Professional Documents
Culture Documents
Renewable Energy
journal homepage: www.elsevier.com/locate/renene
a r t i c l e i n f o a b s t r a c t
Article history: Wind power generation has been widely adopted due to its renewable nature and decreasing capital cost
Received 12 April 2021 per kW. However, existing equipment ages rapidly, leading to higher failure rates, greater operation and
Received in revised form maintenance costs, and worsening safety conditions, calling for improved condition monitoring and fault
23 January 2022
diagnosis for wind turbines. Past methods utilize physical models, but they are only successful in lab-
Accepted 10 April 2022
Available online 14 April 2022
oratory environments. As increasing data are becoming available, there are methods applying machine
learning without careful discrimination, leading to low accuracy. To solve this problem, first this paper
proposes to conduct unsupervised learning to understand data properties, e.g., structural density. Sub-
Index Terms:
Bearing fault diagnosis
sequently, the sensitivity analysis is conducted to extract the significant features and to avoid overfitting.
Feature extraction The sensitivity of various features that are characteristics of wind turbine bearings may vary significantly
Machine learning under different working conditions. During such a process, the piece-wise properties are studied to
Nonstationary signals improve supervised learning. By combining the properties of data and regression, a three-stage learning
Time-frequency analysis algorithm is proposed to refine and learn the most useful information for turbine bearing fault diagnosis.
The proposed framework is validated by using real data from diversified data sets for nonstationary
vibration signals of bearings.
© 2022 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/).
https://doi.org/10.1016/j.renene.2022.04.061
0960-1481/© 2022 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/).
B. Cui, Y. Weng and N. Zhang Renewable Energy 191 (2022) 987e997
To reduce the failure rate, it is crucial to apply reliable and cost- turbine gearbox [31]. Several studies were proposed to detect the
effective condition monitoring (CM) techniques so that the best wind turbine blades faults (such as cracks, erosion, and pitch angle
maintenance intervention can be determined in time to allow the twists) using ML [32].
machine to return to work as early as possible [5,6]. Additionally, For this work, the focus is to study the fault diagnosis of wind
improved CM leads to a better understanding of the design short- turbine bearing in wind energy. Schlechtingen et al. predicted the
comings, which can be used to improve WT design [7]. For this bearing failure of wind turbine generators by comparative analysis
purpose, CM systems are increasingly installed. Past research of neural network and regression using the data of bearing tem-
shows great potential for economic benefit. Such gains are proven perature [33]. They found that the nonlinear NN approaches
to be substantial and largely dependent on the fault detection rate outperform the regression models. Bangalore et al. used an artificial
[8,9]. Therefore, CM systems grew from 25% in 2012 to 36.1% in neural network approach to predict the early fault in the bearing of
2015. The key purpose of CM is to monitor and identify bearing a planetary gearbox in wind turbine by bearing temperature mea-
failures on wind turbines. For example, the gearbox and generator surement [34]. They conducted case studies in which their method
are studied due to the high percentage of replacement over the predicts a gearbox bearing fault a week before it was indicated by
wind farm life and to considerable equipment downtime, leading to the vibration based condition monitoring system. Bach-Andersen
substantial production loss. Recent studies show that bearings et al. use convolutional neural networks to classify expert-labeled
cause approximately 70% of gearbox downtime and 21%e70% of vibration data containing main bearing faults (rotor, planetary
generator downtime [10]. stage, and helical stage bearing faults) [35]. They find that the deep
In addition, corrective or preventive maintenance is inefficient CNN approach achieves a better performance on the test data than
to avoid these failure rates. Corrective maintenance is only per- a logistic classifier or a shallow multilayer perceptron.
formed after the misfunction has occurred, so it would not be able In this work, an approach is proposed to diagnose the bearings
to avoid any failures. Preventive maintenance is time-based way used in wind turbines based on customized ML framework. For
that faces two main challenges: under- and over-maintenance. In such a design, an observation is firstly done that the wind is char-
other words, the challenges are that the system performance is not acterized by its speed and direction, which are affected by several
effectively monitored, resulting in unexpected failures, or there is factors, e.g., geographic location, climate characteristics, height
an excessive number of maintenance activities, resulting in waste of above ground, and surface topography. For example, wind turbines
resources. Therefore, it is crucial to use condition-based mainte- interact with the wind, capturing a portion of its kinetic energy and
nance on bearings and to develop methods that can be used to converting it into useable energy. Therefore, the modeling is
diagnose faults early. In general, bearing diagnosis is a well- customized with these factors for knowledge enhancement. In the
recognized field of research; however, this is not the case for ma- process of fault diagnosis, the selection of characteristic parameters
chine operating under nonstationary loads, such as wind turbines significantly influences the diagnostic accuracy. Thus, a systematic
[11,12]. In the case of varying load/speed, vibrational signals method is presented to assess the prediction accuracy of each
generated by rolling element bearings are affected by operational feature to find the most effective combination for fault diagnosis of
factors, making the diagnosis difficult [13e15]. These difficulties wind turbine bearings based on neighborhood component analysis
arise from the variation in vibration-based diagnostic features (NCA). Various statistical features are used in this work from
caused mostly by load/speed variation (operation factors), low notable category differences in time-domain, frequency-domain,
energy of sought-after features, and low signal-to-noise levels and time-frequency. For this purpose, the hybrid-domain feature
[16,17]. Analysis of the signal from the main bearing is even more parameters defined on multiple domains are introduced to describe
difficult due to the very low rotational speed of the main shaft. the state of the rolling bearing. Furthermore, the fault diagnosis of
In recent years, machine learning (ML) approaches, such as wind turbine bearings is performed based on time-frequency
support vector machine (SVM), artificial neural networks (ANN), analysis and a supervised learning method.
and Deep Neural Networks (DNN), have been widely employed to To demonstrate the performance, the proposed method is vali-
resolve the complex nonlinear and uncertainty problems across dated on different data sets and conditions. The experimental re-
various disciplines. Successful applications of ML dealing with wind sults prove that the proposed method can effectively extract useful
energy industry can be found in many recent studies [18e20]. features accurately and efficiently for nonstationary vibrational
These researches can be divided into four general categories: pre- bearing signals. Further, the accuracy of fault diagnosis is signifi-
diction, design optimization, optimal control, and fault detection cantly higher than that of past methods.
[21]. The important parameters prediction of wind power genera- The remaining paper is organized as follows. Section II analyzes
tion based on ML, such as wind speed and wind power, has received the impacts of working conditions on wind turbine bearings. Sec-
extensive research [22e24]. Plenty of research has been done on tion III performs the data-driven feature extraction and selection
deploying ML to improve the design of wind energy systems, such based on neighborhood component analysis. Section IV develops a
as wind turbine design and wind farm design [25]. At the same bearing fault diagnosis method using ML. Section V conducts ex-
time, ML technologies are showing sustainable strategies in the periments for numerical validation. Section VI compares the per-
field of control, and are shown in related research on max power formance between the proposed method and regression-based
tracking control [26], pitch angle control [27], speed control [28] method. Section VII concludes the paper.
and so on during the process of wind power generation. Research
on fault diagnosis of wind energy using ML has also received 2. Impacts of working conditions on wind turbine bearings
continuous and widespread attention. Leahy et al. diagnosed and
predicted the wind turbine generator faults from SCADA data using Wind turbines (WTs) are exposed to a variable environment and
support vector machines [29]. High recall and low precision were operate 24 h a day, 7 days a week and experience rapid changes in
also found for the diagnostic and prognostic cases. Tang et al. built a temperature, air pressure, wind shear, wind speed, and total load.
classifier to identify gear, bearings, shaft and general transmission Due to these variations, the bearings of wind turbines undergo
failures based on manifold learning and Shannon Wavelet Support continuously changing loads. Therefore, the characteristics of the
Vector Machines [30]. A standard SVM with a radial basis function working conditions of wind turbine bearings are first studied to
kernel obtained an accuracy of 76%. Jiang et al. proposed the mul- understand the causes of bearing faults. The observations will
tiscale convolutional neural networks for fault diagnosis of wind subsequently be incorporated into our feature pool.
988
B. Cui, Y. Weng and N. Zhang Renewable Energy 191 (2022) 987e997
2.1. Effect of nonstationary mechanical load 2.3. Types and impacts of the bearing faults
Mechanical loads cause fatigue damage to several parts of wind According to the characteristics of vibration signals during
turbine bearings, thereby reducing the useful life of the system. operation, the faults of rolling bearings can be mainly divided into
There are two types of mechanical loads: static and dynamic. Static two categories: surface damage faults and wear faults.
loads result from mean wind speed and determine the absolute
mechanical load level. Dynamic loads, induced by turbulence and 1) Surface damage faults: During the operation of the bearing, the
gusts, are more related to the fatigue damage. rolling ball, the outer race, and the inner race, they are all sub-
jected to the action of periodic pulsating loads, resulting in
1) Mean wind speed: The mean wind speed determines the basic cyclically changing contact stress. When the number of stress
working load of a wind turbine. This speed is quantified by the cycles reaches a certain value, fatigue damage occurs on the
hourly mean of wind speeds. It is also used to determine the rolling ball or the working surfaces of the inner and outer race.
economic viability of a wind energy project. The probability This phenomenon caused by impact load and alternating stress,
distribution of the mean wind speed is predicted from mea- also known as bearing fatigue failure, is the main cause of rolling
surements collected during several years. The experimentally bearing faults. When bearing fatigue failure occurs, significant
obtained wind distribution can be approximated by a Weibull shock load, vibration, and noise will be generated.
distribution given by
For bearings in wind turbines, this type of fatigue failure is
k
further aggravated due to the non-linear and uncertain non-
Vm
k Vm k1 C stationary loads, especially sudden shock loads. When such fail-
pðVm Þ ¼ e ; (1)
C C ures are not detected at an early stage, the surface damage will
further increase, resulting in complete bearing invalid and even
where k and C are the shape and scale coefficients, respectively. equipment accidents. Therefore, it is important to diagnose the
These coefficients are adjusted to match the wind data at a fatigue failure of rolling bearings.
particular site. The Weibull probability function reveals that large
wind speeds rarely occur, whereas moderate winds are more 2) Wear faults: In the normal use of the bearing, the friction be-
frequent. tween the bearing balls, the outer and the inner race will lead to
the occurrence of wear failures. This kind of failure usually takes
2) Turbulence: Turbulence includes all wind speed fluctuations a long time to occur because it is a gradual process. In general,
with frequencies above the spectral gap. Therefore, turbulence wear faults will not cause bearing damage immediately, and the
contains all components in the range from seconds to minutes. degree of harm is far less than surface damage faults.
In general, turbulence exerts a minor impact on the annual
energy capture. However, it has major impacts on aerodynamic
loads and power quality. 3. Feature parameter selection for improving the diagnosis
accuracy
Statistical algorithms are used to analyze the time signals of Time-domain Frequency-
vibration sensors. The employed approach involves calculating feature parameters domain feature
statistical values for a trend analysis and learning more about the parameters
xmax ¼ max(xi) FC ¼ PH
shape of the vibration signal. It is assumed that vibration signals fk sðkÞ
PN p5 ¼ Pk¼1
taken from intact bearings are normally distributed since the sonic n¼1 xðnÞxðnÞ
H
k¼1 sðkÞ
P
pulses will be generated by the rolling elements and surfaces in a 2p N n¼1 xðnÞ
2
random way. If a fault starts to develop on the bearing, sonic pulses xpp ¼ max(xi) min(xi) FMS ¼ p6 ¼
PN sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 PH
n¼1 xðnÞ 2
with higher energy (spikes) can be found in the vibration signals. P k¼1 ðfk p5 Þ sðkÞ
4p 2 Nn¼1 xðnÞ
2
H
The presence of those spikes can be identified with the statistical rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffi vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 XN 2 FRMS ¼ FMS uPH 2
analysis algorithms as described below. Therefore, these statistical xrms ¼ u f k sðkÞ
N i¼1 i
x p7 ¼ t Pk¼1 H
algorithms provide an absolute measure of the condition of the k¼1 sðkÞ
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
respective bearing; there is no training phase required to learn the 1 XN FV ¼ FMS ðFC Þ2 uPH 4
u
x ¼ x f k sðkÞ
baseline values, which correspond to a fault-free condition of the N i¼1 i p8 ¼ tPk¼1 H 2
k¼1 f k sðkÞ
component. The disadvantages of these algorithms include the xrms PH p9 ¼
Sf ¼ k¼1 sðkÞ
relatively poor selectivity, which means that a change in the value 1 XN p1 ¼ PH 2
j xj H k¼1 fk sðkÞ
does not indicate a certain faulty component of a bearing, for N i¼1 i qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PH PH 4
example, the inner ring surface. k¼1 sðkÞ k¼1 f k sðkÞ
xmax p2 ¼ p6
The various statistical characteristics are used in this study from If ¼ p10 ¼
1 XN PH p5
j xj p1 Þ2
k¼1 ðsðkÞ
notable category differences in the time-domain, frequency- N i¼1 i
H1
domain, and time-frequency features. For example, the hybrid- xmax
CLf ¼ X pffiffiffiffiffiffiffi p3 ¼ p11 ¼
1 N PH PH
domain feature parameters defined in multiple domains are j jxi jj2 k¼1 ðsðkÞ p1 Þ3 3
k¼1 ðfk p5 Þ sðkÞ
N i¼1 pffiffiffiffiffiffi 3
introduced to describe the state of the rolling bearing. Hð p2 Þ Hp36
1 XN p4 ¼ p12 ¼
ðx xÞ4
i¼1 i PH 4 PH 4
1) Time-domain feature parameters: The crest factor has a value of Ck ¼ N 4 k¼1 ðsðkÞ p1 Þ k¼1 ðfk p5 Þ sðkÞ
xrms
approximately 3 for an ideal normally distributed signal. The Hðp2 Þ2 HF64
xmax
reason, the maximum value in a normally distributed signal (the CF ¼
xrms
so-called white noise) is nearly 3s, whereas the Root Mean where xi is a signal series for n ¼ 1, 2, where s(k) is a spectrum for k ¼ 1, 2, …, H;
Square (RMS) is 1s. If a fault is developing in a bearing, the spike …, N; N is the number of data H is the number of spectrum lines; and fk is
amplitudes, and therefore the maximum value of the signal, points. the frequency value of the kth spectrum
line.
increase. In an early occurrence of the faulty condition, the RMS
value will not be affected since the spike energy is small relative
to the overall energy of the signal. Therefore, the RMS value does
not change significantly, and the crest factor increases. The fault
detection and generation of condition monitoring system (CMS)
alarms are performed by trend analysis of the crest factor.
Although the calculation of the statistical values of vibrational
signals are independent of the measurement time, an appro-
priate time window should be measured, for example, 1 s. In any
case, the measurement time should be a multiple of the grid
frequency (20 ms in Europe) because this will help eliminate
signal distortions related to grid EMC effects.
2) Frequency-domain feature parameters: Energy values of the
intrinsic mode function (IMF) components are obtained from
vibration acceleration signals of the bearing data. To effectively
assess different fault locations and different degrees of perfor-
mance degradation of a rolling bearing with a unified assess-
ment index, a state assessment method based on the relative
compensation distance of multiple domain features and locally
linear embedding is adopted [39]. The time-domain, frequency-
domain and time-frequency features are listed in Table 1, and
Fig. 1 is a visualization of four parameters in the table. Fig. 1. Feature weights for classification.
3.2. Neighborhood component analysis (NCA) for feature selection Consider a multiclass classification problem with a training set
containing n observations
Neighborhood component analysis (NCA) is a nonparametric and
embedded method for selecting features with the goal of maxi- S ¼ fðxi ; yi Þ; i ¼ 1; 2; …; ng; (6)
mizing the prediction accuracy of regression and classification al-
gorithms. The method of NCA feature selection with regularization where xi 2Rp are the feature vectors, yi 2 {1, 2, …, c} are the class
learns feature weights for minimization of an objective function that labels, and c is the number of classes. The aim is to learn a classifier
measures the average leave-one-out classification or regression loss f : Rp /f1; 2; …; cg that accepts a feature vector and makes a pre-
over the training data. In this work, the NCA method is used to select diction f(x) of the true label y.
the feature parameters for classification of bearing faults. Consider a randomized classifier that
990
B. Cui, Y. Weng and N. Zhang Renewable Energy 191 (2022) 987e997
992
B. Cui, Y. Weng and N. Zhang Renewable Energy 191 (2022) 987e997
Fig. 2. Flow chart of bearing fault diagnosis based on feature extraction and ML.
993
B. Cui, Y. Weng and N. Zhang Renewable Energy 191 (2022) 987e997
Fig. 5. Classification of outer and inner race bearing faults with peak-peak values.
Method Classification rate The classification rate of bearing fault for each method by steps
SVM 75% was shown in Table 3. Note, it is easy to classify normal bearings
Naive Bayes classifier 70.83% and faulty bearings with classification rate 100% for all the pro-
KNN 87.5%
posed ML methods. This means that the proposed method works
ANN 83.33%
994
B. Cui, Y. Weng and N. Zhang Renewable Energy 191 (2022) 987e997
very well for fault detection, and can distinguish between normal into discrete labels, e.g., fault or no-fault. The comparison between
and faulty bearings at a very high classification rate. The classifi- the proposed ML methods and regression-based method is done
cation rate decreased in the Step 2 stage, but the rate of all methods finally.
was still above 80%. Next, the classification performance of the four
algorithms further decreased and showed significant differences 6.2. Regression-based modeling of bearing fault
for classifying the inner race fault and the outer race fault. The KNN
method achieved a good performance with classification rate above The public dataset obtained from the Bearing Data Center of
90% in the Step 3 stage. The classification rate of Naive Bayes Case Western Reserve University were used to compare the pro-
method is relatively poor, below 70%. The results of the average rate posed method with regression-based method [47]. For comparison,
show that the KNN method has the best classification performance, the regression model was trained with the same training and
which is slightly higher than that of the ANN algorithm. testing data used by the proposed ML model. In general, both
parametric modeling techniques (linear regression and nonlinear
5.5. Analysis and discussion regression) and non-parametric modeling techniques (ANN) are
employed to predict the bearing fault. The bearing vibration signal
By comparing the classification accuracy in Tables 2 and 3, it after the feature extraction and selection process is used as the
shows that the classification performance of ANN and KNN input data of the regression model. The status of bearing is divided
methods can be improved by using the step-by-step classification into normal with label 0, ball fault with label 1, inner race fault with
method. At the same time, the KNN method produces the best label 2, and outer race fault with label 3. The label value of the
classification performance in both its entirety and step-by-step bearing is used as the output data of the regression model. The
classification. Thus the KNN algorithm can provide good results training of the regression model was performed in MATLAB version
for bearing fault diagnosis. 2019a. Subsequently, data samples selected randomly are input into
the trained regression model to predict the corresponding bearing
state.
6. Comparison between the proposed method and regression
based method
6.3. Evaluation criteria
6.1. Regression-based fault detection method
To analyze the difference of each model clearly, the evaluation
Regression based fault detection identifies how signals and criteria are introduced for comparison. Mean absolute error (MAE)
features are related to a state output in different components. This and root mean square error (RMSE) are used to evaluate the per-
relationship is captured by fitting regression models when the formance. The mathematical expressions are as follows:
system is in a healthy state. Therefore, this regression model is also
1 Xn
known as the normal behavior model (NBM). When new data is MAE ¼ ^t r
ryt y (28)
n t¼1
obtained, the NBM prediction value of state variable is compared
with the measured values. The difference between the predicted
and actual measured values is called residual, and the residual are
used for monitoring the condition. Residuals that exceed a given
Table 4
error threshold indicate malfunctioning.
MAE and RMSE for the prediction methods.
For the fault detection in wind turbine, the residuals indicate
how close the wind turbine's behavior resembles the one observed Method MAE RMSE
during training of the NBM. It is reasonable to assume that in a Linear regression 0.5426 0.6574
normal functioning wind turbine, the residuals are homogeneously Nonlinear regression 0.5052 0.6258
randomly distributed. Therefore, trends or excessively large re- ANN 0.5589 0.6658
Table 3
Classification rate of bearing fault for each method by steps.
995
B. Cui, Y. Weng and N. Zhang Renewable Energy 191 (2022) 987e997
Table 5
Comparison between the actual value and predicted value of bearing state produced with nonlinear regression.
Actual value 1 1 1 1 0 2 2 2
Predicted value 1.7699 1.4837 1.8018 1.2616 0.5165 2.6506 2.5833 2.1380
Actual value 2 2 3 3 3 3 3 3
Predicted value 2.6871 2.1910 2.5462 2.9518 2.5096 1.2209 2.2482 2.7204
Actual value 3 0 0 0 1 1 2 0
Predicted value 2.9060 1.0983 0.3385 0.4730 1.3737 0.8637 2.1730 0.5525
The trained regression models were evaluated with MAE and The authors declare that they have no known competing
RMSE criteria, and the results were listed in Table 4. Notice that the financial interests or personal relationships that could have
MAE and RMSE values corresponding to the nonlinear regression appeared to influence the work reported in this paper.
model are the smallest, that is, the accuracy of the nonlinear
regression model is the highest. The labels of actual values and the
Acknowledgements
predictive values are shown in Fig. 8. It can be seen that the actual
value of the label and the predictive value generated by the KNN
This work was supported in part by the Department of Energy
method are both discrete data. The predicted value of the KNN
under grants DE-AR00001858-1631 and DE-EE0009355, the Na-
produced three prediction errors, and the others were consistent
tional Science Foundation (NSF) under the grants ECCS-1810537
with the actual value. The label that is predicted based on the
and ECCS-2048288.
regression method is a continuous label. A threshold needs to be
given to determine the state of the bearing.
The comparison between the actual value and predicted value of References
bearing label produced with nonlinear regression is shown in [1] Q. Xu, C. Kang, N. Zhang, Y. Ding, Q. Xia, R. Sun, J. Xu, A probabilistic method
Table 5. Currently, how to determine the threshold is the key to for determining grid-accommodable wind power capacity based on multi-
fault diagnosis. For example, if the threshold for outer race faults is scenario system operation simulation, IEEE Trans. Smart Grid 7 (1) (Jan 2016)
400e409.
set to 2.5, five bearings with outer race fault will be correctly
[2] E. Gonzalez, B. Stephen, D. Infield, J.J. Melero, Using high-frequency scada data
identified. At this time, the detection ratio of outer race fault was for wind turbine performance monitoring: a sensitivity study, Renew. Energy
71.4%. But unfortunately, the other three bearings with inner race 131 (Feb 2019) 841e853.
anomalies were mistakenly believed to have outer race fault. [3] J. Carroll, A. McDonald, I. Dinwoodie, D. McMillan, M. Revie, I. Lazakis, Avail-
ability, operation and maintenance costs of offshore wind turbines with
Obviously, the design of thresholds heavily depends on the de- different drive train configurations, Wind Energy 20 (2) (Jul 2017) 361e378.
signer's experience and knowledge. At the same time, the threshold [4] D. Milborrow, Operation and maintenance costs compared and revealed,
setting will also change for different wind farm environments. The Wind Stats 19 (3) (2006) 3.
[5] W. Liu, B. Tang, J. Han, X. Lu, N. Hu, Z. He, The structure healthy condition
proposed ML method can automatically complete the classification monitoring and fault diagnosis methods in wind turbines: a review, Renew.
of faults without the need to set thresholds and provides a diag- Sustain. Energy Rev. 44 (Jan 2015) 466e472.
nostic accuracy about 90% that is produced by KNN. Therefore, the [6] G. Pedro, T. Mark, P. Pinar, P. Mayorkinos, Condition monitoring of wind
turbines: techniques and methods, Renew. Energy 46 (2012) 169e178.
proposed method outperforms the regression-based method for [7] P. Tchakoua, R. Wamkeue, M. Ouhrouche, F. Slaoui-Hasnaoui, T. Tameghe,
bearing fault diagnosis. G. Ekemb, Wind turbine condition monitoring: state-of-the-art review, new
trends, and future challenger, Energies 7 (4) (Apr 2014) 2595e2630.
[8] K. Fischer, F. Besnard, L. Bertling, Reliability-Centered Maintenance for Wind
Turbines Based on Statistical Analysis and Practical Experience 27 (1) (Mar
7. Conclusion 2012) 184e195.
[9] A. May, D. McMillan, S. Thons, Economic analysis of condition monitoring
systems for offshore wind turbine sub-systems, IET Renew. Power Gener. 9 (8)
This paper proposes a novel scheme for wind turbine bearing (Nov 2015) 900e907.
fault diagnosis based on data mining and ML. The proposed method [10] H.D.M. de Azevedo, A.M. Araújo, N. Bouchonneau, A review of wind turbine
works directly on original data obtained from vibration signals of bearing condition monitoring: state of the art and challenges, Renew. Sustain.
Energy Rev. 56 (Apr 2016) 368e379.
bearings under different operation conditions. There are two main [11] Z. Feng, M. Liang, Fault diagnosis of wind turbine planetary gearbox under
steps in the proposed method. (1) A systematic method is pre- nonstationary conditions via adaptive optimal kernel time-frequency analysis,
sented to assess the prediction accuracy of each characteristic to Renew. Energy 66 (Jan 2014) 468e477.
[12] X. Chen, Z. Feng, Iterative generalized time-frequency reassignment for
find the combination of the most effective features for fault diag-
planetary gearbox fault diagnosis under nonstationary conditions, Mech. Syst.
nosis of wind turbine bearings. (2) ML is utilized to perform the Signal Process. 80 (May 2016) 429e444.
fault diagnosis of wind turbine bearings. The experiments [13] Z. Feng, W. Zhu, D. Zhang, Time-frequency demodulation analysis via vold-
demonstrate that this method can effectively extract features from kalman filter for wind turbine planetary gearbox fault diagnosis under
nonstationary speeds, Mech. Syst. Signal Process. 128 (Apr 2019) 93e109.
nonstationary signals, obtaining excellent results in the fault [14] I. Vamsi, G. Sabareesh, P. Penumakala, Comparison of condition monitoring
diagnosis of wind turbine bearings. techniques in assessing fault severity for a wind turbine gearbox under non-
996
B. Cui, Y. Weng and N. Zhang Renewable Energy 191 (2022) 987e997
stationary loading, Mech. Syst. Signal Process. 124 (Jan 2019) 1e20. [31] G. Jiang, H. He, J. Yan, P. Xie, Multiscale convolutional neural networks for
[15] R.U. Maheswari, R. Umamaheswari, Trends in non-stationary signal process- fault diagnosis of wind turbine gearbox, IEEE Trans. Ind. Electron. 66 (4)
ing techniques applied to vibration analysis of wind turbine drive train e a (2018) 3196e3207.
contemporary survey, Mech. Syst. Signal Process. 85 (2017) 296e311. [32] K. Chandrasekhar, N. Stevanovic, E.J. Cross, N. Dervilis, K. Worden, Damage
[16] Y. Guo, J. Keller, Investigation of High-Speed Shaft Bearing Loads in Wind detection in operational wind turbine blades using a new approach based on
Turbine Gearboxes through Dynamometer Testing, vol. 21, Wind energy, machine learning, Renew. Energy 168 (2021) 1249e1264.
2018, pp. 139e150. [33] M. Schlechtingen, I.F. Santos, Comparative analysis of neural network and
[17] E. Marino, A. Giusti, L. Manuel, Offshore wind turbine fatigue loads: the in- regression based condition monitoring approaches for wind turbine fault
fluence of alternative wave modeling for different turbulent and mean winds, detection, Mech. Syst. Signal Process. 25 (5) (2011) 1849e1875.
Renew. Energy 102 (2017) 157e169. [34] P. Bangalore, L.B. Tjernberg, An artificial neural network approach for early
[18] G. Helbing, M. Ritter, Deep learning for fault detection in wind turbines, fault detection of gearbox bearings, IEEE Trans. Smart Grid 6 (2) (2015)
Renew. Sustain. Energy Rev. 98 (2018) 189e198. 980e987.
[19] A. Stetco, F. Dinmohammadi, X. Zhao, V. Robu, D. Flynn, M. Barnes, J. Keane, [35] M. Bach-Andersen, B. Rømer-Odgaard, O. Winther, Deep learning for auto-
G. Nenadic, Machine learning methods for wind turbine condition moni- mated drivetrain fault detection, Wind Energy 21 (1) (2018) 29e41.
toring: a review, Renew. Energy 133 (Oct 2019) 620e635. [36] F.D. Bianchi, H.D. Battista, R.J. Mantz, Wind Turbine Control Systems. Plus 0.5em
[20] Y. Wang, R. Zou, F. Liu, L. Zhang, Q. Liu, A review of wind speed and wind Minus 0.4em, Springer, 2007.
power forecasting with deep neural networks, Appl. Energy 304 (2021). [37] I. Guyon, A. Elisseeff, An introduction to variable and feature selection,
[21] F. Elyasichamazkoti, A. Khajehpoor, Application of machine learning for wind J. Mach. Learn. Res. 3 (2003) 1157e1182.
energy from design to energy-water nexus: a survey, Energy Nexus 2 (2021). [38] P. Marti-Puig, A. Blanco-M, J. Cardenas, J. Cusido, J. Sole-Casals, Effects of the
[22] M.M. Nezhad, A. Heydari, E. Pirshayan, D. Groppi, D.A. Garcia, A novel fore- Pre-processing Algorithms in Fault Diagnosis of Wind Turbines, Environ-
casting model for wind speed assessment using sentinel family satellites mental Modelling and Software, Feb 2018, 0e1.
images and machine learning method, Renew. Energy 179 (Aug 2021) [39] S. Kang, D. Ma, Y. Wang, C. Lan, Q. Chen, V. Mikulovich, Method of assessing
2198e2211. the state of a rolling bearing based on the relative compensation distance of
[23] A. Marvuglia, A. Messineo, Monitoring of wind farms' power curves using multiple-domain features and locally linear embedding, Mech. Syst. Signal
machine learning techniques, Appl. Energy 98 (2012) 574e583. Process. 86 (Oct 2017) 40e57.
[24] M. Sharifzadeh, A. Sikinioti-Lock, N. Shah, Machine-learning methods for in- [40] W. Yang, K. Wang, W. Zuo, Neighborhood component feature selection for
tegrated renewable power generation: a comparative study of artificial neural high-dimensional data, J. Comput. 7 (1) (Jan 2012) 161e168.
networks, support vector regression, and Gaussian process regression, Renew. [41] R. Liu, B. Yang, E. Zio, X. Chen, Artificial intelligence for fault diagnosis of
Sustain. Energy Rev. 108 (2019) 513e538. rotating machinery: a review, Mech. Syst. Signal Process. 108 (Feb 2018)
[25] B. Cheng, J. Du, Y. Yao, Machine learning methods to assist structure design 33e47.
and optimization of dual darrieus wind turbines, Energy (2021), https:// [42] R. Zhao, R. Yan, Z. Chen, K. Mao, P. Wang, R.X. Gao, Deep learning and its
doi.org/10.1016/j.energy.2021.122643. applications to machine health monitoring, Mech. Syst. Signal Process. 115
[26] S. Sabzevari, A. Karimpour, M. Monfared, M.N. Sistani, Mppt control of wind (Jan 2019) 213e237.
turbines by direct adaptive fuzzy-pi controller and using ann-pso wind speed [43] W. Liu, Z. Wang, J. Han, G. Wang, Wind turbine fault diagnosis method based
estimator, J. Renew. Sustain. Energy 9 (2017), 013302. on diagonal spectrum and clustering binary tree svm, Renew. Energy 50 (Feb
[27] G. Abdalrahman, W. Melek, F.-S. Lien, Pitch angle control for a small-scale 2013) 1e6.
darrieus vertical axis wind turbine with straight blades (h-type vawt), [44] Y. Wang, S. Kang, Y. Jiang, G. Yang, L. Song, V. Mikulovich, Classification of
Renew. Energy 114 (2017) 1353e1362, 013302. fault location and the degree of performance degradation of a rolling bearing
[28] H.M. Hasanien, S.M. Muyeen, Speed control of grid-connected switched based on an improved hyper-sphere-structured multi-class support vector
reluctance generator driven by variable speed wind turbine using adaptive machine, Mech. Syst. Signal Process. 29 (May 2012) 404e414.
neural network controller, Elec. Power Syst. Res. 84 (2012) 206e213. [45] V. Muralidharan, V. Sugumaran, A comparative study of naive bayes classifier
[29] K. Leahy, R. Hu, I. Konstantakopoulos, C. Spanos, A. Agogino, D. O'Sullivan, and bayes net classifier for fault diagnosis of monoblock centrifugal pump
Diagnosing and predicting wind turbine faults from scada data using support using wavelet analysis, Appl. Soft Comput. 12 (8) (2012) 2023e2029.
vector machines, Int. J. Prognostics Health Manag. 9 (2018) 1e11. [46] D. He, R. Li, J. Zhu, Plastic bearing fault diagnosis based on a two-step data
[30] B. Tang, T. Song, F. Li, L. Deng, Fault diagnosis for a wind turbine transmission mining approach, IEEE Trans. Ind. Electron. 60 (8) (2013) 3429e3440.
system based on manifold learning and shannon wavelet support vector [47] Bearing Data Center, Case Western Reserve University [Online]. Available:
machine, Renew. Energy 62 (2014) 1e9. https://engineering.case.edu/bearingdatacenter.
997