ORIGINAL RESEARCH

published: 19 November 2021

doi: 10.3389/fenrg.2021.799039

Monitoring and Identifying Wind

Turbine Generator Bearing Faults
Using Deep Belief Network and EWMA
Control Charts
Huajin Li 1,2, Jiahao Deng 3*, Shuang Yuan 2, Peng Feng 1 and Dimuthu D. K. Arachchige 3
1
School of Architecture and Civil Engineering, Chengdu University, Chengdu, China, 2State Key Laboratory of Geo-hazard
Prevention and Geo-environment Protection, Chengdu University of Technology, Chengdu, China, 3College of Computing and
Digital Media, DePaul University, Chicago, IL, United States

Wind turbines are widely installed as the new source of cleaner energy production.
Dynamic and random stress imposed on the generator bearing of a wind turbine may
lead to overheating and failure. In this paper, a data-driven approach for condition
monitoring of generator bearings using temporal temperature data is presented. Four
algorithms, the support vector regression machine, neural network, extreme learning
machine, and the deep belief network are applied to model the bearing behavior.
Comparative analysis of the models has demonstrated that the deep belief network is
Edited by:
Xun Shen, most accurate. It has been observed that the bearing failure is preceded by a change in the
Tokyo University of Agriculture and prediction error of bearing temperature. An exponentially-weighted moving average
Technology, Japan
(EWMA) control chart is deployed to trend the error. Then a binary vector containing
Reviewed by:
Zhiyu Sun,
the abnormal errors and the normal residuals are generated for classifying failures. LS-SVM
The University of Iowa, United States based classification models are developed to classify the fault bearings and the normal
Heming Huang, ones. The proposed approach has been validated with the data collected from 11 wind
Wuhan University, China
turbines.
*Correspondence:
Jiahao Deng Keywords: bearing failure, condition monitoring, deep belief network, EWMA control chart, SCADA data analysis
[email protected]

Specialty section: 1 INTRODUCTION

This article was submitted to
Wind Energy, Wind energy is the fastest growing form of renewable energy. Continuous operations in all
a section of the journal environmental conditions contribute to failures of wind turbine components, assemblies, and
Frontiers in Energy Research systems. The generator of a wind turbine is one of the most failure-prone assemblies due to the
Received: 21 October 2021 variable loads (Kusiak and Verma, 2012). Bearing failures account for more than 40% of the overall
Accepted: 01 November 2021 wind turbine generator failures leading to unexpected energy losses (Tavner et al., 2012). Hence, a
Published: 19 November 2021 solution for effective condition monitoring of generator bearings and early identification of failure
Citation: symptoms is needed.
Li H, Deng J, Yuan S, Feng P and Deteriorating performance of a generator bearing manifests itself on abnormal changes of the
Arachchige DDK (2021) Monitoring vibration signal, torque, and bearing temperature (Yang et al., 2017; Feng et al., 2020). Vibration
and Identifying Wind Turbine
analysis and data-driven approaches have been applied for condition monitoring of generator
Generator Bearing Faults Using Deep
Belief Network and EWMA
bearings (Yang et al., 2018). The frequently used classical vibration analysis approaches include
Control Charts. Fourier transformation (Klein et al., 2001), wavelet transform (Yan et al., 2014), Hilbert-Huang
Front. Energy Res. 9:799039. transform (Peng et al., 2005; Huang and Wu 2008), and empirical model decomposition (EMD)
doi: 10.3389/fenrg.2021.799039 (Huang et al., 2008). Other models have been developed. Teng et al. (2016) utilized a complex

Frontiers in Energy Research | www.frontiersin.org 1 November 2021 | Volume 9 | Article 799039

Li et al. Original Research About Energy

Gaussian wavelet to obtain the multi-scale enveloping

spectrogram for extracting weak features. Lei et al. (2013)
applied an ant colony algorithm to form adaptive stochastic
resonance method for failure detection. Peeters et al. (2018)
integrated automated spectrum editing procedure, band-pass
filtering and envelop analysis to detect bearing failures based
on the vibration signal. Vibration analysis approaches are
valuable in monitoring and diagnosis of generator bearing
failures. However, high frequency data from multiple
FIGURE 1 | The restricted Boltzmann machine.
vibration sensors is needed to perform such analysis.
However, at present high frequency data is not available
from industrial turbines due to the excessive cost and data
sharing practices.
Most commercial wind turbines are equipped with the SCADA data indicates that that bearing failure is preceded by
supervisory control and data acquisition (SCADA) systems the error shift. The exponentially weighted moving average
collecting data that can be used to model behavior of (EWMA) control chart is applied to monitor the error shift. A
generator bearings. Kusiak and Verma. (2012) applied a temporal binary vector is generated in real-time, and a final
neural network to model bearing temperature for failure failure classification model is developed. The benefits of the
prediction and identification. Guo. (2012) introduced proposed approach are demonstrated with computational
nonlinear state estimate technique (NSET) for temperature- experiments.
based failure detection. Yang et al. (2013) applied correlation
analysis and quantitative assessment based on the SCADA
data. The published literature indicates that the data-driven
methods provide robust bearing monitoring solutions for 2 RESEARCH METHODOLOGY
wind turbines.
Deep learning is a recent addition to the modeling suite The use of deep-learning algorithms in prediction and
with promising applications in multiple domains (Ouyang condition monitoring is growing (LeCun et al., 2015). Deep
et al., 2017; Sun et al., 2020a; Sun et al., 2020b; Shen et al., learning originates from the research in neural networks.
2021a; He et al., 2018; Li et al., 2018). The deep learning Deep-learning algorithms avoid the local optima dilemma
algorithms are capable of extracting in-depth features and and contains superior power in extracting globally robust
patterns within the training dataset (Gritsenko et al., 2017; features from the dataset (Deng and Yu 2013; Qiu et al., 2017).
Ouyang et al., 2019; Li et al., 2020; Shen et al., 2021b; Shen and
Raksincharoensak 2021). Within the wind energy sector, it 2.1 Deep Belief Network
has been applied in the prediction tasks of wind speed (Hu In this research, a deep belief network (DBN) is applied to model
et al., 2016), wind power (Wang et al., 2017), and wind the generator bearing temperature. Proposed by Hinton et al.
direction (Wang et al., 2016a; Li et al., 2021a). Extensive (2006), the classical DBN algorithm multilayers of restricted
research has also been published using the deep-learning Boltzmann machines (RBMs) and a logistic regression layer
approaches: Wang et al. (2016b) developed deep auto- (Wang et al., 2016c).
encoders to compress the time-series SCADA dataset and The restricted Boltzmann machine (RBM) is a commonly used
the blade breakages are extracted from the deep-learned generative stochastic neural network (Hinton et al., 2006). It
features. Yang et al. (2018) applied stacked Restricted includes a visible layer of binary-valued neurons and a hidden
Boltzmann Machines (RBMs) to capture the system-wide layer of Boolean neurons (see Figure 1). The connection between
patterns and then performed condition monitoring with the hidden layer and the visible layer is bidirectional and
promising results. Bach-Andersen et al. (2018) selected 1- symmetrical. There are no inter-connections between neurons
dimensional convolutional neural networks (CNN) to extract in the same layer.
temporal features to classify failures of gearbox bearings. Training a single restricted Boltzmann machine (RBM)
Overall, deep-learning algorithms support development of involves the weight matrix between the two layers. The
higher complexity models. configuration of weight matrix is based on the energy function
In this research, a deep-learning approach is explored to expressed in Eq. 1 (Wang et al., 2016c). The joint distribution of a
monitor generator bearings. A deep belief network (DBN) visible layer vector and the hidden layer vector is expressed in Eq.
integrated with back-propagation (B-P) fine-tuning and layer- 2 (Hinton et al., 2006). The activation functions of neurons in the
wise training is developed to model normal generator bearing visible and hidden layer are presented in Eqs 3, 4 (Hinton et al.,
temperature using SCADA data. Four data-driven models 2006
predicting normal bearing temperature are constructed.
Their performance is assessed with the absolute percentage nv nh nv nh

error (APE), the mean absolute percentage error (MAPE) and E(v, h)  −
ai vi −
bj hj −

hj wj,i vi , (1)
i1 j1 i1 j1
the root mean square error (RMSE). The analysis of industrial

Frontiers in Energy Research | www.frontiersin.org 2 November 2021 | Volume 9 | Article 799039

Li et al. Original Research About Energy

L(a, b, w) 
log P(v, h), (5)

where: a is the bias vector of the visible layer; b is the bias vector of
the hidden layer; and w is the weight matrix between the two
layers. The parameters of the objective function (a, b, w) are
updated based on the gradients of the function expressed in Eqs
6–8. The updating rules are formulated in Eqs 9–11 (Hinton
et al., 2006).
zlog P(v, h)
 〈vi hi 〉P(h|v) − 〈vi hi 〉recon, (6)
zwj,i
zlog P(v, h)
 〈vi 〉P(h|v) − 〈vi 〉recon, (7)
zai
zlog P(v, h)
 〈hi 〉P(h|v) − 〈hi 〉recon (8)
zbj
FIGURE 2 | Architecture of the deep belief network.
wi+1  wi + η〈vi hi 〉P(h|v) − 〈vi hi 〉recon , (9)
bi+1  bi + η〈vi 〉P(h|v) − 〈vi 〉recon , (10)
e−E(v,h) ai+1  ai + η〈hi 〉P(h|v) − 〈hi 〉recon , (11)
P(v, h)  , (2)
  e−E(v,h)
v h where: η is the learning rate; 〈〉P(h|v) is the expectation of the
nh conditional distribution with respect to the original input data;
P(vi  1|h)  sig⎛ ⎠,
⎝αi +
wj,i hj ⎞ (3) 〈〉recon is the i-step reconstructed distribution obtained by the
j1 alternating Gibbs sampling scheme. The expectation of the
nv reconstructed distribution is computed following the rules of
P(hi  1|v)  sig⎛ ⎠,
⎝bj +
wj,i vi ⎞ (4) contrastive divergence (Hinton, 2002).
i1

where: vi is the number of neurons in the visible layer; hi is the 2.3 Data-Driven Algorithms
number of Boolean neurons within the hidden layer; wj,i is the Performance of the deep belief network (DBN) is compared
weight matrix between the visible layer and hidden layer; ai and bi with three algorithms, support vector regression machine
are the biases of the two layers; and sig() denotes the logistic (SVR), neural network (NN), and extreme learning
sigmoid function. Hence, the weight matrix and the layer biases machine (ELM).
are obtained in a layer-wise unsupervised pre-training described The support vector regression machine (SVR) is considered in
in the Section 2.2. this study includes a Gaussian kernel function (Drucker et al.,
1997). The values of the model parameters (c and γ) are selected
2.2 Layer-wise Pre-training based on the 10-fold cross-validation. The neural network (NN)
A deep belief network (DBN) includes multiple layers of contains two hidden layers. By testing on a small portion of the
restricted Boltzmann machines (RBMs) (Ouyang et al., 2019). training data, the sigmoid activation function is selected based on
Figure 2 shows the architecture of the proposed DBN. The first the satisfactory performance. The extreme learning machine
RBM of the DBN model consisting of a visible and a hidden layer (ELM) algorithm (Liang et al., 2006) is utilized to model the
(hidden layer 1) is pre-trained as an independent RBM. Then, normal bearing temperature. As a single-hidden layer feed-
the weight matrix of the first RBM is computed. The output of forward network, the ELM learning model is expressed in Eqs
the first RBM becomes the input to the second RBM that 12, 13 (Liang et al., 2006).
includes two layers. The first layer (hidden layer 1) is treated
fL
xj   oj , ∀j, (12)
as a visible layer of the second RBM while the second
L
layer (hidden layer 2) is treated as the hidden layer. The

βi G
ai , bi , xj   tj , j  1, 2, ..., N, (13)
weight matrix of the second RBM is computed. Hence, the i1
weight matrices between the remaining hidden layers are
obtained iteratively. where: xj represents the input parameters; oj represents the
Training each restricted Boltzmann machine (RBM) is predicted output values; fL() is the non-linear function
accomplished with a stochastic gradient descent method representing the ELM algorithm; ai is the weight vector
(Hinton et al., 2006). Based on vector Eq. 2 of the joint connecting the ith hidden node and the input nodes; bi is the
distribution function between the visible and hidden layer, the threshold of the ith hidden node; βi is the weight vector
objective function of the stochastic gradient descend method is connecting the ith hidden node and the output nodes; and tj is
expressed in Eq. 5 (Wang et al., 2016c). the actual output value.

Frontiers in Energy Research | www.frontiersin.org 3 November 2021 | Volume 9 | Article 799039

Li et al. Original Research About Energy


λ1 − (1 − λ)2t 
UCL(t)  μAPE + L p σ APE , (18)
(2 − λ)N

λ1 − (1 − λ)2t 
LCL(t)  μAPE − L p σ APE , (19)
(2 − λ)N
where: µAPE is the mean of absolute percentage error (APE); σ APE
is the standard deviation of APE; N denotes number of samples.
According to Horng Shiau and Ya-Chen. (2005), the value of the
parameter L is commonly set to 3 and λ is usually set to 0.2.

2.6 Binary Vectors Generated by Control

Chart
The EWMA control charts used statistical thresholds to label the
prediction error (residuals) as normal and abnormal. The normal
FIGURE 3 | Vectorization of residuals using EWMA control charts. residual usually denotes the bearing temperature is within the
normal range and the wind turbine is at healthy status. On the
other hand, the abnormal values often indicate abnormal bearing
temperature change and it can be the warning signal for bearing
failures. Hence, in this research, the normal and abnormal
2.4 Performance Evaluation Metrics residuals identified by the EWMA control charts are
To assess prediction accuracy of the deep belief network, three
transformed into binary vectors as described in Figure 3 as
performance evaluation metrics are computed: the absolute
follows.
percentage error (APE) Eq. 14, the mean absolute percentage
According to Figure 3, the statistical thresholds classified the
error (MAPE) Eq. 15, and the root mean square error (RMSE)
residuals into normal and abnormal ones. Each data point can be
Eq. 16.
simply labeled as 0 (normal) and 1 (abnormal). Hence, the binary
 
oj − tj  vectors can be generated in real-time and be utilized as the
APE    p 100%, (14)
 tj  inputs for the final classification models introduced in the
  Section 2.7.
1 N oj − tj 
MAPE 
 , (15)
N i1  tj  2.7 Classification Models

 Using the real-time vectors generated by the EWMA control
1 N  2 charts, the final failure classification models are constructed in
RMSE 
oj − tj  , (16)
Nj1 this research. Here, the dimension of the input vector is
determined as 20 which represents all normal/abnormal
where: oj is the jth predicted generator bearing temperature; tj is prediction residuals of bearing temperatures. In total of four
the jth actual generator bearing temperature; N denotes the state-of-art machine learning algorithms including support
number of data points. vector machine (SVM), least-square support vector machine
(LS-SVM), extreme learning machine (ELM), and kernel-based
extreme learning machine (KELM) are selected to classify the
2.5 Exponentially Weighted Moving Average vectors representing generator bearing failures and vectors from
Control Chart normal bearing behaviors.
The increasing value of the prediction bearing temperature error The SVM is the state-of-art supervised learning algorithm
of a data-driven model reflects deterioration of the generator used for classification and function approximation (Cherkassky
bearing conditions. In this research, an exponentially weighted and Ma, 2004). It is based on kernel functions and it avoids the
moving average (EWMA) (Jones et al., 2001) control chart is difficulty of using linear functions in the high dimensional
applied to monitor the error. The weighted average of the past parameter space, and the optimization problem is transformed
bearing temperatures reduces the noise and allows detecting small into a dual convex quadratic programming problem.
process shifts. The LS-SVM is developed based on statistical theory and
To compute the upper and lower confidence limits of the considered as the improved version of SVM (Zhu et al., 2018).
EWMA control chart, the EWMAt is obtained from Eq. 17 Compared with the vanilla SVM, the LS-SVM modifies the
(Wang et al., 2016b). The upper and lower confidence limits inequality constraint in the SVM to the equality constraint.
can be computed from Eqs 18, 19 (Horng Shiau and Ya-Chen Meanwhile, the training error square is used to replace the
2005). slack variable in order to transform quadratic programming
problem into the linear equation problem for greatly
EWMAt  λ p APEt + (1 − λ) p EWMAt−1 , (17) improving the speed and accuracy of model parameters. The

Frontiers in Energy Research | www.frontiersin.org 4 November 2021 | Volume 9 | Article 799039

Li et al. Original Research About Energy

TABLE 1 | Dataset description. TABLE 2 | Dataset description.

Turbine id. Bearing temperature Bearing failure Failure times Parameter BTA WGS RFA
Min (°C) Max (°C)
Generator phase-1 winding temperature 100 10 0.1
Generator phase-2 winding temperature 98 10 0.1
A 14 68 No
Generator air temperature 97 10 0.09
B 11 87 Yes 8
Generator rear temperature 96 9 0.09
C 0 63 No
Generator phase-3 winding temperature 96 9 0.1
D 3 68 No
Water cooler temperature 91 8 0.11
E 13 71 No
Phase compensation panel temperature 78 7 0.07
F 13 69 No
Nacelle temperature 78 6 0.05
G 0 73 No
H 6 90 Yes 2
I 9 86 Yes 5
J 14 71 No
K 7 85 Yes 7
(Kohavi and John, 1997), boosting-tree algorithm (BTA)
(Sbihi, 2007), and the relief algorithm (RA) (Liu et al., 2018)
were applied to select the most relevant parameters for predicting
LS-SVM has the unique superiority in dealing with the small- the generator bearing temperature. The wrapper approach uses
sample learning problem. supervised learning to perform 10-fold cross validation in
The ELM is a feedforward neural network which contains the selecting relevant parameters. The boosting-tree algorithm
input layer, the output layer and one single hidden layer. evaluates the importance of parameters by constructing a
Compared with other computationally expensive and time- sequence of decision trees and computing the prediction
consuming neural networks, the ELM adopts Penn Moore residuals. The relief algorithm selects the parameter set by
pseudo inverse to determine the weights and biases between detecting conditional dependence between the parameters. The
the hidden layer and output layer (Li et al., 2021b). This eight most important parameters selected by the three data-
method enables ELM to learn faster and attain higher mining algorithms are listed in Table 2.
generalization capability compared with other neural networks.
The KELM uses the kernel method over the vanilla ELM and it 3.3 Modeling Bearing Behavior
solves the problem of random initialization of ELM and has high Data from three wind turbines (i.e., Turbine C, Turbine D,
classification accuracy (Pandey et al., 2018; Ouyang 2021), good Turbine E) have been merged to train the neural network,
generalization ability and high degree of robustness. The support vector regression machine, the extreme-learning
Gaussian kernel function is the most frequently used kernel machine presented in Section 2.2, and the proposed deep
function and thus is selected in this study. belief network (DBN). Data collected from Turbine A, B, F
and G are used as validation dataset to validate prediction
performance of the proposed DBN algorithm. Data from
3 COMPUTATIONAL ANALYSIS Turbine G, J, I and K are used as testing dataset respectively.
To design the DBN, the number of hidden neurons in each layer
The data used in this research has been collected from SCADA is set at 10% of the training data (Mitchell, 1999). The data from
systems of a large wind farm. The data 10 min resolution data the remaining 2 healthy turbines (i.e., Turbine 9 and 11) are
from 11 wind turbines is used to investigate failure of a generator designated as test datasets to evaluate performance of the four
bearing. Two bearing failure instances have been reported during algorithms.
the period covered by the dataset. Table 3 presents prediction results produced by the four
algorithms based for the test and validation datasets. The
3.1 Dataset Description and Preprocessing mean absolute percentage error (MAPE) and the root mean
The ranges of the generator bearing temperature of the 11 square errors (RMSE) produced by the DBN algorithm are the
turbines are provided in Table 1. The bearing failure incidents smallest which confirms the accuracy of the DBN model. This
are also included in Table 1. Based on the maintenance records, superior performance may be attributed to the layer-wise pre-
Turbine B, H, I, and K have been affected by bearing failures and training.
are not considered for modeling normal bearing behavior Figure 4 illustrates prediction error from testing and
discussed in the Section 3.2. Rather they are selected to test validation produced by the deep belief network (DBN). The
abnormal behavior of the bearing temperature. APEs of healthy wind turbines and turbines with bearing
failures demonstrate different behaviors. Hence, the
3.2 Parameter Selection emerging bearing failure is indicated by the APE of the
To capture the normal behavior of a generator bearing, 33 DBN model.
parameters relevant to the bearing temperature have been
initially considered. Using domain expertise, the number of 3.4 Condition Monitoring
parameters of interest was reduced to 12. Next, three In this section, behavior of the prediction error associated with
algorithms (i.e., the wrapper with genetic search (WGS) the bearing failure is discussed. The APE was monitored for

Frontiers in Energy Research | www.frontiersin.org 5 November 2021 | Volume 9 | Article 799039

Li et al. Original Research About Energy

TABLE 3 | Performance evaluation of four algorithms.

Algorithm Validation Testing

Turbine B, H Turbine A, F Turbine G, J Turbine I, K
MAPE (%) RMSE MAPE (%) RMSE MAPE (%) RMSE MAPE (%) RMSE

SVR 1.27 0.93 0.96 0.56 4.95 4.27 4.46 3.98

NN 0.65 0.41 0.42 0.34 2.57 2.12 2.29 2.04
ELM 0.94 0.52 0.92 0.51 3.55 2.86 4.81 4.22
DBN 0.63 0.49 0.33 0.23 2.38 1.80 2.23 2.01

FIGURE 4 | The absolute percentage error produced by the deep belief network.

1 week prior to the bearing failure. The upper confidence 1 week prior to the bearing failure and an early alarm is
limit (UCL) and the lower confidence limit (LCL) of the issued. According to the results presented in Figure 6,
exponentially-weighted moving average (EWMA) control bearing failures are visible several days ahead of
chart are computed from Eqs 18, 19 of Section 2.4. the occurrence. The proposed approach provides
The monitored examples of healthy turbines and the sufficient time to react and thus minimize power loss and
turbines with emerging bearing failures are illustrated in downtime.
Figures 5, 6. The outcomes of the EWMAs are transformed into the real-
Figure 5 illustrates the EWMA charts of healthy time binary vectors and then the bearing failure classification
turbines (Turbine G and J) while Figure 6 shows the wind models are developed to classify the actual failures. However,
turbines (Turbine I and K) with problematic generator in the temporal domain, the optimal size of the EWMA
bearings of the same wind farm. In Figure 5, all statistics vectors are uncertain. Hence, this research performed several
fall within the control limits which indicates normal bearing experiments by trying difference size of the EWMA vectors
behavior. Meanwhile, outliers in Figure 6 begin to emerge (i.e., K  10, 20, 30, 40). All algorithms introduced in the Section

Frontiers in Energy Research | www.frontiersin.org 6 November 2021 | Volume 9 | Article 799039

Li et al. Original Research About Energy

FIGURE 5 | The EWMA control charts of two healthy turbines.

FIGURE 6 | The EWMA control charts of two turbines with bearing failures.

FIGURE 7 | The AUCs of all classification algorithms under different dimensions of EWMA vectors.

2.7 are tested and the computational results are illustrated in performance when K  20 and thus it is selected as the
Figure 7 below. The AUC is selected as the measurement It is optimal setting for the dimension of the input EWMA vector
obvious that all algorithms reached their peak classification in our study.

Frontiers in Energy Research | www.frontiersin.org 7 November 2021 | Volume 9 | Article 799039

Li et al. Original Research About Energy

FIGURE 8 | ROC curves of the bearing failure classification outcome.

TABLE 4 | Summary of bearing fault classification results.

Classifier Accuracy Sensitivity Specificity AUC

Mean 95% C.I. Mean 95% C.I. Mean 95% C.I. Mean 95% C.I.

SVM 0.74 (0.71–0.77) 0.64 (0.58–0.70) 0.87 (0.80–0.94) 0.76 (0.72–0.80)

LSSVM 0.83 (0.78–0.88) 0.77 (0.72–0.83) 0.94 (0.91–0.97) 0.88 (0.81–0.95)
ELM 0.71 (0.65–0.77) 0.64 (0.61–0.67) 0.79 (0.72–0.86) 0.69 (0.60–0.78)
KELM 0.76 (0.72–0.80) 0.71 (0.65–0.77) 0.82 (0.78–0.86) 0.75 (0.70–0.80)

As illustrated in Figure 8 below, the ROC curves for the four the statistical outliers. Instead, in this research, the machine-
state-of-art algorithms are obtained with respect to the testing learning classifiers enables the automation of this process. In sum,
dataset. Among them, the LS-SVM achieves the highest area it can be widely applied in wind farms for condition
under the ROC curve (AUC) as 0.88 which demonstrates its monitoring tasks.
superior performance in classifying bearing failures from the On the other hand, there are also few shortcomings at
binary vector mixed with normal and abnormal prediction current stage. For example, the sensor errors can be a
residuals. Meanwhile, the other performance metrics including misleading factor that cause false classification of mechanical
accuracy, sensitivity and specificity along with the 95% failures. The reliability of the SCADA sensors is not considered in
confidence intervals are also provided in Table 4. The LS- this framework. This can be a future direction of our current
SVM still performs best among all algorithms tested according research.
to all evaluation metrics. Hence, using the vectors generated from
the DBN and EWMA control charts, the LS-SVM is capable of
classify the majority of the bearing failures in the temporal 5 CONCLUSION
domain.
In this research, a deep-learning based condition-monitoring
framework to identify bearing failures was presented in this
4 DISCUSSION study. Historical data collected from healthy wind turbines
was utilized to develop a model predicting bearing
The condition-monitoring framework proposed in this study has temperature with a deep belief network. Data from both
provided promising results using field SCADA data. Overall, the healthy wind turbines and turbines to the bearing failures are
advantages of the proposed framework can be summarized into served as the testing dataset. Comparative analysis demonstrated
the following three points: First, it uses deep belief network as the that the deep belief network model was more accurate in
backbone regressor. It has shown superior power in extracting predicting generator bearing failures. An exponentially-
temporal abnormal features from the dataset. Second, the weighted moving-average control chart was applied to capture
framework is designed to be implemented on SCADA data shifts in prediction error. The control charts generated binary
which is the standard data collection system for almost all vectors lead to identification of the emerging bearing failure in
wind farms across the globe. Hence, it can be widely real-time in the temporal domain.
implemented on practice. Third, the classification part can Computational results reported in the paper validated
save a lor of labor and time. Conventional control chart-based accuracy of the deep-learning framework in condition
identification of mechanical failures requires humans to detect monitoring of wind turbine generator bearings. In the future

Frontiers in Energy Research | www.frontiersin.org 8 November 2021 | Volume 9 | Article 799039

Li et al. Original Research About Energy

research, analysis of high frequency vibration data may be and writing-original draft. All authors have read and agreed to the
coupled with the bearing temperature data for multi-scale published version of the manuscript.
condition monitoring.

FUNDING
DATA AVAILABILITY STATEMENT
This research is supported by National Key Research and
The raw data supporting the conclusion of this article will be Development Program of China (2018YFC1505105), the
made available by the authors, without undue reservation. Opening fund of State Key Laboratory of Geohazard
Prevention and Geoenvironment Protection (Chengdu
University of Technology) (Grant No. SKLGP2021K014), the
AUTHOR CONTRIBUTIONS “Miaozi project” of scientific and technological innovation
in Sichuan Province, China (Grant No. 2021090), the Project
HL conceptualized the study, contributed to the study from Sichuan Mineral Resources Research Center
methodology, and wrote the original draft. JD contributed to (SCKCZY2021-YB009), and the Open Research Subject of
the study methodology, data curation and investigation. SY and Key Laboratory of Fluid and Power Machinery
PF contributed to data analysis and investigation. HL contributed (Xihua University), Ministry of Education (Grant No.
to software and formal analysis. DA contributed to investigation LTDL2021-011).

Huang, N. E., and Wu, Z. (2008). A Review on Hilbert-Huang Transform: Method

REFERENCES and its Applications to Geophysical Studies. Rev. Geophys. 46 (2). doi:10.1029/
2007rg000228
Bach-Andersen, M., Rømer-Odgaard, B., and Winther, O. (2018). Deep Learning Huang, T. L., Ren, W. X., and Lou, M. L. (2008). “The Orthogonal Hilbert-Huang
for Automated Drivetrain Fault Detection. Wind Energy 21 (1), 29–41. Transform and its Application in Earthquake Motion Recordings Analysis,” in
doi:10.1002/we.2142 Proceedings of the 14th World Conference on Earthquake Engineering (Beijing,
Cherkassky, V., and Ma, Y. (2004). Practical Selection of SVM Parameters and China, October 2008, 12–17.
Noise Estimation for SVM Regression. Neural networks 17 (1), 113–126. Jones, L. A., Champ, C. W., and Rigdon, S. E. (2001). The Performance of
doi:10.1016/s0893-6080(03)00169-2 Exponentially Weighted Moving Average Charts with Estimated Parameters.
Deng, L., and Yu, D. (2013). Deep Learning for Signal and Information Processing. Technometrics 43 (2), 156–167. doi:10.1198/004017001750386279
IEEE Signal. Process. Mag. 28 (1), 145–154. doi:10.1109/MSP.2010.939038 Klein, R., Ingman, D., and Braun, S. (2001). Non-stationary Signals: Phase-Energy
Drucker, H., Burges, C., Kaufman, L., Smola, A., and Vapnik, V. (1997). Support Approach-Theory and Simulations. Mech. Syst. Signal Process. 15 (6),
Vector Regression Machines. Adv. Neural Inf. Process. Syst. 9, 155–161. 1061–1089. doi:10.1006/mssp.2001.1398
Feng, P., Wei, M., Dai, F., Tang, R., Qiu, H., and Gong, J. (2020). DEM Kohavi, R., and John, G. H. (1997). Wrappers for Feature Subset Selection. Artif.
Investigation on the Mechanical Behaviors of Flawed Specimens Subjected intelligence 97 (1-2), 273–324. doi:10.1016/s0004-3702(97)00043-x
to Coupled Static-Dynamic Loads. Soil Dyn. Earthquake Eng. 135, 106–220. Kusiak, A., and Verma, A. (2012). Analyzing Bearing Faults in Wind Turbines: a
doi:10.1016/j.soildyn.2020.106220 Data-Mining Approach. Renew. Energ. 48, 110–116. doi:10.1016/
Gritsenko, A., Sun, Z., Baek, S., Miche, Y., Hu, R., and Lendasse, A. (2017). j.renene.2012.04.020
“Deformable Surface Registration with Extreme Learning Machines,” in LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep Learning. nature 521 (7553),
International Conference on Extreme Learning Machine (Cham: Springer), 436–444. doi:10.1038/nature14539
304–316. doi:10.1007/978-3-030-01520-6_28 Lei, Y., Li, N., Lin, J., and Wang, S. (2013). Fault Diagnosis of Rotating Machinery
Guo, P. (2012). Wind Turbine Generator Bearing Condition Monitoring with Based on an Adaptive Ensemble Empirical Mode Decomposition. Sensors
NEST Method,” in Proceedings of the 2012 24th Chinese Control and Decision (Basel) 13 (1-2), 16950–16964. doi:10.3390/s131216950
Conference (CCDC) (IEEE), Taiyuan, China, May 2012, 235–239. 10.1109/ Li, H., Deng, J., Feng, P., Pu, C., Arachchige, D. D. K., and Cheng, Q. (2021a).
CCDC.2012.6244033. Short-Term Nacelle Orientation Forecasting Using Bilinear Transformation
He, Y., Fei, F., Wang, W., Song, X., Sun, Z., and Baek, S. (2018). “Predicting and ICEEMDAN Framework. Front. Energ. Res. 9, 780928. doi:10.3389/
Manufactured Shapes of a Projection Micro-stereolithography Process via fenrg.2021.780928
Convolutional Encoder-Decoder Networks,” in Proceedings of the Li, H., He, Y., Yang, H., Wei, Y., Li, S., and Xu, J. (2021b). Rainfall Prediction Using
International Design Engineering Technical Conferences and Computers Optimally Pruned Extreme Learning Machines. Nat. Hazards 108 (1), 799–817.
and Information in Engineering Conference (American Society of doi:10.1007/s11069-021-04706-9
Mechanical Engineers (ASME)), V01BT02A033. doi:10.1115/detc2018-85458 Li, H., Xu, Q., He, Y., and Deng, J. (2018). Prediction of Landslide Displacement
Hinton, G. E., Osindero, S., and Teh, Y.-W. (2006). A Fast Learning Algorithm for with an Ensemble-Based Extreme Learning Machine and Copula Models.
Deep Belief Nets. Neural Comput. 18 (7), 1527–1554. doi:10.1162/ Landslides 15 (10), 2047–2059. doi:10.1007/s10346-018-1020-2
neco.2006.18.7.1527 Li, H., Xu, Q., He, Y., Fan, X., and Li, S. (2020). Modeling and Predicting Reservoir
Hinton, G. E. (2002). Training Products of Experts by Minimizing Contrastive Landslide Displacement with Deep Belief Network and EWMA Control Charts:
Divergence. Neural Comput. 14 (8), 1771–1800. doi:10.1162/ a Case Study in Three Gorges Reservoir. Landslides 17 (3), 693–707.
089976602760128018 doi:10.1007/s10346-019-01312-6
Horng Shiau, J.-J., and Ya-Chen, H. (2005). Robustness of the EWMA Control Liang, N. Y., Saratchandran, P., Huang, G. B., and Sundararajan, N. (2006).
Chart to Non-normality for Autocorrelated Processes. Qual. Technology Classification of Mental Tasks from EEG Signals Using Extreme Learning
Quantitative Management 2 (2), 125–146. doi:10.1080/ Machine. Int. J. Neural Syst. 16 (1), 29–38. doi:10.1142/S0129065706000482
16843703.2005.11673089 Liu, W., Gao, P., Yu, W., Qu, Z., and Yang, C. (2018). Quantum Relief Algorithm.
Hu, Q., Zhang, R., and Zhou, Y. (2016). Transfer Learning for Short-Term Wind Quan. Inf. Process. 17 (10), 1–15. doi:10.1007/s11128-018-2048-x
Speed Prediction with Deep Neural Networks. Renew. Energ. 85, 83–95. Mitchell, T. M. (1999). Machine Learning and Data Mining. Commun. ACM 42
doi:10.1016/j.renene.2015.06.034 (11), 30–36. doi:10.1145/319382.319388

Frontiers in Energy Research | www.frontiersin.org 9 November 2021 | Volume 9 | Article 799039

Li et al. Original Research About Energy

Ouyang, T. (2021). Feature Learning for Stacked ELM via Low-Rank Matrix Teng, W., Ding, X., Zhang, X., Liu, Y., and Ma, Z. (2016). Multi-fault Detection and
Factorization. Neurocomputing 448, 82–93. doi:10.1016/ Failure Analysis of Wind Turbine Gearbox Using Complex Wavelet Transform.
j.neucom.2021.03.110 Renew. Energ. 93, 591–598. doi:10.1016/j.renene.2016.03.025
Ouyang, T., He, Y., Li, H., Sun, Z., and Baek, S. (2019). Modeling and Forecasting Wang, H.-z., Li, G.-q., Wang, G.-b., Peng, J.-c., Jiang, H., and Liu, Y.-t. (2017). Deep
Short-Term Power Load with Copula Model and Deep Belief Network. IEEE Learning Based Ensemble Approach for Probabilistic Wind Power Forecasting.
Trans. Emerg. Top. Comput. Intell. 3 (2), 127–136. doi:10.1109/ Appl. Energ. 188, 56–70. doi:10.1016/j.apenergy.2016.11.111
tetci.2018.2880511 Wang, H. Z., Wang, G. B., Li, G. Q., Peng, J. C., and Liu, Y. T. (2016c). Deep Belief
Ouyang, T., Kusiak, A., and He, Y. (2017). Modeling Wind-Turbine Power Curve: Network Based Deterministic and Probabilistic Wind Speed Forecasting
A Data Partitioning and Mining Approach. Renew. Energ. 102, 1–8. Approach. Appl. Energ. 182, 80–93. doi:10.1016/j.apenergy.2016.08.108
doi:10.1016/j.renene.2016.10.032 Wang, L., Zhang, Z., Long, H., Xu, J., and Liu, R. (2016a). Wind Turbine Gearbox
Pandey, P., Patel, V., George, N. V., and Mallajosyula, S. S. (2018). KELM- Failure Identification with Deep Neural Networks. IEEE Trans. Ind. Inform. 13
CPPpred: Kernel Extreme Learning Machine Based Prediction Model for (3), 1360–1368.
Cell-Penetrating Peptides. J. Proteome Res. 17 (9), 3214–3222. doi:10.1021/ Wang, L., Zhang, Z., Xu, J., and Liu, R. (2016b). Wind Turbine Blade Breakage
acs.jproteome.8b00322 Monitoring with Deep Autoencoders. IEEE Trans. Smart Grid 9 (4), 2824–2833.
Peeters, C., Guillaume, P., and Helsen, J. (2018). Vibration-based Bearing Fault Yan, R., Gao, R. X., and Chen, X. (2014). Wavelets for Fault Diagnosis of Rotary
Detection for Operations and Maintenance Cost Reduction in Wind Energy. Machines: a Review with Applications. Signal. Process. 96, 1–15. doi:10.1016/
Renew. Energ. 116, 74–87. doi:10.1016/j.renene.2017.01.056 j.sigpro.2013.04.015
Peng, Z., Peter, W., and Chu, F. (2005). An Improved Hilbert–Huang Transform Yang, B., Liu, R., and Chen, X. (2017). Fault Diagnosis for a Wind Turbine
and its Application in Vibration Signal Analysis. J. sound vibration 286 (1-2), Generator Bearing via Sparse Representation and Shift-Invariant K-Svd. IEEE
187–205. doi:10.1016/j.jsv.2004.10.005 Trans. Ind. Inform. 23 (5), 91–99. doi:10.1109/tii.2017.2662215
Qiu, X., Ren, Y., Suganthan, P. N., and Amaratunga, G. A. J. (2017). Empirical Yang, W., Court, R., and Jiang, J. (2013). Wind Turbine Condition Monitoring by
Mode Decomposition Based Ensemble Deep Learning for Load Demand Time the Approach of SCADA Data Analysis. Renew. Energ. 53, 365–376.
Series Forecasting. Appl. Soft Comput. 54, 246–255. doi:10.1016/ doi:10.1016/j.renene.2012.11.030
j.asoc.2017.01.015 Yang, W., Liu, C., and Jiang, D. (2018). An Unsupervised Spatiotemporal Graphical
Sbihi, A. (2007). A Best First Search Exact Algorithm for the Multiple-Choice Modeling Approach for Wind Turbine Condition Monitoring. Renew. Energ.
Multidimensional Knapsack Problem. J. Comb. Optim 13 (4), 337–351. 127, 230–241. doi:10.1016/j.renene.2018.04.059
doi:10.1007/s10878-006-9035-3 Zhu, X., Xu, Q., Tang, M., Li, H., and Liu, F. (2018). A Hybrid Machine Learning and
Shen, X., Ouyang, T., Khajorntraidet, C., Li, Y., Li, S., and Zhuang, J. (2021a). Computing Model for Forecasting Displacement of Multifactor-Induced Landslides.
Mixture Density Networks-Based Knock Simulator. IEEE/ASME Trans. Neural Comput. Applic 30 (12), 3825–3835. doi:10.1007/s00521-017-2968-x
Mechatronics, 1. 10.1109/TMECH.2021.3059775..
Shen, X., Ouyang, T., Yang, N., and Zhuang, J. (2021b). Sample-Based Neural Conflict of Interest: The authors declare that the research was conducted in the
Approximation Approach for Probabilistic Constrained Programs. IEEE Trans. absence of any commercial or financial relationships that could be construed as a
Neural Networks Learn. Syst.. doi:10.1109/tnnls.2021.3102323 potential conflict of interest.
Shen, X., and Raksincharoensak, P. (2021). Pedestrian-Aware Statistical Risk
Assessment. IEEE Trans. Intell. Transportation Syst.. doi:10.1109/ Publisher’s Note: All claims expressed in this article are solely those of the authors
tits.2021.3074522 and do not necessarily represent those of their affiliated organizations, or those of
Sun, Z., He, Y., Gritsenko, A., Lendasse, A., and Baek, S. (2020a). Embedded the publisher, the editors and the reviewers. Any product that may be evaluated in
Spectral Descriptors: Learning the point-wise Correspondence Metric via this article, or claim that may be made by its manufacturer, is not guaranteed or
Siamese Neural Networks. J. Comput. Des. Eng. 7 (1), 18–29. doi:10.1093/ endorsed by the publisher.
jcde/qwaa003
Sun, Z., Rooke, E., Charton, J., He, Y., Lu, J., and Baek, S. (2020b). Zernet: Copyright © 2021 Li, Deng, Yuan, Feng and Arachchige. This is an open-access
Convolutional Neural Networks on Arbitrary Surfaces via Zernike Local article distributed under the terms of the Creative Commons Attribution License (CC
tangent Space Estimation. Computer Graphics Forum 39 (6), 204–216. BY). The use, distribution or reproduction in other forums is permitted, provided the
doi:10.1111/cgf.14012 original author(s) and the copyright owner(s) are credited and that the original
Tavner, P. J., Greenwood, D. M., Whittle, M. W. G., Gindele, R., Faulstich, S., and publication in this journal is cited, in accordance with accepted academic practice.
Hahn, B. (2012). Study of Weather and Location Effects on Wind Turbine No use, distribution or reproduction is permitted which does not comply with
Failure Rates. Wind Energy 16 (2), 175–187. doi:10.1002/we.538 these terms.

Frontiers in Energy Research | www.frontiersin.org 10 November 2021 | Volume 9 | Article 799039