Available online at www.sciencedirect.

com

ScienceDirect
Procedia Engineering 144 (2016) 305 – 311

12th International Conference on Vibration Problems, ICOVP 2015

Fault Size Estimation Using Vibration Signatures in a Wind

Turbine Test-Rig
Sailendu Biswal, Jithin Donny George, G.R Sabareesh*
Department of Mechanical Engineering, BITS-Pilani Hyderabad Campus, Hyderabad,500078,India

Abstract

Fault size evaluation has become more significant in recent years to determine the fault size or the severity of fault apart from the
fault detection for prediction of remaining useful life. The present investigation focuses on the fault size estimation of gear root
crack in a wind turbine test rig using vibration signatures. A wind turbine test rig was developed at BITS-Pilani, Hyderabad
Campus to simulate the working of a wind turbine. Time domain vibration signals are acquired using accelerometers for the
healthy as well as faulty components. Discrete Wavelet Transform of vibration signature is performed and features are extracted
from the statistical analysis of wavelet coefficients and the extracted features are used as inputs in an ANN (Artificial Neural
Network) to effectively predict the size of gear root crack.
© 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
© 2016 The Authors. Published by Elsevier Ltd.
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-reviewunder
Peer-review under responsibility
responsibility of organizing
of the the organizing committee
committee of ICOVP
of ICOVP 2015 2015.

Keywords: Wind Turbine, Vibration Signature, Discrete Wavelet Transform, Fault Size Estimation, ANN

1. Introduction

Wind power has come up as an emerging source of renewable energy and it’s expected to amount to 25% of
world’s total energy by 2035.However, the operating and maintenance cost of wind turbine is estimated to be 20-
35% of the total power generation cost during its lifetime. Condition Based Maintenance (CBM) has become more
significant in wind turbine health monitoring due to its convincing ability to reduce Life Cycle Cost (LCC).
Predictive maintenance of critical components of wind turbine such as gearbox , bearings, shaft has become matter

* Corresponding author. Tel.: +91-8185919590; fax: +91-40-66303655.

E-mail address: sabareesh@hyderabad.bits-pilani.ac.in

1877-7058 © 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the organizing committee of ICOVP 2015
doi:10.1016/j.proeng.2016.05.137
306 Sailendu Biswal et al. / Procedia Engineering 144 (2016) 305 – 311

of concern as unscheduled maintenance cost of these components will be significantly larger than the scheduled
maintenance cost [1].The wind turbine gearbox is vulnerable towards failure due to high amount of load, low
speed and both fluctuating speed and load[2].The critical components used in wind turbine gearbox is often
susceptible to catastrophic failure due to development of localized faults .Gear faults such as wear ,misalignment
etc. can be categorised as distributed fault whereas pitting ,chipping, crack etc. are categorized as local faults[3].
Local faults or cracks of varying sizes often occur in gears of wind turbine. Nature of fault as well as its intensity is
characterized by the depth and width of these cracks. However it has become more significant to determine the fault
size or the severity of fault apart from the fault detection for prediction of remaining useful life[4].
Signal processing of raw vibration data obtained from critical component has been a crucial factor in successful
fault detection as well as in fault size estimation. Many research and investigations has been made for fault detection
in gears as they are considered as one of the vital component in a wind turbine, automotive or other rotating
machineries for power transmission. The fault detection becomes a challenging affair in a wind turbine gearbox due
to presence of multi stage of gears and the fault features are covered up by background random noise and vibration
[2].
The time domain analysis technique such as time synchronous averaging (TSA) technique was implemented [5-7]
for gearbox fault diagnostics due to its ability to diminish the effects of vibration signal which are not synchronous
to the shaft and gear mesh frequencies and at the same time can amplify required part of signal over the noise.
Statistical parameters such as skewness, kurtosis, root mean square (rms) value, crest factor, shape factor, clearance
factor, impulse factor etc. has been used for fault detection and fault severity prediction [3, 8].Apart from earlier
mentioned statistical parameters some special statistical parameters which includes FM0,FM4,NA4,NB4,M6A etc.
has become advantageous particularly for gear fault detection[3, 4, 7]. Frequency domain analysis such as cepstral
analysis [9] and spectrum analysis [10] has been effectively implemented for gear fault detection. Frequency domain
features such as energy ratio, energy operator, mean frequency, frequency centre, root mean square frequency,
standard deviation frequency was used along with other time domain features for fault level estimation[3, 11].
Vibration signal obtained from a gear having local fault contains amplitude and phase modulation. Faulty teeth of a
gear often gives rise to sidebands which results from modulation of gear meshing frequency. Sidebands occur around
a central frequency and it spreads over a wide frequency range due to its short time period of impact. The central
frequency is also known as carrier frequency. The sideband is the region of importance containing local fault
features, thus a properly chosen narrow band spectrum analyser could be very useful for fault detection[12].
However recent fault detection techniques for wind turbine gear box fault detection gives a preference for time
frequency analysis. Vibration signatures obtained from wind turbine gear box used to be nonstationary. The time
frequency analysis is desirable for nonstationary signal analysis[13].The time frequency analysis techniques such
as time-frequency demodulation [14],Wigner-Ville Distribution(WVD) [13], Empirical Mode Decomposition(EMD)
[15],wavelet analysis[13, 16-20],Hilbert –Huang Transform[21, 22] etc. has been utilized for fault detection in gear.
However, wavelet transformation is found to be a better time frequency analysis technique practised for
nonstationary vibration signals from wind turbine. Fault diagnosis in wind turbine was performed based on Morlet
wavelet transformation [13, 23] and multiwavelet denoising [24].

The fault diagnosis and fault size estimation of wind turbine components is a matter of great concern in order to
provide an effective maintenance solution as well as it helps in reducing the overall operating and maintenance cost.
The time frequency analysis of vibration signature is often accomplished by performing feature extraction [25],
followed by feature classification. Artificial intelligent methods based on Artificial Neural Network
(ANN)[26],Support Vector Machine(SVM) [8, 27, 28],Immune Genetic Algorithm(IGA)[29]etc. were successfully
implemented for feature classification in wind turbine fault diagnosis.
This paper discusses the fault size estimation for a gear root crack in a wind turbine test rig using Discrete
Wavelet Transform (DWT) and Artificial Neural Network (ANN).
 Sailendu Biswal et al. / Procedia Engineering 144 (2016) 305 – 311 307

2. Experimental Setup

A wind turbine test rig was developed at BITS-Pilani Hyderabad Campus to perform condition monitoring
studies as shown in Fig. 1. [30]. The wind turbine test rig comprises of three stage gear train having an overall speed
ratio of 48:1. A 1hp, 3phase AC motor was used in place of generator to perform the experiment and the speed of
the motor was controlled by variable frequency drive .Three test cases including one for healthy condition and other
two cases for faulty condition were performed. In order to simulate the faulty condition, two different root cracks
were introduced to the gear tooth individually. The difference in root cracks was characterized by the depth of the
crack. Using wire cut EDM (Electro Discharge Machining) the cracks were created at the root of gear as shown in
Fig. 2. (a) and (b). The crack parameters are specified in the Table 1. The experiment was conducted at 1200 rpm
and the vibration signatures were collected using an industrial accelerometer and National Instrument Data
Acquisition System (NI DAQ 9234) having 10000 sample length and with a sampling rate of 2400Hz .The pinion
was having 20 teeth with module 3 , 200 pressure angle and involute profile. The Gear Meshing Frequency (GMF)
was calculated to be 400Hz using GMF= (t*rpm)/60 where‘t’ is number of gear teeth and ‘rpm’ refers to the speed
of pinion. For each test case, 80 number of discrete time domain signal samples were collected and these time
domain vibration signals were further analyzed to predict the fault size in the gear of wind turbine test rig.

Fig. 1. Wind Turbine Test Rig

a b

Fig. 2.(a) Pinion with 1.5mm Depth Root Crack ; (b) Pinion with 2.5mm Depth Root Crack
308 Sailendu Biswal et al. / Procedia Engineering 144 (2016) 305 – 311

Table 1.Showing Fault Parameters

Fault Type Length(l in mm) Width(w in mm) Depth(d in mm)

Root Crack-1 30 0.25 1.5
Root Crack-2 30 0.25 2.5

3. Wavelet Analysis

3.1. Continuous Wavelet Transform (CWT)

The CWT of signal (t) is represented below

(1)

The mother wavelet is described by

(2)

Where τ is translation and s is scale.

CWT is a highly redundant representation as it computes many more coefficients than necessary. In signal
analysis, a compact representation is desirable to avoid the requirement of high computational capacity. DWT
provides compact representation with minimum number of coefficients. In CWT, the translation and scaling were
arbitrarily chosen where as in DWT the translation and scaling is mathematically restricted. DWT can be defined as
CWT evaluated at specific scale and translation. In DWT the translation is proportional to width of wavelet.

3.2. Discontinuous Wavelet Transform (DWT)

(3)

The mother wavelet is described by

(4)

The features from raw vibration signals were extracted using multilevel decomposition and an orthogonal wavelet
db5 (Daubeschies Wavelet) was used as mother wavelet.Db5 possess highest potential to extract the buried pattern
from vibration signatures from gear box [31]. Daubeschies wavelet is an orthogonal wavelet having compact support
as well as high vanishing moment. For an orthogonal wavelet, vanishing moment is coupled with compact support,
so a high vanishing moment ensures smooth and wide wavelet. A multi-level wavelet decomposition consisting of
five level was used to obtain the wavelet coefficients. The original signal was decomposed into approximation and
details using wavelet analysis as shown in Fig. 3. , Fig. 4. and Fig. 5. respectively. Approximation and details were
considered as output of a low pass and high pass filter respectively, whereas the approximation contains more useful
information about the signal being analysed than the details [32] .The approximation coefficients of level 5 denoted
as ‘a5’ were collected from wavelet analysis of vibration signal .Statistical analysis was performed on the
approximation coefficients and statistical feature such as mean, standard deviation, variance , root mean square(rms)
value, skewness , kurtosis, maximum value and minimum value were collected and these features will serve as input
to ANN.
 Sailendu Biswal et al. / Procedia Engineering 144 (2016) 305 – 311 309

Fig. 3. Signal Decomposition for Healthy Condition Fig. 4. Signal Decomposition for Faulty Gear with 1.5mm Depth
Crack

Fig. 5. Signal Decomposition for Faulty Gear with 2.5mm Depth Crack Fig. 6. Process Flow of Crack Size Estimation
310 Sailendu Biswal et al. / Procedia Engineering 144 (2016) 305 – 311

4. Results and Discussion

Machine learning algorithms such as Artificial Neural Network (ANN) are capable of predicting the faults based
on their size. The fault size was predicted using a feed forward back propagating ANN with five hidden layers.
Gradient descent with momentum algorithm was used to train the network. The use of gradient descent with
momentum algorithm causes ease attainment of global minima [33]. The ANN had considered mean squared error
between the neural network output and target values performance. The statistical features were used as input to the
ANN. The ANN was trained using 168 sets of data. In each iteration gradient of mean squared error was evaluated
using back propagation to adjust the weights and biases. The predicted values of fault depth were evaluated from the
ANN. The percentage of inaccuracy between predicted and actual fault size was calculated as shown in the Table 2.
below.

Table 2. Table Showing Predicted Fault Size and % of Error

Sl. No Actual Depth of Predicted Depth of % error

Crack(mm) Crack(mm)
1 1.5 1.323 11.8%
2 2.5 2.252 9.92%

The features obtained from wavelet transform of time domain vibration signal were used as input to the ANN and
the gear tooth root crack depth was predicted with 88.2% and 90.08% accuracy for 1.5mm depth and 2.5mm
respectively. The results obtained will be helpful for determining the severity of fault as well as remaining useful
life prediction.

4. Summary

The fault size for different test cases of gear root crack in a wind turbine test rig were predicted using vibration
analysis .The original signal was decomposed to get approximation and details using discrete wavelet analysis.
Statistical analysis of approximation coefficients were performed to extract features. Using features as input to the
ANN the fault size for gear root cracks were determined successfully.

Acknowledgements

This study was funded by RIG, BITS-Pilani, Hyderabad Campus, which is gratefully acknowledged.

References

[1] S. Soua, P. Van Lieshout, A. Perera, T.-H. Gan, and B. Bridge, Determination of the combined vibrational and acoustic emission signature
of a wind turbine gearbox and generator shaft in service as a pre-requisite for effective condition monitoring, Renewable Energy,51 (2013)
pp. 175-181.
[2] L. Yang, L. Xie, J. Wang, D. Wang, and Q. Miao, Current Progress on Wind Turbine Gearbox Condition Monitoring and Health
Evaluation,ASME 2013 International Mechanical Engineering Congress and Exposition,(2013) pp. V04BT04A047-V04BT04A047.
[3] Y. Lei and M. J. Zuo, Gear crack level identification based on weighted K nearest neighbor classification algorithm, Mechanical Systems
and Signal Processing, 23 (2009) 1535-1547.
 Sailendu Biswal et al. / Procedia Engineering 144 (2016) 305 – 311 311

[4] S. Choi and C. J. Li, Estimation of gear tooth transverse crack size from vibration by fusing selected gear condition indices, Measurement
Science and Technology, 17(2006) 2395.
[5] R. Stewart, Some useful data analysis techniques for gearbox diagnostics, University of Southampton, 1977.
[6] N. Ahamed, Y. Pandya, and A. Parey, Spur gear tooth root crack detection using time synchronous averaging under fluctuating speed,J.
Measurement, 52(2014) 1-11.
[7] H. J. Decker and D. G. Lewicki, Spiral bevel pinion crack detection in a helicopter gearbox, 2003.
[8] V. Sugumaran, G. Sabareesh, and K. Ramachandran, Fault diagnostics of roller bearing using kernel based neighborhood score multi-class
support vector machine, J. Expert Systems with Applications, 34(2008) 3090-3098.
[9] T. Fakhfakh, F. Chaari, and M. Haddar, Numerical and experimental analysis of a gear system with teeth defects, The International Journal
of Advanced Manufacturing Technology, 25(2005) 542-550.
[10] R. Randall, A new method of modeling gear faults, Journal of Mechanical Design, 104(1982) 259-267.
[11] J. Hines, D. Muench, J. Keller, and A. Garga, Effects of time-synchronous averaging implementations on HUMS features for UH-60A
planetary carrier cracking, Annual Forum Proceedings-American Helicopter Society, (2005) 218.
[12] W. Wang, Early detection of gear tooth cracking using the resonance demodulation technique, J. Mechanical Systems and Signal Processing,
15( 2001) 887-903.
[13] B. Tang, W. Liu, and T. Song, Wind turbine fault diagnosis based on Morlet wavelet transformation and Wigner-Ville distribution, J.
Renewable Energy, 35(2010) 2862-2866.
[14] Z. Feng, X. Chen, M. Liang, and F. Ma, Time–frequency demodulation analysis based on iterative generalized demodulation for fault
diagnosis of planetary gearbox under nonstationary conditions, J. Mechanical Systems and Signal Processing, 62(2015) 54-74.
[15] I. Antoniadou, G. Manson, N. Dervilis, T. Barszcz, W. Staszewski, and K. Worden, Condition monitoring of a wind turbine gearbox using
the empirical mode decomposition method and outlier analysis, in Sixth European Workshop on Structural Health Monitoring, Dresden,
Germany, 2012, 3–6 July.
[16] X. Zheng and H. Xu, Fault Diagnosis of Wind Turbine Rolling Bearing Based on Wavelet and Hilbert transforms.
[17] Q. Huang, D. Jiang, L. Hong, and Y. Ding, Application of wavelet neural networks on vibration fault diagnosis for wind turbine gearbox,
Advances in Neural Networks-ISNN 2008, ed: Springer, 2008, pp. 313-320.
[18] A. Belsak and J. Flasker, Wavelet analysis for gear crack identification, Engineering Failure Analysis, 16(2009) 1983-1990,.
[19] Y. Lu, J. Tang, and H. Luo, "Wind turbine gearbox fault detection using multiple sensors with features level data fusion," Journal of
Engineering for Gas Turbines and Power, vol. 134, p. 042501, 2012.
[20] S. Radhika, G. R. Sabareesh, G. Jagadanand, and V. Sugumaran, "Precise wavelet for current signature in 3Φ IM, J. Expert Syst. Appl.,
37(2010) 450-455.
[21] J. Cheng, D. Yu, J. Tang, and Y. Yang, Application of frequency family separation method based upon EMD and local Hilbert energy
spectrum method to gear fault diagnosis, J. Mechanism and Machine Theory, 43(2008) 712-723.
[22] T. Wu, J. Chen, and C. Wang, Characterization of gear faults in variable rotating speed using Hilbert-Huang Transform and instantaneous
dimensionless frequency normalization, J. Mechanical Systems and Signal Processing, 30(2012) 103-122.
[23] Y. Jiang, B. Tang, Y. Qin, and W. Liu, Feature extraction method of wind turbine based on adaptive Morlet wavelet and SVD, J. Renewable
energy, 36(2011) 2146-2153.
[24] H. Sun, Y. Zi, and Z. He, Wind turbine fault detection using multiwavelet denoising with the data-driven block threshold, J. Applied
Acoustics, 77(2014) 122-129.
[25] M. Kang, J. Kim, and J.-M. Kim, Reliable fault diagnosis for incipient low-speed bearings using fault feature analysis based on a binary bat
algorithm, J. Information Sciences, 294(2015) 423-438.
[26] S. Yang, W. Li, and C. Wang, The intelligent fault diagnosis of wind turbine gearbox based on artificial neural network,Condition
Monitoring and Diagnosis, 2008. CMD 2008. International Conference on, 2008, pp. 1327-1330.
[27] L. Wenyi, W. Zhenfeng, H. Jiguang, and W. Guangfeng,Wind turbine fault diagnosis method based on diagonal spectrum and clustering
binary tree SVM, J. Renewable Energy, 50(2013) 1-6.
[28] G. Sabareesh, V. Sugumaran, and K. Ramachandran, Fault diagnosis of a taper roller bearing through histogram features and proximal
support vector machines, IEEE international conference on signal and image processing, December, 2006.
[29] F. Chen, B. Tang, and R. Chen, A novel fault diagnosis model for gearbox based on wavelet support vector machine with immune genetic
algorithm, J. Measurement, 46(2013) 220-232.
[30] S. Biswal and G. Sabareesh, Design and development of a wind turbine test rig for condition monitoring studies, Industrial Instrumentation
and Control (ICIC), 2015 International Conference on, 2015, pp. 891-896.
[31] N. Saravanan and K. Ramachandran, Incipient gear box fault diagnosis using discrete wavelet transform (DWT) for feature extraction and
classification using artificial neural network (ANN), J. Expert Systems with Applications, 37(2010) 4168-4181.
[32] C. Smith, C. M. Akujuobi, P. Hamory, and K. Kloesel, An approach to vibration analysis using wavelets in an application of aircraft health
monitoring, J. Mechanical Systems and Signal Processing, 21(2007) 1255-1272.
[33] L. Mendelsohn, MATLAB Neural Network Toolbox, product review, Technical Analysis of Stocks & Commodities, vol. 12, 1994