Wind Turbine Anomaly Identification Based on Improved Deep Belief Network with SCADA Data
This article presents a novel fault diagnosis algorithm based on the whale optimization algorithm (WOA)-deep belief networks (DBN) for wind turbines (WTs) using the data collected from the supervisory control and data acquisition (SCADA) system. Through the domain knowledge and Pearson correlation, the input parameters of the prediction models are selected. Three different types of prediction models, namely, the wind turbine, the wind power gearbox, and the wind power generator, are used to predict the health condition of the WT equipment. In this article, the prediction accuracy of the models built with these SCADA sample data is discussed. In order to implement fault monitoring and abnormal state determination of the wind power equipment, the exponential weighted moving average (EWMA) threshold is used to monitor the trend of reconstruction errors. The proposed method is used for 2 MW wind turbines with doubly fed induction generators in a real-world wind farm, and experimental results show that the proposed method is effective in the fault diagnosis of wind turbines.
Wind power is considered one of the most promising forms of renewable energy sources, which has been widely used in the world [1–3]. However, it results in a higher failure rate than expected because the wind turbine (WT) is exposed to adverse environmental conditions. The operation and maintenance (O&M) costs account for approximately 15–20% of the gross income of onshore wind farms and 30–35% of offshore wind farms . Therefore, condition monitoring (CM) and fault diagnosis technology can achieve preventive maintenance before the malfunction happens. Many kinds of condition monitoring and fault diagnosis techniques, such as acoustic emission [5–8], vibration analysis [9–12], strain measurement , electrical effects analysis , and ultrasonic testing techniques , have been developed to minimize downtime as well as the O&M costs. The recent review in  included a comprehensive survey of the existing fault diagnosis technologies for wind turbines with qualitative and quantitative methods, and the fault diagnosis system is also described in this paper. However, most of these methods have not been universally used in wind farms because they are mostly based on the analysis of mechanical signals emitted during the operation process and require a large number of extra sensors, thus resulting in high-deployment costs.
As an important part of wind power equipment, the supervisory control and data acquisition (SCADA) can remotely collect equipment operating data . It has the advantages of providing complete information, improving efficiency, and correctly mastering the operating status of the system. Moreover, it can also help to monitor the health condition of wind turbines. However, currently, there is no effective state evaluation algorithm in most SCADA systems. Only when the sampled data exceed the given threshold can the alarm signal of the wind turbine components be given. At this point, the fault may have deteriorated to a certain level. In order to further improve the fault diagnosis rate and reduce the high-deployment costs, it is of significant academic value and application prospect to establish an evaluation algorithm and model using SCADA data.
In recent years, the measurement data that come from the SCADA system of a wind farm is used for condition monitoring and fault diagnosis [18–21]. On the one hand, it has achieved remarkable results in reducing the wind turbines downtimes and improving the reliability of wind power. On the other hand, it has become profitable to integrate multisource data signals including mechanical and electrical signals in the wind turbines for analysis. A large number of data mining and abnormal identification methods have been applied to improve diagnostic accuracy upon integrating different condition parameters of the SCADA system. In [22–24], the prediction models composed of conditional parameters of the wind turbine are established by the neural networks. In , the historical SCADA data, its current SCADA data, and some wind turbines' current SCADA data from other wind farms are used to train and test these models. The study in  proposes an online wind turbine fault detection system through the multilayer neural network technology using SCADA data. The study cases show the effectiveness of this method. However, in these researches, the input parameters of the model are selected based on domain knowledge, without considering the correlation between the parameters which are associated with other signals that are measured simultaneously. In , a method for cointegration analysis of wind turbine fault diagnosis is presented based on the SCADA data. In this current research, six condition parameters are selected for cointegration verification, and abnormal problems can be detected through this method. It is worth pointing out that the proposed approach in  assumes that the cointegration relationship is unique; however, in actual situation, the cointegration relationship may be diverse. In , an AutoRegressive with eXogenous input model for wind turbine generator bearing fault is established using SCADA data. However, the choice of manual intervention parameters in this method noticeably affects the accuracy of the model. A deep learning method of a layerwise encoder network for the CM of wind turbine main bearing through SCADA data is proposed in . However, the selection of the number of neurons in the hidden layer and the learning rate parameters have a significant impact on the prediction accuracy.
For different target condition parameters, the different sample data selected from the WT SCADA system will have a significant impact on diagnostic accuracy. However, to the best of our knowledge, little research has been carried out on the effects of the selected condition parameters for WT fault diagnosis model. Aiming to these problems, in this article, a novel fault diagnosis algorithm for WTs based on the whale optimization algorithm (WOA)-deep belief networks (DBN) using SCADA data is proposed. The proposed novel fault diagnosis algorithm can automatically analyze a large number of SCADA data with a low sampling rate. Furthermore, the input condition parameters of this fault diagnosis model are selected based on domain knowledge and the correlation between conditional parameters. This method effectively solves the problem of difficult fault diagnosis of the wind turbine, reduces the installation cost of fault diagnosis equipment, and improves the economic benefit of the wind farm.
The structure of the article is presented in the following manner. The wind turbine and the SCADA system are introduced in Section 2. Section 3 describes the condition parameter selection, the model performance analysis, and the exponentially weighted moving average (EWMA) thresholds method. The whale optimization algorithm, deep belief networks technology, and the WOA-DBN approach are presented in detail in Section 4. In Section 5, a framework of the WT condition monitoring using SCADA data is discussed. In Section 6, the field example is investigated to prove the effectiveness of the proposed method. The conclusions are summarized in Section 7.
2. Wind Turbine and SCADA System
2.1. Description of Wind Turbine
As illustrated in Figure 1, the wind turbine studied in this article is mainly composed of the rotor system, main shaft, gearbox, generator, converter, and so on. And the capacity of the system is 2 MW. The wind energy captured by the rotor system blades is converted into mechanical energy, and then, the main bearings transmit mechanical energy to the generator through the gearbox, while increasing lower rotor speed of the wind turbine to a higher speed. Finally, the generator converts mechanical energy into electrical energy. The simulation and the wind turbine operational principle are discussed [27, 28].
Wind turbines work in complex and varied environments consistently, resulting in frequent failures and high O&M costs, and it should be noted that the longest mean time to repair corresponds to the blade failure, followed by the gearbox and the generator .
2.2. Wind Turbine SCADA System
The SCADA system is a data acquisition and monitoring control system, which plays a vital role in the collection, management, and transmission of wind turbine real-time data. The data acquisition of each WT is then transmitted to the data information to the monitoring control system through equipment such as optical fiber and switches. As shown in Figure 2, the monitoring control system provides the interface, picture, and service operation for the wind farm operators to understand all kinds of information of the wind field. The monitoring software provides users with online monitoring, alarm reception, remote control, and operational data charts, using a friendly and flexible user graphical interface, which can real-time monitor the relevant data of wind turbine operation.
However, relying solely on the SCADA alarm system does not accurately grasp the operation status of the wind turbine. Due to the wide setting of the alarm threshold, when the system issues an alarm, the fault may have deteriorated to a serious degree, which cannot prevent the failure. In order to improve the fault diagnosis rate and prevent faults in advance, the SCADA data used in this study were derived from the 2 MW onshore wind turbines of a wind farm in Central China. The continuous monitoring data from March 2018 to September 2018 were collected.
3. Prediction Model Development
3.1. Parameter Description and Processing
Up to 97 different types of condition parameters are measured and provided in the SCADA system of a wind farm, including wind speed, wind power, and temperature of different components. However, not every parameter plays a decisive role. For various fault classifications, the choice of parameters will be different, which will lead to different diagnostic models. In order to improve accuracy rates and reduce the diagnostic time, the input condition parameters for each model are selected based on domain knowledge and the correlation between parameters. For example, the relationship between wind speed and the wind power, and the relationship between the oil sump temperature of the wind gearbox and wind power are shown in Figures 3(a) and 3(b), respectively.
The selection of input condition parameters has a high impact on simplifying the model and ensuring prediction accuracy. Firstly, according to the judgment of expert domain knowledge, a total of 49 condition parameters are selected. Secondly, after judging the correlation coefficient as shown in equation (1), where denotes the Pearson correlation, is the length of the condition parameter, and are the sample point corresponding to different parameters, and denote the average of the condition parameters, and different parameters are selected as shown in equation (1). Finally, the selected conditional parameters are modeled to assess the health of the wind turbine. In this step, according to the expert domain knowledge and the size of correlation between the condition parameters, the type and number of parameters selected are different for different prediction models, which is beneficial to the increase in prediction accuracy and the decrease in prediction time.
3.2. Performance Analysis
As shown in equations (2)–(4), the root mean square error (RMSE), the mean absolute percentage error (MAPE), and the mean absolute error (MAE) are used to evaluate the performance of the predictive model.where and represent the actual and predicted values of the sample and is the number of the test samples.
3.3. The Reconstruction Errors and Exponentially Weighted Moving Average Thresholds
The reconstruction error between the actual value and the predicted value of the dataset is used as the detection quantity of the wind turbine state trend. The EWMA  threshold is used to monitor the trend of the reconstruction error to realize the fault monitoring and abnormal state determination of the wind power equipment because the trend of reconstruction error will change significantly under normal operating conditions and fault conditions. The reconstruction error is given by
Based on the above analysis, the threshold control chart is sensitive to detect a small drift of the mean value in the process, and the exponential weighted moving average is used to set the control line to monitor the fluctuation of reconstruction error. The statistics () of the EWMA control chart arewhere represents the weight of history on the current EWMA statistic, , and denotes the time. The mean and variance of the statistics are obtained by equation (7), where and denote the mean and variance of the statistics, respectively, is the mean of the of the wind power equipment, and represents the sampling length.
Therefore, the upper and lower bound functions of EWMA control charts for wind power equipment based on time are expressed aswhere is a value of 2 and 3, in which 2 is taken as the fault warning limit and 3 is taken as the fault detection warning limit, is a value of 3, and the equipment is in the safe zone when . The principle is shown in Figure 4.
4. Whale Optimization Algorithm and Deep Belief Network Approach
4.1. Deep Belief Network Approach
Deep belief network, which was first proposed by Hinton , is a kind of generative depth model, and it has become a widely used deep structure in the field of research and applied science [31–33]. As shown in Figure 5, the structure of DBN consists of a number of restricted Boltzmann machines (RBM) stacked layer by layer [31, 34], while extracting important information and critical features through deep data mining. Based on the RBM learning rule, the data conversion process from the visible layer to the hidden layer is completed by a sigmoid activation function . The hidden layer of the first RBM is considered a visible layer of the second RBM and so on. The training process continues until the final RBM is completed. Trained RBMs are stacked to form DBN; for the DBN classification model with hidden layers, the joint probability () of the whole DBN is given aswhere is the conditional distribution between the hidden layer and the hidden layer.
Figure 5 shows the RBM as the development of Boltzmann machine (BM) proposed by Ackley in , which was further introduced in . From Figure 5, it can be seen that RBM consists of a visual layer containing visible units (, , …) and a hidden layer containing hidden units (, , …, ). Typically, the visible units follow the Bernoulli or Gaussian distribution, whereas the hidden units usually follow the Bernoulli distribution. The state of the RBM is generally evaluated using an energy function:where denotes network parameters, is the weight between hidden unit and visible unit , and and are the bias of the hidden variable and the visual variable, respectively. The negative and positive phases of the RBM learning can be mathematically expressed aswhere the function is sigmoid . Based on the contrast divergence algorithm, the update parameters of the RBM can be obtained aswhere is the learning rate and represents the number of the RBMs.
4.2. The Whale Optimization Algorithm
The WOA, which mimics the social behavior of humpback whales, is known as a nature-inspired metaheuristic optimization algorithm, proposed by Mirjalili and Lewis . As shown in Figure 6, this social feeding behavior of humpback whales is called the bubble-net predation strategy. This optimization algorithm is becoming more and more popular in engineering applications because it does not require gradient information, relies on relatively simple concepts, is easy to implement, and can bypass local optimization to find global optimality . As can be seen from Figure 6, whales attack their prey by spiraling upward and contracting their encirclement.
A mathematical model of the WOA algorithm can be obtained by combining the three aspects of encircling prey, bubble-net attacking predation strategy, and hunting prey. The encircling prey behavior can be represented bywhere is the current iteration, and denote coefficient vectors, represents the position vector, and indicates the position vector of the best solution currently obtained, and it should be updated in each iteration if there is a better solution. The and are obtained aswhere denotes a random vector in [0, 1] and linearly drops from 2 to 0 throughout the process.
According to the behavior of the humpback whale, the mathematical model of the bubble-net attacking predation strategy is aswhere is a constant that defines the logarithmic spiral shape and is a random number between −1 and 1. It is worth noting that the whale swims in the spiral shape to the prey while shrinking the encirclement. To model this synchronous behavior, it is assumed that there is a probability of to choose the contraction enclosing mechanism and a probability selection spiral model of to update the position of the whale during optimization. The mathematical model can be expressed as
On hunting prey stage, the mathematical model is given aswhere denotes a random position vector (a random whale) selected from the current population.
The WOA algorithm begins with a set of random solutions. In each iteration, the search agent updates its location based on a randomly selected search agent or the best solution obtained so far. The random search agent is selected when , and the best solution is chosen when for updating the search agent location. According to the value, the WOA algorithm can switch between either spiral motion or circular motion. Finally, the WOA algorithm is terminated according to the satisfaction of termination criteria.
4.3. Deep Belief Networks Optimized by Whale Optimization Algorithm
The random initialization of DBN’s network parameters is easy to cause local optimization. For this reason, this article uses the WOA to globally optimize the DBN network parameters and provide the DBN reliable initialization parameters, and the global optimal solution is terminated based on the maximum number of iterations. In order to reduce the fluctuation of prediction accuracy caused by the artificial selection of the DBN network parameters, the WOA is used to optimize the learning rate of the DBN network () and the number of the hidden neurons in each layer of the RBM () to obtain the highest accuracy in the global sample space.
In this article, the WOA algorithm is used to optimize the parameter combination of the DBN model, and the fitness function is selected as equation (22), where and represent the actual and predicted values of the sample and denotes the number of hidden neurons in layer of the RBM, respectively.
The flowchart of the WOA-DBN algorithm is shown in Figure 7, and it can be summarized and described as follows:(i)Step 1. Initialize the WOA algorithm parameters. The maximum number of iterations, the population size of whales, and the variable dimensions and their ranges are set.(ii)Step 2. Randomly initialize the generated whale's position.(iii)Step 3. Calculate the fitness of the individual group, according to the objective function (equation (19)), and find the position of the current global optimal solution as the target position.(iv)Step 4. Update the individual whale location.(v)Step 5. Determine whether the algorithm satisfies termination conditions. If the maximum number of iterations is reached, the optimal parameters of DBN are output to establish the prediction model; otherwise, return to Step 3.(vi)Step 6. Obtain the best parameters. When the maximum number of iterations is reached, the global optimal solution is obtained.(vii)Step 7. Generation of the DBN prediction model. According to the model training, the fault prediction results are obtained.(viii)Step 8. The validity of the model is verified by examples.
5. The Framework of the WT Condition Monitoring Using SCADA Data
5.1. Implementation of the Proposed Method
In this work, the WOA-DBN approach is employed for WT fault diagnosis using SCADA data. Figure 8 shows the application of the present proposed method, and it can be summarized as follows:(i)Step 1. Collect the SCADA data. As the research object, the same type of wind turbine was selected, and the original operation data of the WTs were extracted from the wind farm SCADA system.(ii)Step 2. Process the SCADA data. Firstly, the condition parameters are selected from the original operational data based on expert domain knowledge. Then, the SCADA data are further processed according to the wind farm accident operation record. Finally, the accident operation data and the error data in the SCADA system are deleted.(iii)Step 3. Select the input and output parameters for each model. (1) Select the output parameters of each model through expert domain knowledge; (2) calculate the correlation between the parameters by the Pearson correlation coefficient (equation (1)); and (3) when the correlation coefficient value between the output parameter and the remaining parameters is greater than a certain value, these parameters are used as input parameters.(iv)Step 4. Normalize the selected SCADA data to get samples. It can obtain the preprocessed data samples to test or train the prediction model after normalizing the SCADA data.(v)Step 5. Randomly divide the samples into testing and training sample. Normalize the selected SCADA data to get samples. The ratio of the testing dataset to the training dataset is randomly set. The testing samples are applied to validate the prediction model, and the training samples are adopted to train this model.(vi)Step 6. Optimize the DBN parameters by the WOA. In order to improve the situation that the DBN method is easy to fall into local optimum, the WOA technology is used to optimize the learning rate and the number of hidden neurons in layer of the RBM.(vii)Step 7. Train the DBN prediction model and analyze its performance. The three prediction models of the whole wind turbine, WT gearbox, and WT generator are trained separately, and the prediction model are evaluated using the RMSE, MAPE, and MAE indicators.(viii)Step 8. Apply the DBN prediction model to the actual wind farm and evaluate the results. The SCADA data of the actual wind field are input into the prediction model to evaluate the condition of the wind turbine, and the economic benefits of the wind farm are analyzed under the method.
6. The Example Verification
The WT example verification units, which are located in the hills at an altitude of about 1,700 meters originate from a wind farm in Central China. As shown in Figure 9, a total of 25 wind turbines (2-MW) are sent to 110 kV step-up substation through three collector lines in this wind farm.
The SCADA analysis data, collected from March 2018 to September 2018, derived from 4 WTs (#3, #11, #12, and #19) with the same type in this article.
6.1. Data Acquisition and Preprocessing
The wind speed of the SCADA data used for modeling is between 3.5 m/s and 25 m/s by statistically sorting the operational data and accident records of the wind farm, and the sampling interval of these data is 10 minutes. In this way, a data matrix with 8209 lines and 97 columns is obtained. Based on domain knowledge, the data can be further processed to obtain a data matrix, which contains 8209 rows and 49 columns.
In order to improve the accuracy of the prediction, the data matrix is further processed by the Pearson correlation coefficient, and the health prediction model of the whole WT is taken as an example to illustrate. Firstly, the wind power is taken as the output, the correlation coefficients of 49 selected conditional parameters are calculated, and those with correlation larger than 0.896 (shown in Figure 10) are selected as input parameters. Secondly, according to the previous step, the condition parameters of three different types of health prediction models can be obtained, as shown in Table 1, where GNT. denotes the generator and T. denotes the temperature. The third row represents the output condition parameters of the three prediction models, and the remaining rows represent the input condition parameters of the prediction models. Figure 11 shows the boxplot of the 12 condition parameters in the whole WT prediction model, where ST.W. represents the stator winding. Finally, the corresponding two columns of parameter values in Table 1 are modeled separately for the WT generator and the WT gearbox, and the performance analysis values of various models are obtained.
6.2. Experimental Results
The WOA-DBN-based method is used to recognize the fault classification of WTs. The programs are run in the MATLAB R2017a environment, where 80% of the samples are randomly exploited for training, and the remaining are used for the test of the fault diagnosis performance. The WOA parameters are set as follows: the maximum number of iterations is set as 10, the population size of whales is set to as 40, the variable dimensions are set as 2, and their ranges are set to as [100 100 1] and [1 1 0.01], respectively. The DBN parameters are set to as follows: the number of train iterations is determined to be 10, the set batch size is selected to be 2, and a network with an output unit is set up. On the basis of these parameters, the performance of the trained WOA-DBN prediction model is shown in Table 2. The predicted results produced by the proposed method are shown in Figures 12(a), 12(b), and 12(c).
After the model training is completed, according to the historical data of the wind farm, the SCADA fault data of the three sets of the wind turbine, gearbox, and generator are searched separately, and the validity of the model for fault detection is verified according to the content proposed in Section 3.3. According to the on-site accident record book, some data of wind turbine, wind power gearbox, and the whole machine before failure are collected, and then, the EWMA method is used to verify whether the WOA-DBN model can judge the health status of wind power equipment in advance. As shown in Figures 12(d), 12(e), and 12(f), the simulation results show that the proposed method can accurately judge the fault types of wind power equipment.
From Figure 12(d), it can be concluded that the fault data samples of the wind turbine generator are predicted by the model, and only the EWMA value of one data sample does not exceed the upper limit of the alarm value. Considering the influence of singularity point, it can be judged that the wind turbine generator has been faulted. According to the fault records of the wind farm, three days after the operation of the unit, the “generator temperature alarm” message issued by the SCADA system indicates that the generator has failed. Similarly, from Figure 12(e), it can be seen that all EWMA of the fault data samples corresponding to the wind power gearbox exceeds its alarm upper limit value, and it can be judged that the gearbox has been faulted and needs to be checked and further processed. The fault alarm results of the whole machine show that 15% of the EWMA value of the fault data sample does not exceed its upper limit of the alarm value as seen in Figure 12(f). Although it is judged by the WOA-DBN method that the probability of failure of the unit is large, in order to avoid the occurrence of fault misjudgment, it should be further judged according to the experience of the field staff whether the unit should be shut down for inspection.
6.3. Economic Analysis
Checking the on-site operation record found that the #12 wind turbine was out of operation for 42 days due to gearbox damage failure at 14 : 05 on August 5, 2018, and the gearbox needed to be replaced. Five days before the failure of the gearbox, the field data are collected, and the method proposed in this article is used for condition monitoring and analysis. It is found that this method can accurately judge the alarm value of the wind power gearbox exceeding the upper limit. If the machine was stopped and checked in time, the whole damage accident of the gearbox could be avoided and only the high-speed axle of the gearbox would need to be replaced. According to the analysis of the wind field, it usually takes about 15 days to replace the high-speed bearing of the gearbox. Taking Nangao Gear as an example, the direct cost of hoisting and purchasing the gearbox is about 116,427 $, while the maximum cost of replacing a high-speed shaft of the gearbox is not more than 36,383 $. The detailed statistics are shown in Table 3, where it can be concluded that the fault prediction method proposed in this article could have reduced the direct economic loss about 115,949 $.
In addition, according to the on-site operation records of #47 wind turbine, at 18 : 20 on January 7, 2019, the #47 wind turbine was shut down for 54 days due to generator failure, and the generator needs to be replaced. 10 days before the generator failure, the field data are collected, and the method proposed in this article to carry out condition monitoring and fault diagnosis analysis. The results show that 80% of the alarm value of the wind turbine is over the upper limit. According to the auxiliary analysis of the on-site operators, if the wind turbine is shut down for inspection in time, the occurrence of the cleaning accident of the generator can be avoided, and only the rear bearing of the generator needs to be replaced. Using the data of wind field for analysis, it usually takes only 3 days to replace the generator rear bearing. Take the wind turbine as an example, the replacement cost of the generator is about 90,154 $, while the maximum cost of replacing the rear bearing is not more than 1600 $. The specific statistical results are shown in Table 4. It can be seen that the fault prediction method proposed in this article can reduce the direct economic loss about 175,498 $.
All in all, the method can effectively monitor the condition of the wind turbine and reduce the operation and maintenance cost of the wind farm.
A novel method for the fault diagnosis of a wind turbine based on the WOA-DBN approach using the SCADA data is presented in this article. The SCADA data are preprocessed based on domain knowledge and the Pearson correlation coefficient. Three prediction models, namely, the whole wind turbine, the wind power gearbox, and the wind power generator, are established in this article. The simulation results show that the proposed method has high prediction accuracy. Field examples further show that this method can realize the condition monitoring of wind turbines and reduce the operation and maintenance costs of wind farms. This method can effectively detect the fault of a wind turbine based on SCADA data and can provide a new idea for the health management of wind turbine in the data center of wind farm. It is a supplement to the deficiency of current fault detection methods. However, in this study, the SCADA data can only be used to judge the failure of the wind turbine equipment. Specifically, some types of failure need to be further verified by relevant technical means, which is a problem to be solved in the follow-up study.
The data used to support the findings of this study are available from the corresponding author upon request.
Conflicts of Interest
The authors declare that there are no conflicts of interest regarding the publication of this article.
This research was supported by the National Basic Research National Natural Science Foundation of China (Grant no. 51977072) and the Natural Science Foundation of China (Grant no. 2018YFB0606005).
R. Ruiz-Cruz, E. N. Sanchez, A. G. Loukianov, and J. A. Ruz-Hernandez, “Real-time neural inverse optimal control for a wind generator,” Ieee Transactions On Sustainable Energy, vol. 10, no. 3, pp. 1172–1183, 2019.View at: Publisher Site | Google Scholar
H. Zhou, S. Yi, and Y. Liu, “A fatigue life prediction method for the drive system of wind turbine using internet of things,” Materials Science and Engineering, vol. 2020, Article ID 9048508, 8 pages, 2020.View at: Publisher Site | Google Scholar
B. Wang, W. Li, and X. Chen, “Improved chicken swarm algorithms based on chaos theory and its application in wind power interval prediction,” Mathematical Problems in Engineering, vol. 2019, Article ID 1240717, 2019.View at: Publisher Site | Google Scholar
L. Hui, Y. Hu, and L. Yang, “Overview of condition monitoring and fault diagnosis for grid-connected high-power wind turbine unit,” Electric Power Automation Equipment, vol. 36, no. 1, pp. 6–16, 2016.View at: Google Scholar
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, vol. 51, pp. 175–181, 2013.View at: Publisher Site | Google Scholar
T. Bouno, T. Yuji, and T. Hamada, “Failure forecast diagnosis of small wind turbine using acoustic emission sensor,” KIEE International Transactions on Electrical Machinery and Energy Conversion Systems, vol. 1, pp. 78–83, 2005.View at: Google Scholar
Y. Zhang, W. Lu, and F. Chu, “Planet gear fault localization for wind turbine gearbox using acoustic emission signals,” Renewable Energy, vol. 109, pp. 449–460, 2017.View at: Publisher Site | Google Scholar
L. Raúl Ruiz de and H. Gonzalez-Carrato, “Sound and vibration-based pattern recognition for wind turbines driving mechanisms,” Renewable Energy, vol. 109, pp. 262–274, 2013.View at: Google Scholar
P. Shi, W. X. Yang, and M. P. Sheng, “An enhanced empirical wavelet transform for features extraction from wind turbine condition monitoring signals,” Energies, vol. 10, pp. 1210–1226, 2017.View at: Publisher Site | Google Scholar
R. Yan, Y. V. Qian, and S. J. Hu, “Wind turbine gearbox fault diagnosis based on wavelet domain stationary subspace analysis,” Journal of Mechanical Engineering, vol. 50, no. 11, pp. 9–16, 2014.View at: Publisher Site | Google Scholar
N. Sawalhi, R. B. Randall, and D. Forrester, “Separation and enhancement of gear and bearing signals for the diagnosis of wind turbine transmission systems,” Wind Energy, vol. 17, no. 5, pp. 729–743, 2014.View at: Publisher Site | Google Scholar
J. Igba, K. Alemzadeh, C. Durugbo, and E. T. Eiriksson, “Analysing RMS and peak values of vibration signals for condition monitoring of wind turbine gearboxes,” Renewable Energy, vol. 91, pp. 90–106, 2016.View at: Publisher Site | Google Scholar
K. Schroeder, W. Ecke, J. Apitz, E. Lembke, and G. Lenschow, “A fibre Bragg grating sensor system monitors operational load in a wind turbine rotor blade,” Measurement Science and Technology, vol. 17, no. 5, pp. 1167–1172, 2006.View at: Publisher Site | Google Scholar
R. R. Schoen, B. K. Lin, T. G. Habetler, J. H. Schlag, and S. Farag, “An unsupervised, on-line system for induction motor fault detection using stator current monitoring,” IEEE Transactions on Industry Applications, vol. 31, no. 6, pp. 1280–1286, 1995.View at: Publisher Site | Google Scholar
E. Jasiuniene, R. Raisutis, and R. Sliteris, “NDT of wind turbine blades using adapted ultrasonic and radiographic techniques,” Insight-Non-Destructive Testing and Condition Monitoring, vol. 51, no. 9, pp. 477–483, 2009.View at: Google Scholar
X. Long, P. Yang, and H. Guo, “Review of fault diagnosis methods for large wind turbines,” Power System Technology, vol. 41, pp. 3480–3491, 2017.View at: Google Scholar
Y. Cui, P. Bangalore, and L. B. Tjernberg, “An Anomaly Detection Approach Based on machine learning and scada data for condition monitoring of wind turbines,” in Proceedings of the 2018 IEEE International Conference on Probabilistic Methods Applied to Power Systems, Boise, ID, USA, June 2018.View at: Google Scholar
M. Du, J. Yi, and J. B. Guo, “Research on the application of neural networks on wind turbine SCADA data analysis,” Power System Technology, vol. 42, pp. 2200–2205, 2018.View at: Google Scholar
M. Cao, Y. Qiu, and Y. Feng, “Study of wind turbine fault diagnosis based on unscented kalman filter and SCADA data,” Energies, vol. 9, pp. 847–854, 2016.View at: Publisher Site | Google Scholar
Y. Zhao, D. Li, and A. Dong, “Fault prediction and diagnosis of wind turbine generators using SCADA data,” Energies, vol. 10, pp. 972–984, 2017.View at: Google Scholar
Y. Liang and R. M. Fang, “On-line evaluation method for wind turbine status based on SCADA and support vector regression,” Automation of Electric Power Systems, vol. 37, no. 14, pp. 7–12, 2013.View at: Google Scholar
P. Sun, J. Li, C. Wang, and X. Lei, “A generalized model for wind turbine anomaly identification based on SCADA data,” Applied Energy, vol. 168, no. 168, pp. 550–567, 2016.View at: Publisher Site | Google Scholar
A. Zaher, S. D. J. McArthur, and D. G. Patel, “Online wind turbine fault detection through automated SCADA data analysis,” Wind Energy, vol. 12, no. 6, pp. 574–593, 2009.View at: Publisher Site | Google Scholar
P. B. Dao, W. J. Staszewski, T. Barszcz, and T. Uhl, “Condition monitoring and fault detection in wind turbines based on cointegration analysis of SCADA data,” Renewable Energy, vol. 116, pp. 107–122, 2018.View at: Publisher Site | Google Scholar
W. G. Garlick, R. Dixon, and S. J. Watson, “A model-based approach to wind turbine condition monitoring using SCADA data,” in Proceedings of the 20th International Conference on Systems Engineering, Coventry, UK, September 2009.View at: Google Scholar
H. S. Zhao and H. H. Liu, “Fault detection of wind turbine main bear based on deep learning network,” Acta Energiae Solaris Sinica, vol. 39, pp. 588–595, 2018.View at: Google Scholar
Y. Qiu, M. Cao, Y. Feng, and D. Infield, “An electro-thermal analysis of a variable-speed doubly-fed induction generator in a wind turbine,” Energies, vol. 8, no. 5, pp. 3386–3402, 2015.View at: Publisher Site | Google Scholar
Y. Qiu, Y. Feng, J. Sun, W. Zhang, and D. Infield, “Applying thermophysics for wind turbine drivetrain fault diagnosis using SCADA data,” IET Renewable Power Generation, vol. 10, no. 5, pp. 661–668, 2016.View at: Publisher Site | Google Scholar
B. Hahn, M. Durstewitz, and K. Rohrig, Reliability of Wind turbines, Springer, Berlin, Germany, 2007.View at: Publisher Site
G. E. Hinton and S. Osindero, “The A fast learning algorithm for deep belief nets,” Neural Computation, vol. 18, pp. 1527–1554, 2006.View at: Publisher Site | Google Scholar
L. Zhang, H. L. Gao, and J. Wen, “A deep learning-based recognition method for degradation monitoring of ball screw with multi-sensor data fusion,” Microelectronics Reliability, vol. 75, pp. 215–222, 2017.View at: Publisher Site | Google Scholar
P. Tamilselvan and P. F. Wang, “Failure diagnosis using deep belief learning based health state classification,” Reliability Engineering & System Safety, vol. 115, pp. 124–135, 2013.View at: Publisher Site | Google Scholar
L. Deng, G. Hinton, and B. Kingsbury, “New types of deep neural network learning for speech recognition and related applications: an overview,” in Proceedings of the IEEE ICA, pp. 8599–8603, Jinan, China, June 2013.View at: Google Scholar
C. M. Wang, “Research on the fault diagnosis method based on deep belief network for main bearing of wind turbine,” Process Automation Instrumentation, vol. 39, pp. 33–37, 2018.View at: Google Scholar
G. E. Hinton, A Practical Guide to Training Restricted Boltzmann Machines, Springer-Verlag, Berlin, Germany, 2012.View at: Publisher Site
D. H. Ackley, G. E. Hinton, and T. J. Sejnowski, “A learning algorithm for Boltzmann machines,” Cognitive Science, vol. 9, pp. 147–169, 1985.View at: Publisher Site | Google Scholar
P. Smolensky, “Information processing in dynamical systems: foundations of Harmony Theory,” in Parallel Distributed Processing: Explorations in the Microstructure of Cognition, D. E. Rumelhart and J. L. Mcclelland, Eds., pp. 194–281, MIT Press, Cambridge, MA, USA, 1986.View at: Google Scholar
S. Mirjalili and A. Lewis, “The whale optimization algorithm,” Advances in Engineering Software, vol. 95, no. 95, pp. 51–67, 2016.View at: Publisher Site | Google Scholar
P. F. Niu, Z. L. Wu, and Y. P. Ma, “Prediction of steam turbine heat consumption rate based on whale optimization algorithm,” CIE Journal, vol. 68, pp. 1049–1057, 2017.View at: Google Scholar