference
from the calculated BPFI, which is in agreement with the low speed data observation of Figure
9.7.
105
x 10
-7 Carrier Frequency at : 230.987 Hz, Sideband Spacing at : 20.0878 Hz
4
X: 20.1
3.5 Y: 3.364e-007
2.5
2
X: 230.7
Y: 1.697e-007
1.5
0.5 X: 461.4
Y: 2.537e-008
0
0 100 200 300 400 500 600
Frequency (Hz)
Figure 9.7. Squared envelope spectrum for data 2_a_10 sensor 7
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
0 100 200 300 400 500 600 700 800 900 1000
Frequency (Hz)
Figure 9.8. Squared envelope spectrum for data 2_c_10 sensor 7
106
Figure 9.9 shows that the suspected BPFI of 345.3 Hz is close to one of the harmonics of the
intermediate shaft (46th harmonic), but is not in fact a harmonic of the intermediate shaft.
0.05
0.04
0.03
0.02
0.01
0
330 335 340 345 350 355
Frequency (Hz)
Figure 9.9. Zoom-in around the BPFI. Harmonic cursors for the ISS
The second bearing fault diagnosis indicated the presence of an inner race fault and possible
roller/cage pitting in the SKF HSS upwind bearing (NU 2220 ECM). This was detected mainly
through sensor eight and was confirmed by testing data from set a through the presence of the
BPFI, its harmonics, and the modulation of these harmonics by the high speed shaft speed. There
are also indications of modulations by the fundamental train frequency (FTF), which can come
from variations between rollers.
In Figure 9.10, the BPFI at 198.08 Hz modulated by a shaft speed of 20.091 Hz is observed. This
matches very closely the calculated BPFI of bearing NU 2220 ECM at a shaft speed of about
20.1 Hz. The FTF harmonics (12.8 Hz) are at the same speed for the NU 2220 bearing, as shown
in Figure 9.11.
107
x 10
-4 Carrier Frequency at : 198.082 Hz, Sideband Spacing at : 20.0906 Hz
14
12
10
-2
140 160 180 200 220 240 260
Frequency (Hz)
Figure 9.10. Squared envelope spectrum for data 2_a_5 sensor 8 showing the BPFI of bearing
NU2220
0.05
0.04
0.03
0.02
X: 22.54
Y: 0.01013
0.01
0
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hz)
Figure 9.11. Squared envelope spectrum for data 2_a_5 sensor 8 showing the FTF harmonics of
bearing NU2220
108
9.3.1.2 Actual Findings and Missed Detection
In the test gearbox failure analysis report [4], it was shown that the IR raceway and rollers of
bearing 3222 had straw-yellow temper colors. The color implies that the temperature reached
about 400F. The root cause of the overheating was probably lubricant starvation. Even though
no spalls were detected, it is likely there was geometric distortion from the overheating. The IR
of bearing NU 2220 ECM had assembly damage at the roller spacing caused by cocking of the
rollers during blind assembly. Debris dents and lines of false brinelling were also observed. The
IR of bearing NU 2220 ECM had corrosion at roller spacing.
The spacer for bearing 32032X outer race had assembly damage at the roller spacing caused by
interference with the bearing rollers during assembly. This damage was missed in our initial (and
later) diagnosis. The ball-pass frequency of the outer race (BPFO), when the HSS speed is 30
Hz is estimated at 105. 9 Hz. Figure 9.12 shows the squared envelope analysis for data c_5,
sensor 6, where the HSS is 30.06 Hz and is present. A frequency at 210.4 Hz appears clearly in
this figure. This is close to 7HSS, but it is also close to 2BPFO. This main evidence seems to
indicate the presence of the fault in the 32032X, but it was not considered strong enough for us
to call the fault. Note also the presence of the 345.1 Hz, which is the BPFI of the NU 2220
bearing. Also, upon inspection of the spectrum comparison of sensor five (Figure 9.13), there is a
strong presence of 210 Hz and a change around this frequency, in particular.
-4
x 10
1.6
1.4
X: 30.06
Y: 0.0001195
1.2
X: 210.4
0.8 Y: 7.143e-005
0.6
X: 345.1
Y: 4.035e-005
0.4
0.2
0
0 100 200 300 400 500 600 700 800 900 1000
Frequency (Hz)
Figure 9.12. Squared envelope spectrum of data 2c_5 sensor 6 showing the shaft speed (30.06
Hz), what appears as 2BPFO for bearing 32032X and the BPFI for bearing NU 2220
109
Frequency (Hz)
Frequency (Hz)
Figure 9.13. Power spectrum density comparison of the high speed data through sensor 5
9.3.2 Gears
9.3.2.1 Initial Diagnosis
Initial diagnosis indicated the possibility of severe damage in the planetary stage of the gearbox.
The main indications came through sensors three and four and were observed more in the low
speed data (set a). It was noticed that the impact pattern came in pairs (roughly separated by 10-
12 teeth on the planetary gear). This was clearly observed in data 2_a, but not very clearly in
2_c. The residual signal from sensor five for data 2a_5 is shown in Figure 9.14, where the
impacts are seen clearly. The analysis of the squared envelope spectrum of the signal, Figure
9.14, is shown in Figure 9.15. The carrier speed (around 0.25 Hz), the spin frequency of the
planetary gear (PGSF) at around 0.625 Hz, and 3PGSF are very clear in Figure 9.15. It was
indicated at the time of this diagnosis that to confirm this, further analysis would be required to
obtain the synchronous average with respect to each planet.
110
6
-2
-4
-6
1 2 3 4 5 6 7
Time (Samples) 5
x 10
-5
x 10
3.5
2.5
2 X: 1.878
Y: 1.747e-005
X: 0.626
1.5
Y: 1.351e-005
X: 0.2504
Y: 9.977e-006
1
0.5
0 2 4 6 8 10 12 14 16 18 20
Frequency (Hz)
Figure 9.15. Squared envelope spectrum of the residual signal shown in Figure 9.14
111
9.3.2.2 Revised Diagnosis
The revised diagnoses included using the healthy set of data and comparing it with the faulty
one, in both the frequency domain (using power spectrum density) and the cepstrum domain. The
revised analysis also included the removal of shaft related components from the synchronously
averaged signals through pre-whitening. Finally synchronous averages were extracted for the sun
gear and each planet using an algorithm earlier developed by DSTO.
9.3.2.2.1 Power spectrum and cepstrum comparisons (healthy and faulty signals)
The scaling for the healthy spectra was given in g's, while the time domain was reported in ms-2.
The resolution of the healthy spectra was determined and used to find the equivalent FFT
transform size to use with the faulty data. When scaling both in dB, reference level 1e-6 was used
for the faulty data and 1e-7 for the healthy data to compensate for the units. Comparisons show
increases at the gear mesh frequencies and sideband families. This is shown clearly for all
sensors (AN3, AN5, AN6 and AN7). Harmonic and sideband cursors show dominant
components and modulations.
Cepstra were generated from the corresponding spectra to give more information on sideband
patterns. The cepstra represent the amplitude of the analytic cepstrum (from the one-sided log
spectrum). This version can also be used on zoom spectra [24].
Figure 9.16 shows the spectrum comparison using the data from sensor three, in the low speed
section of the gearbox. There is a noticeable increase (more than 20 dB) in the HSS, the epicyclic
mesh frequency, and its sidebands. Most noticeable are the sidebands at the planet pass
frequency around the epicyclic mesh frequency in the fault case. This is evident in the cepstrum
comparison presented in Figure 9.17. Note the second rahmonic in the healthy case,
corresponding to 1 times the carrier speed, which is unexplained. It is possible that it has
something to do with the far from hunting tooth design of the planetary section and means that
the particular tooth combinations occur much more frequently than usual.
In Figure 9.18 and Figure 9.19, the spectrum and cepstrum comparisons based on the data from
sensor five, with generator speed 30 Hz, are presented. The fact that the cepstrum does not
change appreciably shows that modulation at ISS (which would come from local faults) did not
occur, and the corresponding lack of sidebands in the faulty spectrum confirms that the faults are
distributed.
The distributed wear of the intermediate shaft pinion, ascribed in the inspection report to the
hunting tooth ratio, is shown in Figure 9.20 in the growth of the harmonics of the IS gear mesh.
Figure 9.21 shows the same spectra, but concentrates on the growth of sidebands around the HS
gear mesh harmonics, and they are spaced at the HS shaft speed. The corresponding cepstra of
Figure 9.22 shows that the local faults causing the sideband generation have grown from nothing
in the healthy condition; whereas, the increased peak corresponding to the ISS probably indicates
some growth of harmonics at this shaft speed since the sidebands were not in evidence. Both the
high speed pinion and gear had localized scuffing, which would explain the strong modulation at
HSS speed.
Data from sensor seven in Figure 9.23 shows the same story as Figure 9.21.
112
Harmonics at epicyclic gear mesh, sidebands at
planet pass frequency (3X carrier)
100
HSS
90
80
70
60
50
40
30
20
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hz)
100
90
80
70
60
50
40
30
20
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hz)
0.8 Carrier
0.6
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3 3.5 4
Quefrency (seconds)
1
nd
0.8
2 rahmonic unexplained (1 carrier speed)
0.6
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3 3.5 4
113
Harmonics at HSS
100
80
60
40
20
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)
100
80
60
40
20
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)
0.5
Rahmonics of ISS
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.5
Quefrency (seconds)
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
114
Harmonics at IS gearmesh
120
110
100
90
80 Faulty
70
60
50
40
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Frequency (Hz)
120
110
100
90
Healthy
80
70
60
50
40
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Frequency (Hz)
115
1
0.8
HSS ISS
0.6
0.4
0.2
0
0 0.05 0.1 0.15 0.2 0.25
Quefrency (seconds)
1
0.8
0.6
0.4
0.2
0
0 0.05 0.1 0.15 0.2 0.25
116
Figure 9.23. Spectrum comparison using the data from sensor 7
9.3.2.2.2 AN6 ISS signal processed to remove effects of HSS
Because the ratio is exactly 4:1, the ISS record was divided into four sections, which were
averaged, recombined, and subtracted. The residual record should contain only information from
the ISS, for example, the shaft harmonics that are not divisible by four and the IS gear mesh
frequency (23X). Time signals and spectra are shown in Figures 9.24 and 9.25.
The HS gear mesh is strongly modulated by the HSS, because damage is more localized. The IS
gear mesh is much distorted, with many harmonics, but it is not modulated. This is compatible
with the distributed damage attributed to the hunting tooth design.
117
10
(a)
5
-5
-10
0 1000 2000 3000 4000 5000 6000 7000 8000
Time (sample no.). One rotation of IS
10
(b)
5
-5
-10
0 1000 2000 3000 4000 5000 6000 7000 8000
Time (sample no.). One rotation of IS
Figure 9.24. Time records from the averaged ISS signals: (a) original, including four rotations of
the HSS; (b) Residual after removal of the HSS average.
4
10
(a)
2
10
0
10
0 20 40 60 80 100 120 140 160 180 200
Harmonics of ISS
4
10
(b)
2
10
0
10
0 20 40 60 80 100 120 140 160 180 200
Harmonics of ISS
Figure 9.25. Spectra of signals of Figure 9.24: (a) original including four rotations of the HSS;
(b) residual after removal of the HSS average.
118
low speed shaft, and on the annulus gear. No faults were reported on the planet gears, but this is
discussed in the next section.
Intermediate Shaft
0.2
-0.2
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
0.2
-0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Planet carrier
1
0.5
-0.5
-1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Planet gear
0.4
0.2
-0.2
-0.4
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Time (s)
Figure 9.26. Whitened synchronously averaged signals corresponding to the periods of all major
gear components in the gearbox, which enhance local faults
Note that the Planet gear result is a composite of all planets, and shows a similar result to the
individual planets in the next section.
9.3.2.2.4 Planet, sun, and annulus gear signatures extracted by Dr. David Forrester using
patented DSTO software#
Signatures for each planet and the sun gear were extracted by shifted and weighted averages of
signals taken as the various planets pass the measurement point [63]. Note that the signature for
each planet tooth is actually a composite of the two teeth meshing simultaneously with the
annulus and sun gears (opposite flanks). This is an important observation since no faults were
reported on the planet gears, even though virtually all Round Robin partners detected apparent
faults corresponding to the planet gear rotation period. This is discussed further below.
The DSTO patented method described in [63] and illustrated below produces average time
signals for each individual planet gear and the sun gear using shifted weighted signals from
passage of each planet past the transducer, with correction for the phase offset of individual teeth
119
for each passage. Note that the average for the sun gear can include the contact of individual
teeth with all planet gears.
The residual mentioned in some figures is the result of removing the regular tooth mesh signal so
as to highlight local faults on a gear.
Figure 9.28. Sun Gear residual of DSTO average data set 2c (high speed, high load)
Figure 9.28 shows the residual signal for the sun gear at high speed and load. This result is
compatible with the observation of the inspection report that the sun gear had localized fretting
corrosion.
120
Figure 9.29. Sun Gear residual of DSTO average data set 2a (low speed, low load)
Even at a very different speed and load, in Figure 9.29, the sun gear signature is almost identical
(though displaced because there is no common phase reference). The same was found for the
individual planet gear signatures discussed below, in Figure 9.30.
Planet 1
Planet 2
Planet 3
Revolutions
Several Round Robin partners found evidence of faults on the planet gears, even though none
were reported in the inspection report, but the above figure implies that the fault patterns are the
same on all three planets. This could be caused by the far from hunting tooth design of the
planetary section. Individual pairs of teeth on the annulus and sun gears, both of which had
121
faults, could mesh simultaneously on opposite sides of a given planet gear relatively frequently.
It would be natural for this to occur identically for all three planets, since all tooth numbers are
divisible by three. This potential explanation should be confirmed by more detailed analysis of
the kinematics of this particular configuration.
Figure 9.31 shows a typical average signal for the annulus gear. The residual in this case did not
clarify the local faults to any great extent. However, the visible variations are compatible with
the results of the inspection report, which found a distributed fault pattern from a local area of
scuffing. Because of the numbers of teeth of all planetary components being divisible by three,
the damage tended to imprint on every third tooth, and many examples this pattern over groups
of three teeth can be seen in the above figure.
When spectra were received for the gearbox in healthy condition, at about the same time as
receiving the inspection report, considerably more detailed analysis could be done as to the
details of the faults on each gear. An exception was the indication of faults on the planet gears,
which were not found on inspection. Much of this analysis could have been done blind, with the
availability of a healthy data for comparison from the outset.
122
Much more detailed analysis was later performed by Dr David Forrester, of DSTO in Australia,
using patented algorithms that allowed for the production of separate averages for the individual
planet gears and sun gear. These corresponded well with the detected faults on most gears, but
once again indicated faults on the planet gears. Other Round Robin partners found the same. It
now seems likely that the misdiagnosis was due to the far from hunting tooth design of the
planetary section, which could mean that faulty teeth on the sun and annulus gears could mesh
simultaneously with a healthy planet gear relatively frequently, and thus give an indication of a
fault on the planet gear. The fact that the three planets had the same fault pattern lends credence
to this interpretation, as the meshing patterns would likely be the same for all three planets (with
all tooth numbers divisible by three). This hypothesis needs to be further investigated before it
can be confirmed.
This highlighted the fact that gear diagnostics is made easier by the adoption of designs as close
as possible to hunting tooth designs, considered good design practice in any case. The exact
4:1 ratio between the IS and HS shafts also made it difficult to separate the faults on gears on
these two shafts, although luckily the second mesh on the ISS was a hunting tooth design.
For the relatively modest speed variations in the test data, it was possible to extract information
on instantaneous speed from the signals themselves in the form of a pseudo-encoder signal.
The signal, typically based on a high speed gear mesh component, can be used for order tracking
and, thus, synchronous averaging of gear signals throughout the gear train. For larger speed
variations, which are not uncommon with pitch controlled wind turbines, it would be necessary
to start with a lower order harmonic of the shaft speed; however, in principle, the process can be
done iteratively to improve the speed correction. This should be tested in the future.
123
10 A Two Stage Fault Detection Framework for Wind Turbine
Gearbox Condition Monitoring
Pingfeng Wang* and Prasanna Tamilselvan
Department of Industrial and Manufacturing Engineering, Wichita State University
*
Corresponding Author Email: [email protected]
10.1 Introduction
Maintaining wind turbines in top operating condition ensures not only a continuous revenue
generation but a reduction in electric power drawn from non-renewable and more polluting
sources. Despite the large capital for establishing a wind farm, the maintenance activities of wind
turbines are the primary contributors for the wind energy costs [64-66]. The need for reduction in
O&M costs are likely to increase due to the rising competition in todays global economy.
Effective health diagnosis of wind turbines provides various benefits, such as improved
reliability and reduced costs for operation and maintenance (O&M). Research on real-time
diagnostics and prognostics, which interpret data acquired by smart sensors and distributed
sensor networks, and utilization of these data streams to make critical O&M decisions offers
significant advancements in creating early awareness of wind turbine health condition before
unexpected failures. The unexpected breakdowns can be prohibitively expensive since they
immediately result in lost production [67-70]. To reduce, and possibly eliminate such problems,
real time condition monitoring is required to avoid sudden catastrophic system failures.
Vibration analysis is the most vastly used mechanism of condition monitoring in wind turbines.
It is mainly applied to identify the current condition of rotating components, such as gearbox,
generator, and main bearing, by installing mechanical sensors on the components. The defects of
the components are estimated based on the vibrations produced by these components during
operation. This chapter presents the vibration based condition monitoring framework with
analytical defect detection method and graphical analysis developed for the CM Round Robin
study.
124
Figure 10.1. Vibration based condition monitoring approach
The pre-processing of vibration data involves three steps, as shown in Table 10.1. The primary
step of the vibration analysis is the calculation of gear and bearing frequencies. The gear
meshing frequencies of all the gears, bearing frequencies such as Ball Passing Frequency Outer
(BPFO), Ball Passing Frequency Inner (BPFI), and Ball Spinning Frequency (BSF) of all the
bearings were determined for both speeds 1200 rpm and 1800 rpm. The next step involves the
identification of the relationship between the sensor and the components. The final step in the
pre-processing is to develop the frequency spectrum from the raw time domain vibration signal
using the fast Fourier transformation (FFT) process for the desired sensors.
Step 1 Calculate Gear Meshing Frequency (GMF) for gears and bearing frequencies
The Round Robin study involves three speed stages: low speed (LS), intermediate speed (IS) and
high speed (HS). In this research, the high frequency rotating components such as intermediate
and high speed stages of the gearbox are only considered for damage detection. Among the total
of 10 sensor values from the Round Robin gearbox, the desired sensors for the IS and HS are AN
5 to AN 9. The relationships between these sensors and the components are determined based on
the location and proximity to the rotating components, as listed in Table 10.2.
125
Table 10.2. Sensor and component relationship
Sensor
Sensor GMF SRF Bearing Damage
Name
The raw vibration signals from the sensors are time domain signals. The defects from the
bearings and gears can be identified from their corresponding desired frequency amplitudes in
the frequency spectrum. The FFT converts the time domain signal into a frequency domain
signal and helps in analyzing each desired frequency based on its amplitude and its harmonics.
Step 1 Determine maximum amplitude values for sidebands and desired frequency
K Lj ALj (21)
=KR = ; AR
K Rj ARj
Table 10.4 shows the different severity levels based on the frequency sideband features. The
severity factor analysis resulted in three severity factor metrics for online defect detection. The
severity factor 1 (SF1) ensures equal spread of the sidebands using the kurtosis ratio metric, as
shown in Equation (22). The threshold kurtosis ratio, KRT, is considered to be 0.6. The value of
SF1<1 denotes the unequal spread of sidebands and vice versa. The severity factor 2 (SF2)
ensures equal maximum amplitude of sidebands on both sides of the desired frequency, as shown
in Equation (23). The threshold amplitude ratio, AT, is considered to be 0.9. The value of SF2< 1
denotes the unequal frequency amplitudes on both sides of the sidebands and vice versa.
Min ( KR j , KR j 1 ) (22)
SF1 =
KRT
Min ( AR j , AR j 1 ) (23)
SF2 =
AT
The severity factor 3 (SF3) ensures that the maximum desired frequency amplitude is higher than
the maximum amplitude of the sideband Amax, as shown in Equation (24), where AF is the
maximum amplitude at the desired frequency. The value of SF3< 1 denotes the frequency
amplitude of the sideband, Amax, which is higher compared to the desired frequency, Amax.
AF (24)
SF3 =
Max ( ALj , ARj )
127
Table 10.5. Severity factor analysis of sensor AN 6 for Case 2a
Desired
Component SF1 SF2 SF3 Low Medium High
Frequency
ISS upwind
BPFI 0.98 0.73 2.85 0 1 0
bearing
ISS downwind
BPFO 0.86 0.42 2.94 0 1 0
bearing
HSS upwind
BPFO 0.58 1.06 0.23 0 0 1
bearing
HSS downwind
BPFI 0.72 0.58 1.67 0 1 0
bearing
The conditions SF1 1, SF2> 1, and SF3> 1 show that the component has a low severity defect.
The severity factor characteristics of the medium severity defect are SF1 1, SF2 1, and SF3>
1 and SF1> 1, SF2 1, and SF3> 1. Similarly, the high severity defect conditions are SF1 1,
SF2 1, and SF3 1; SF1> 1, SF2 1, and SF3 1; SF1 1, SF2> 1, SF3 1, and SF1> 1,
SF2> 1, SF3 1. Based on these rules, the severity levels and the failure modes of the
components are identified based on the each sensor. The severity factor analysis of sensor AN 6
for 2a case is listed in Table 10.5.
128
The desired component matrix and the severity factor levels of all the components are utilized
for developing a defect severity matrix. The severity ratio of component u at severity level g, Sug
is represented as Equation (26), where g represents the different severity levels, such as low,
medium, and high, and Sugm represents the severity level of component u at level g through
sensor m.
M
(26)
S ugm
Sug = m =1 3 M
S
g 1=
= m 1
ugm
The defect severity matrix, DS, represents the defect component and its severity level in the
matrix format as shown in Equation (27), where rows of the matrix represent each desired
component and columns represent the severity level of the components, such as low, medium
and high.
129
2a 10 AN6 ISS Radial
0.16
0.14
0.12
0.1
|Y(f)|
0.08
0.06
0.04
0.02
0
330 335 340 345 350 355 360
Frequency (Hz)
Figure 10.2. HSS downwind bearing BPFO graphical analysis
10.5 Results
The results from the online analytical defect detection method are used as inputs to the graphical
verification. The failure modes and their severity levels from the multi-dimensional vibration-
based sensory signals are verified graphically, and the results are unified to the component level,
with their corresponding severity levels, as shown in Table 10.6.
The tabulated results were identified before the receiving knowledge of the actual failure modes.
The possible number of failures that can be identified from the vibration analysis for this Round
Robin study is about seven. The proposed condition monitoring approach identified five failures
and their severity levels. Moreover, the failures identified by the proposed vibration analysis
approach do not have any false identification. The accuracy of the condition monitoring
approach is due to its two fold analysis process, i.e. the analytical identification and the graphical
verification. The preliminary results from the analytical identification are further fine-tuned
using the graphical verification to avoid false identifications.
130
As a summary, this research showed that the developed vibration based two stage fault detection
framework that integrates both analytical diagnostics and graphical diagnostics is quite effective
for analyzing gear and bearing faults in wind turbine transmissions, as proved by the CM Round
Robin study results. With successful studies and lessons learned on the drivetrain CM, the
research can be extended to a probabilistic complex system design framework that potentially
can quantify the functionality, reliability, uncertainty, and cost/benefits of condition monitoring
techniques. It can integrate them into a system-level wind turbine design practice, as a
fundamental solution of enhancing reliability and reducing life cycle cost.
131
11 Recommended Practices
Based on the comparison of diagnostics results provided by sixteen partners during the blind
study stage, as presented in Chapter 1, and the detailed results from eight of the sixteen partners,
as presented in Chapters 3-10, it is clear that there is still room for the industry to improve
vibration analysis algorithms. Some algorithms presented in this report have not been widely
adopted in commercialized vibration-based condition monitoring (CM) systems. If adopted, they
can lead to increased accuracy of vibration-based wind turbine drivetrain condition monitoring.
They may potentially help increase the cost effectiveness of wind turbine condition monitoring
techniques. In addition, based on the lessons learned in this study, some recommended practices
were provided by several partners, especially Impact Technologies, National Instruments, and
STC Consultants. They will be discussed in this chapter. It is hoped that these recommended
practices can be considered in future research and development efforts within the wind turbine
condition monitoring community.
This study, however, was challenged by poor speed measurements. As pointed out by almost all
of the project partners, there was no once per revolution signal, which is valuable for time
synchronous averaging for gear health condition diagnostics. The measured high speed shaft rpm
also showed oscillations, which could be worked around but was a challenge. A once per
revolution signal could have been generated based on the raw encoder readings throughout the
test and provided to all Round Robin project partners. It was not attempted so the project
partners could have enough time to conduct the blind stage data analysis. However, it is
recommended that a once per revolution signal be provided in gear health monitoring since
tachometer signal acquisition requires converting pulse trains into a series of timestamps and
speeds. In general, to isolate mechanical vibration frequencies from one another, accurate speeds
and angular positions of the shaft's rotation are critical. The GL guidelines [72] call for high
resolution speed measurements as part of an instrumentation system in the field. Further, to track
vibration frequencies with respect to speed, both vibration and speed measurements should be
made simultaneously or clocked from the same base clock. In the Round Robin project, speed
information came from the high speed shaft. Adding a tachometer to monitor the input rotor shaft
might provide a more accurate result for the lower speed components.
This Round Robin study focuses on the gearbox, as it is the only component with disassembled
information. If the main bearing is considered for study, accelerometers, with a measurement
frequency range down to 0.1 Hz, are recommended. Alternatively, new sensing techniques, such
as acoustic emission [73], can be investigated. This recommendation may also be applicable to
the ring gear. In addition, for some bearing locations and types inside the gearbox, it may be
worth evaluating axial-mounted accelerometers.
132
In this study, only the frequency domain baseline data was provided. Also, the test data was
collected from one test gearbox for a short period of time. Diagnostics results could have been
further improved by providing the baseline data in the time domain from a bigger population of
gearboxes of the same model, and a longer data acquisition window of months or years.
The data acquisition system developed for this study was put together to facilitate the Round
Robin research. Its main emphasis was to collect high resolution raw time series data, which can
be provided to vibration analysts for diagnosis of the monitored gearbox condition. When
deploying a condition monitoring system in the field, it is important to balance the amount of
data, communications, and timing of the data acquisition. Communications may be expensive
when using a cellular modem, or slow using a 900 MHz radio. The amount of data storage on-
board an embedded data acquisition system may be limited. For this balancing task, a machine
with a check trigger state containing continuous acquisition and analysis of incoming sensory
data can be investigated. When the monitored wind turbine is operating, and measured values
have changed by a specific percentage or delta, then both pre-trigger and post-trigger sensory
data is stored to a binary file with a complete descriptive set including enterprise, wind farm, and
wind turbine. By combining both periodic data recordings with data change driver recordings, a
complete picture of the wind turbine is possible, using just the right amount of data.
Though no sensor faults were present in the test data sets shared by NREL for this study, it is
generally recommended to perform sensor validation before using vibration data for condition
monitoring of rotating components. GL uses a similar guideline. [72] This will help reduce
ambiguity between sensor and mechanical faults and reduce false alarms.
The GRC test turbine operates at two relatively constant speeds. For variable speed wind
turbines, the GL guidelines call for order tracking using measured speed [72].
Due to the complexity in gearbox design and the dynamic operating conditions, an integrated
approach must be taken that uses diagnostic information from all components (gears, bearings,
and shafts) as a whole. In other words, the analysis needs to integrate all available diagnostic
information to confidently detect and isolate problems via a high level reasoning / classifier
method. In addition, vast differences in speed/torque could cause dramatic differences in results
of vibration analysis. The accuracy can be increased by comparing results from similar steady
state operating conditions, e.g., the operational-category concept [74], by applying techniques
that are less sensitive to the effects, or by normalizing results. Conventional gear diagnostic
features are limited for planetary gear component fault detection because of the changing
transmission path, moving fault location, and modulation. As such, special consideration should
be given to fault detection of these components. For example, multiple cycles of the hunting
133
tooth ratio should be recorded for each planetary component. In addition, diagnostic approaches
that are specifically designed for planetary gear fault detection should be used.
Fusing vibration results with those from other sensors would help complete the diagnostic
coverage. In particular, oil debris, oil temperature, and casing temperature [72,76] would provide
additional evidence of impending failures. Modeling/simulation of expected vibration at different
operating and fault conditions is useful and a good compliment to seeded fault and accelerated
life testing.
134
Appendix A Project Partners
Brel & Kjr Vibro A/S
GE Bently Nevada
Impact Technologies
IVC Technologies
National Instruments
Purdue University
Schenck Corporation
Sentient Corporation
University of Cincinnati
University of Connecticut
University of Iowa
135
References
1. Motavalli, J. (2004). Catching the Wind. The Environmental Magazine; Dec 31, 2004.
http://www.emagazine.com/archive/2176
2. Link, L.; LaCava, W.; van Dam, Jeroen.; McNiff, B.; Sheng, S.; Wallen, R.; McDade, M.;
Lambert, S.; Butterfield, S.; Oyague, Francisco. Gearbox Reliability Collaborative Project
Report: Findings from Phase 1 and Phase 2 Testing. NREL/TP-5000-51885. Golden, CO:
National Renewable Energy Laboratory, June 2011.
http://www.nrel.gov/docs/fy11osti/51885.pdf.
3. Sheng, S.; Link, L.; LaCava, W.; van Dam, J.; McNiff, B.; Veers, P.; Keller, J.; Butterfield, S.;
Oyague, Francisco. Wind Turbine Drivetrain Condition Monitoring During GRC Phase 1 and
Phase 2 Testing. NREL/TP-5000-52748. Golden, CO: National Renewable Energy Laboratory,
October, 2011. http://www.nrel.gov/docs/fy12osti/52748.pdf
5. Musial,W.; McNiff, B. Wind Turbine Testing in the NREL Dynamometer Test Bed.
American Wind Energy Associations WindPower 2000 Conference; April 30-May 4, 2000, Palm
Springs, CA. NREL/CP-500-28411. Golden, CO: National Renewable Energy Laboratory, June
2000 . http://www.nrel.gov/docs/fy00osti/28411.pdf
7. Frarey, J.L. The History and Application of the Envelope Detector. Technology Showcase:
Integrated Monitoring, Diagnostics and Failure Prevention, Proceeding of a Joint Conference,
Mobile, Alabama, and April 22-26, 1996. Haymarket, VA: Society for Machinery Failure
Prevention Technology. http://www.dtic.mil/cgi-bin/GetTRDoc?AD=ADP010154
8. Broderick, J.J.; Burchill, R.F.; Clark, H.L. Design and Fabrication of Prototype System for
Early Warning of Impending Bearing Failure. NASA-CR-123717, MTI-71TR1. Hunstville, AL:
National Aeronautics and Space Administration (NASA), January 1972.
http://ntrs.nasa.gov/archive/nasa/casi.ntrs.nasa.gov/19720020851_1972020851.pdf
9. Luo, H.; Qiu, H.; Ghanime, G.; Hirz, M.; van der Merwe, G. Synthesized Synchronous
Sampling Technique for Differential Bearing Damage Detection. Journal of Engineering for
Gas Turbines and Power; Vol.132, Issue 7, 2010. http://dx.doi.org/10.1115/1.4000092
10. Hochmann, D.; Sadok, M. Theory of Synchronous Averaging. Proceedings of the 2004
IEEE Aerospace Conference; March 6-13, 2004, Big Sky, Montana. Washington, DC: IEEE,
2004; pp. 3636-3653. http://dx.doi.org/10.1109/AERO.2004.1368181
136
11. Hanna, J.; Hatch, C.; Kalb, M.; Weiss, A.; Luo, H. Detection of Wind Turbine Gear Tooth
Defects Using Sideband Energy Ratio. China Wind Power 2011;October, 19-21, 2011,
Beijing, China. http://www.ge-mcs.com/download/monitoring/SER-Technical-Paper.pdf
12. Marple, L. Jr., Computing the Discrete-Time Analytic Signal via FFT. IEEE Transactions
on Signal Processing, Vol. 47, No. 9, 1999; pp.2600-2603. http://dx.doi.org/10.1109/78.782222
13. Luo, H.; Fang, X.; Ertas, B. Hilbert transform and its engineering applications. AIAA
Journal, Vol. 47, No. 4, 2009 pp. 923-932. http://dx.doi.org/10.2514/1.37649
14. Bagnoli, S.; Capitani, R.; Citti, P. Comparison of Accelerometer and Acoustic Emission
signals as Diagnostic Tools in Assessing Bearing Damage, Proceedings of the 2nd International
Conference: Condition Monitoring; 24-25 May 1988, London, UK. Cranfield, England: BHRA,
1988.
15. Braun, S.; Datner, B. Analysis of Roller/Ball Bearing Vibrations. Transactions of the
ASME, Vol. 101, 1979; pp. 118-125.
16. Harris, T.A. Rolling Bearing Analysis. 4th edition. New York, NY: John Wiley & Sons, York,
2001; pp. 993-1000.
17. Braun, S.G.; Seth. B.B. On the Extraction and Filtering of Signals Acquired from Rotating
Machines. Journal of Sound and Vibration, Vol. 65, No. 1, 1979; pp. 37-50.
18. Lebold, M.; McClintic, K.; Campbell, R.; Byington, C.; Maynard, K. Review of Vibration
Analysis Methods for Gearbox Diagnostics and Prognostics. Proceedings of the 54th Meeting of
the Society for Machinery Failure Prevention Technology; May 1-4, 2000,Virginia Beach, VA.
Haymarket, VA: Society for Machinery Failure Prevention Technology, 1985.
http://php.scripts.psu.edu/users/k/p/kpm128/pubs/FeatureTutorialBody18.PDF
19. Szczepanik, A. "Time Synchronous Averaging of Ball Mill Vibrations." Mechanical Systems
and Signal Processing; Vol. 3, No. 1, 1989, pp. 99-107. http://dx.doi.org/10.1016/0888-
3270(89)90025-3
20. Ma, J. "Energy Operator and Other Demodulation Approaches to Gear Defect Detection."
Proceedings of the 49th Meeting of the Society of Mechanical Failures Prevention Technology;
April 18-20, Virginia Beach, VA. Haymarket, VA: Society for Machinery Failure Prevention
Technology, 1995; pp. 127-140.
21. Favaloro, S.C. A Preliminary Evaluation of Some Gear Diagnostics Using Vibration
Analysis. ARL-AERO-RPOP-TM-427. Melbourne, Australia: Aeronautical Research
Laboratory, 1985.
22. Watson, M. J.; Sheldon, J.S.; Lee, H. Novel Joint Time Frequency Vibration Diagnostics of
Turbine Engine Accessories. ASME Turbo Expo 2010: Power for Land, sea, and Air; June 14-
18, 2010, Glasgow, UK. New York, NY: ASME, 2010; pp. 353-363.
http://dx.doi.org/10.1115/GT2010-23539
137
23. McFadden, P. A Revised Model for the Extraction of Periodic Waveforms by Time Domain
Averaging. Mechanical Systems and Signal Processing; Vol. 1, No. 1, 1999; pp. 83-95.
http://dx.doi.org/10.1016/0888-3270(87)90085-9
24. Randall, R.B. Vibration-based Condition Monitoring. Hoboken, NJ: Wiley, 2011.
25. McFadden, P.; Smith, J. A Signal Processing Technique for Detecting Local Defects in a
Gear from a Signal Average of the Vibration. Proceedings of the Institution of Mechanical
Engineers, Part C: Journal of Mechanical Engineering Science; Vol. 199, No. 4, 1985; pp. 287-
292.
26. Zakrajsek, J.J.; Townsend, D.P.; Decker, H.J. An Analysis of Gear Fault Detection Method
as Applied to Pitting Fatigue Failure Damage. 47th Mechanical Failure Prevention Group; April
13-15, 1993, Virginia Beach, VA. NASA Technical Memorandum 105950. Cleveland, OH:
National Aeronautics and Space Administration (NASA), Lewis Research Center, 1993.
http://www.dtic.mil/cgi-bin/GetTRDoc?AD=ADA263115.
27. Mark, W.D.; Lee, H.; Patrick, R.; Coker, J.D. A Simple Frequency-Domain Algorithm for
Early Detection of Damage Gear Teeth. Mechanical Systems and Signal Processing; Vol. 24,
No. 8, 2010; pp. 2807-2823. http://dx.doi.org/10.1016/j.ymssp.2010.04.004
28. Randall, R.B.; Antoni, J. Rolling Element Bearing Diagnostics-A Tutorial. Mechanical
Systems and Signal Processing; Vol. 25, No. 2, 2011; pp. 485-520.
http://dx.doi.org/10.1016/j.ymssp.2010.07.017.
29. McFadden, P.D. Window Functions for the Calculation of the Time Domain Averages of
the Vibration of the Individual Planet Gears and Sun Gear in an Epicyclic Gearbox. Journal of
Vibration and Acoustics; Vol. 116, No. 2, 1994. http://dx.doi.org/10.1115/1.2930410
30. Antoni, J.; Randall, R.B. The Spectral Kurtosis: Application to the Vibratory Surveillance
and Diagnostics of Rotating Machines. Mechanical Systems and Signal Processing; Vol. 20,
No. 2, 2006; pp. 308-331. http://dx.doi.org/10.1016/j.ymssp.2004.09.002.
31. Combet, F.; Gelman, L. Novel Adaptation of the Demodulation Technology for Gear
Damage Detection to Variable Amplitudes for Mesh Harmonics. Mechanical Systems and
Signal Processing; Vol. 25, No. 3, 2011; pp. 839-845.
http://dx.doi.org/10.1016/j.ymssp.2010.07.008
32. El Badaoui, M.; Guillet, F.; Daniere, J. New Applications of the Real Cepstrum to Gear
Signals, Including Definition of a Robust Fault Indicator. Mechanical Systems and Signal
Processing; Vol. 18, No. 5, 2004; pp. 1031-1046. http://dx.doi.org/10.1016/j.ymssp.2004.01.005
33. McFadden, P.D.; Smith, J.D. Vibration Monitoring of Rolling Element Bearings by the
High-Frequency Resonance Technique-A Review. Tribology International; Vol. 17, No. 1,
1984; pp. 3-10. http://dx.doi.org/10.1016/0301-679X(84)90076-8
138
34. Combet, F.; Gelman, L. Optimal Filtering of Gear Signals for Early Damage Detection
Based on the Spectral Kurtosis. Mechanical Systems and Signal Processing; Vol. 23, No. 3,
2009; pp. 652-668. http://dx.doi.org/10.1016/j.ymssp.2008.08.002
35. Bonnardot, F.; El Badaoui, M.; Randall, R.B.; Daniere, J.; Guillet, F. Use of the
Acceleration Signal of a Paradox in Order to Perform Angular Resampling (with Limited Speed
Fluctuation). Mechanical Systems and Signal Processing; Vol. 19, No. 4, 2005; pp. 766-785.
http://dx.doi.org/10.1016/j.ymssp.2004.05.001
37. Dempsey, P.J. A Comparison of Vibration and Oil Debris Gear Damage Detection Methods
Applied to Pitting Damage. 13th International Congress on Condition Monitoring and
Diagnostic Engineering Management; December 3-8, 2000, Houston, Texas. NASA/TM-2000-
210371. Cleveland, OH: National Aeronautics and Space Administration (NASA), Glenn
Research Center, 2000; 18 pp.
http://ntrs.nasa.gov/archive/nasa/casi.ntrs.nasa.gov/20000120403_2000175883.pdf
38. McFadden, P.D. Detecting Fatigue Cracks in Gears by Amplitude and Phase Demodulation
of the Meshing Vibration. Journal of Vibration, Acoustics, Stress, and Reliability in Design;
Vol. 108, No. 2, 1986; pp. 165-170. http://dx.doi.org/10.1115/1.3269317
39. Forrester, D. A Method for the Separation of Epicyclic Planet Gear Vibration Signatures.
Proceedings of the Acoustical and Vibratory Surveillance Methods and Diagnostic Techniques;
Senlis, France. 1998; pp. 539-548.
40. Samuel, P.D.; Conroy, J.K.; Pines, D.J. Planetary Transmission Diagnostics. NASA/CR-
2004-213068. Cleveland, OH: National Aeronautics and Space Administration (NASA), Glenn
Research Center, May 2004.
http://ntrs.nasa.gov/archive/nasa/casi.ntrs.nasa.gov/20040073522_2004069093.pdf
41. Ciang, C.C.; Lee, J.-R.; Bang, H.-J. Structural Health Monitoring for a Wind Turbine
System: A Review of Damage Detection Methods. Measurement Science and Technology; Vol.
19, No. 12, 2008. http://dx.doi.org/10.1088/0957-0233/19/12/122001
42. McFadden, P.D. Detecting Fatigue Cracks in Gears by Amplitude and Phase Demodulation
of the Meshing Vibration. Journal of Vibration, Acoustics, Stress, and Reliability in Design;
Vol. 108, No. 2, 1986; pp. 165-170. http://dx.doi.org/10.1115/1.3269317
43. Lin, J.; Zuo, M.J. Gearbox Fault Diagnosis Using Adaptive Wavelet Filter. Mechanical
Systems and Signal Processing; Vol. 17, No. 6, 2003; pp. 1259-1269.
http://dx.doi.org/10.1006/mssp.2002.1507
139
44. Barszcz, T.; Randall, R.B. Application of Spectral Kurtosis for Detection of a Tooth Crack
in the Planetary Gear of a Wind Turbine. Mechanical Systems and Signal Processing; Vol. 23,
No. 4, 2009; pp. 1352-1365. http://dx.doi.org/10.1016/j.ymssp.2008.07.019
45. Yan, R.; Gao, R.X. Multi-Scale Enveloping Spectrogram for Vibration Analysis in Bearing
Defect Diagnosis. Tribology International; Vol. 42, No. 2, 2009; pp. 293-302.
http://dx.doi.org/10.1016/j.triboint.2008.06.013.
46. Tsai, C.-S.; Hsieh, C.-T.; Huang, S.-J. Enhancement of Damage-Detection of Wind Turbine
Blades via CWT-Based Approaches. IEEE Transactions on Energy Conversion; Vol. 21, No. 3,
2006; pp. 776-781. http://dx.doi.org/10.1109/TEC.2006.875436
47. Watson, S.J.; Xiang, B.J.; Yang, W.; Tavner, P.J.; Crabtree, C.J. Condition Monitoring of
the Power Output of Wind Turbine Generators Using Wavelets. IEEE Transactions on Energy
Conversion; Vol. 25, No. 3, 2010; pp. 715-721. http://dx.doi.org/10.1109/TEC.2010.2040083
48. Mohanty, A.R.; Kar, C. Fault Detection in a Multistage Gearbox by Demodulation of Motor
Current Waveform. IEEE Transactions on Industrial Electronics; Vol. 53, No. 4, 2006; pp.
1285-1297. http://dx.doi.org/10.1109/TIE.2006.878303.
49. Wang, W.J.; McFadden, P.D. Application of Wavelets to Gearbox Vibration Signals for
Fault Detection. Journal of Sound and Vibration; Vol. 192, No, 5, 1996; pp. 927-939.
http://dx.doi.org/10.1006/jsvi.1996.0226
50. Yuan, X.; Cai, L. Variable Amplitude Fourier Series with its Applicationh in Gearbox
Diagnosis-Part I: Principle and Simulation. Mechanical Systems and Signal Processing; Vol.
19, No. 5, 2005; pp. 1055-1066. http://dx.doi.org/10.1016/j.ymssp.2004.10.011
51. Yuan, X.; Cai, L. Variable Amplitude Fourier Series with its Applicationh in Gearbox
Diagnosis-Part I: Principle and Simulation. Mechanical Systems and Signal Processing; Vol.
19, No. 5, 2005; pp. 1067-1081. http://dx.doi.org/10.1016/j.ymssp.2004.10.012
52. Tan, P.-N.; Steinbach, M.; Kumar, V. Introduction to Data Mining. Boston, MA: Pearson
Addison Wesley, 2006.
53. Devijver, P.A.; Kittler, J. Pattern Recognition: A Statistical Approach. Englewood Cliffs, NJ:
Prentice-Hall, 1982.
54. Mosteller, F. A k-Sample Slippage Test for an Extreme Population. The Annals of
Mathematical Statistics; Vol. 19, No. 1, 1948; pp. 58-65. http://www.jstor.org/stable/2236056
55. Antoni, J. The Spectral Kurtosis: A Useful Tool for Characterising Non-Stationary Signals.
Mechanical Systems and Signal Processing; Vol. 20, No. 2, 2006; pp. 282-307.
http://dx.doi.org/10.1016/j.ymssp.2004.09.001
140
56. Sawalhi, N.; Randall, R.B. Semi-Automated Bearing Diagnostics-Three Case Studies.
Comadem Conference;June 13-15, 2007, Faro, Portugal. Coimbra : Inst. de Telecomunicaes,
2007.
57. Sawalhi, N.; Randall, R. B. Signal Pre-Whitening for Fault Detection Enhancement and
Surveillance in Rolling Element Bearings. Eighth International Conference on Condition
Monitoring and Machinery Failure Prevention Technologies; June 20-22,2011, St David's Hotel,
Cardiff, UK, 20-22 June.
58. Antoni, J. Differential Diagnosis of Gear and Bearing Faults. Journal of Vibration and
Acoustics; Vol. 124, No. 2, 2002; pp. 165-171. http://dx.doi.org/10.1115/1.1456906
59. Sawalhi, N.; Randall, R.B. Spectral Kurtosis Optimization for Rolling Element Bearings.
ISSPA 2005: The 8th International Symposium on Signal Processing and its Applications; August
28-31, 2005, Sydney, Australia. Piscataway, NJ: IEEE, 2005; pp. 839-842.
http://dx.doi.org/10.1109/ISSPA.2005.1581069
60. Urbanek, J.; Barszcz, T.; Sawalhi, N.; Randall, R.B. Comparison of Amplitude Based and
Phase Based Methods for Speed Tracking in Application to Wind Turbines. IMEKO TC 10,
11th Workshop on Smart Diagnostics of Structures; October 18-20, Krakow, Poland.
61. Sawalhi, N.; Randall, R.B. Localised Fault Diagnosis in Rolling Element Bearings in
Gearboxes. The Fifth International Conference on Condition Monitoring and Machinery Failure
Prevention Technologies; July 15-18, Edinburgh, UK, 2008. Oxford, UK: Coxmoor Publishing
Company, 2008.
62. Randall, R.B.; Sawalhi, N.; Coats, M. A Comparison of Methods for Separation of
Deterministic and Random Signals. International Journal of Condition Monitoring; Vol. 1, No.
1; 2011; pp. 11-19. http://dx.doi.org/10.1784/204764211798089048
63. Forrester, D.; Blunt, D. Analysis of Epicyclic Gearbox Vibration. DSTO HUMS;
Conference, Melbourne. 2003.
64. Tamilselvan, P.; Wang, P.; and Twomey, J. Quantification of Economic and Environmental
Benefits for Prognosis Informed Wind Farm Operation and Maintenance. ISERC 2012, 62nd
Annual IIE Industrial Systems and Engineering Research Conference; May 19-23, 2012,
Orlando, FL, USA.
65. Tamilselvan, P.; Wang, Y.; and Wang, P. Optimization of Wind Turbines Operation and
Maintenance Using Failure Prognosis. PHM 2012, IEEE 2012 Prognostics and Health
Management; June 18-21, 2012, Denver, CO, USA.
66. Tamilselvan, P.; Wang, P.; and Twomey, J. Prognosis Informed Stochastic Decision
Making Framework for Operation and Maintenance of Wind Turbines. ISFA 2012, ASME 2012
International Symposium on Flexible Automation; June 18-20, 2012, St. Louis, MO, USA.
141
67. Tamilselvan, P.; Wang , P.; and Jayaraman, R. Diagnostics with Unexampled Faulty States
Using a Two-Fold Classification Method. PHM 2012, 2012 IEEE International Conference on
Prognostics and Health Management; June 18-21, 2012, Denver, CO, USA.
68. Tamilselvan, P.; and Wang P. A Hybrid Inference Approach for Health Diagnostics with
Unexampled Faulty States. AIAA 2012-1784, 53th AIAA/ASME/ASCE /AHS/ASC Structures,
Structural Dynamics, and Materials Conference; April 23-26, 2012, Honolulu, Hawaii, USA.
69. Tamilselvan, P.; and Wang, P. Structural Health Diagnosis Using Deep Belief Network
Based State Classification. AIAA 2012-1783, 53th AIAA/ASME/ASCE /AHS/ASC Structures,
Structural Dynamics, and Materials Conference; April 23-26, 2012, Honolulu, Hawaii, USA.
70. Tamilselvan, P.; Wang, P.; and Youn, B. Multi-Sensor Health Diagnosis Using Deep Belief
Network Based State Classification. ASME 2011 International Design Engineering Technical
Conference; August 29-31, 2011, Washington DC, USA.
71. Spectra Quest Tech Note. Analyzing Gearbox Degradation Using Time-Frequency
Signature Analysis. March, 2006.
72. Germanischer Lloyd. Guidelines for the Certification of Condition Monitoring Systems for
Wind Turbines. Hamburg, Germany, 2007.
73. Smulders, A. Challenges of Condition Monitoring for Wind Turbines and Successful
Techniques. Wind Turbine Condition Monitoring; October 8-9, 2009, Omni Interlocken Resort,
Broomfield, CO. Golden, CO: National Renewable Energy Laboratory, 2009.
http://wind.nrel.gov/public/Wind_Turbine_Condition_Monitoring_Workshop_2009/3.2.Smulder
s.Challenges_of_CM_for_WT_and_Successful_Techniques.pdf
74. Gellermann, T.; Walter, G. Requirements for Condition Monitoring Systems for Wind
Turbine. AZT Report No. 03.01.068, 2003.
75. Wade, D. and Larsen Christopher. "Measurement of Gearbox Surface Frequency Repsonse
Functions for HUMS Algorithm Improvement. Presented at the American Helicopter Society
68th Annual Forum, Fort Worth, TX. May 1-3, 2012.
142