Data-Driven Methodologies Used for Wind Turbines O&M Tasks: History
View Latest Version
Please note this is an old version of this entry, which may differ significantly from the current revision.
Contributor:
Ravi Kumar Pandit
,
Davide Astolfi
,
Isidro Durazo Cardenas

 Wind power is one of the most sustainable and eco-friendly energy sources. With the rapid development of wind turbines (WTs), there is an increasing need to lower the Cost of Energy (COE) of wind power. Predictive maintenance techniques that leverage past failures to learn from and forecast failure and the remaining usable life of various wind turbines can significantly reduce operation and maintenance (O&M) expenses. In this article, recent advancements in data-driven models for condition monitoring and predictive maintenance of wind turbines’ critical components (e.g., bearing, gearbox, generator, blade pitch) are reviewed. The entry categorizes these models according to data-driven procedures, such as data descriptions, data pre-processing, feature extraction and selection, model selection (classification, regression), validation, and decision making. The findings after reviewing extensive relevant articles suggest that (a) SCADA (supervisory control and data acquisition) data are widely used as they are available at low cost and are extremely practical (due to the 10 min averaging time), but their use is in some sense nonspecific. (b) Unstructured data and pre-processing remain a significant challenge and consume a significant time of whole machine learning model development. (c) The trade-off between the complexity of the vibration analysis and the applicability of the results deserves further development, especially with regards to drivetrain faults. (d) Most of the proposed techniques focus on gearbox and bearings, and there is a need to apply these models to other wind turbine components. 

  • wind turbine
  • predictive maintenance
  • big data computation

1. Bearing Failure

An important factor in unexpected maintenance, repairs, and replacement downtime in energy generation is bearing failures in wind turbines. This primary cost failure type increases the O&M costs for the energy operator as well as the customer’s electricity bills. According to the Gearbox Reliability Database (GRD) of the National Renewable Energy Laboratory (NREL), 76% of gearbox failures were attributable to bearing failures, while 17% were attributable to gear failures. This demonstrates the value of dependable bearings and gearboxes for the functioning of wind turbines to the economy and society [1].
Common causes for bearing failure are excessive load, fatigue, contamination, misalignments, overheating etc. The latter phenomenon is addressed in several papers, as for example, in [2]. As the approach considered is fault estimation, it can be categorized as model-based and data-driven. In that method, the aim was to determine the bearing fault at least 33 days prior using statistical features of residuals evaluating Bayesian state prediction. An artificial neural network (ANN) was chosen for modelling the temperatures of the main bearing component. Prediction of the event bearing over-temperature was possible, but over a limited set of time series, it could give confidence one month prior to the failure of the bearing. Further analysis is necessary for the accuracy of other event failures for different time series.
Jian et al. introduced deep learning for WT condition monitoring in [3]. An adaptive elastic network, a convolutional neural network (CNN), and an LSTM (Long Short-Term Memory) were coupled to perform feature extraction, dimension reduction, and classification. The gradient explosion and overfitting problems were resolved by this technique, lowering the prediction error. Before the data gathered by SCADA are analyzed, the suitable variables associated with the transmission-bearing temperature must be selected as the research object. Gearbox-bearing temperatures are impacted differently by variables with various relationships [3]. In [4], LSTM is used to solve the problem of the exploding gradient and vanishing gradient when the layer of the network increases and the subsequent node perceptions for the previous nodes become weak. In [5], a prognosis indicator for the health status of the main bearing of wind turbines is formulated based on the number of times the residual between the measured component temperature and the model-based estimate exceeds a certain threshold. SCADA data are employed for this aim through artificial neural network modelling. In [6], a physics-domain method for high-speed shaft axial crack prognosis is formulated and validated by using SCADA data with an averaging time of ten minutes. The frictional energy and the electrical power are employed as damage metrics, and it is shown that the advantage of physics-domain models is not only the capability of estimating damage probabilities but also a deeper insight into the failure mechanism. In [7], a mixture of physics-domain and data-driven modelling is employed and applied to two test cases: a planetary and a high-speed bearing fault. It is shown that the combined approach is superior to a purely data-driven method for fault prognosis.
The huge data problem in WT can be resolved by using sparse representation techniques to greatly compress the observed signal into a few nonzero coefficients as a signal projection on dictionaries. Typically, measured current/vibration signals are used to extract the defect characteristics of wind turbine bearings. It is crucial to keep in mind that this approach will fall short if the model cannot accurately capture the mathematical operation of the dictionary. Although the K-singular value decomposition (K-SVD) and general principal component analysis (GPCA) are more adaptable to describe signal data, the learning procedure is difficult and time consuming. To determine the failure of high-speed shaft bearing (HSSB), a vibration-based diagnosis methodology for wind turbine high-speed bearing is proposed using principal component analysis (PCA) [8]. Though this method fails to predict the exact date of failure in advance, it showed good accuracy in monitoring the health of the component. In [9], vibrations collected with a frequency of the order of 16,000 Hz by eight acceleration transducers placed in the drivetrain are processed through an artificial neural network in order to estimate the remaining useful life of high-speed shaft bearings. M. Kordestani et al. [10] proposed a fault detection and diagnosis (FDD) method consisting of feature extraction/feature selection and an adaptive neuro-fuzzy inference system (ANFIS) method. The feature extraction and selection phase identified proper features to capture the nonlinear dynamics of the failure. Then, the ANFIS classifier was used to diagnose the failure type using the extracted features.
To identify wind turbine bearing issues, Ref. [11] provide a feature selection and learning vector quantization (LVQ) neural network technique combination. The right features are extracted using Empirical Mode Decomposition (EMD). The LVQ neural network is then utilized to categorize different failures. The results of the experimental tests show that the suggested fault diagnosis approach is highly accurate. According to [12], the modulation signal bispectrum (MSB) detector is used to identify bearing problems in DFIGs of WT. Overlapping segmentation is suggested as a way to increase computational accuracy with sparse data. The MSB algorithm was discovered to be an efficient, space-saving method to retrieve modulation information from data, while traditional methods based on a single spectrum were concerned only with the amplitude. Quadratic phase coupling (QPC) and amplitude modulation (AM) were caused by vibration caused by bearing faults.
The work in [13] deals with the estimation of the remaining useful lifetime of wind turbine bearings through the analysis of vibration data collected with a frequency of 97,000 kHz. The building blocks of the algorithm are wavelet transform pre-processing, Bayesian state-space modelling, and particle filter. In [14], the remaining useful lifetime of a high-speed shaft wind turbine bearing is estimated based on processing the statistical features of vibration signals collected with a frequency in the order of 100 Hz. Vibration data from a real-world wind turbine are analyzed, and the prognostic capability of several signal processing techniques is compared. Data are collected at nominal speed and are sampled at 97,656 Hz for 6 s. It is shown that spectral kurtosis followed by envelope analysis provides early fault detection compared to the other techniques employed. In [15], the angular velocity error at the various stages of the gearbox is selected as a target to monitor for individuating bearing faults. In [16], the proposed approach is based on the co-integration of multiple industrial data types, with different sampling times. Using SCADA data averaged on a 10-minute basis, the main bearing fault is identified by monitoring the residuals between measured and model-estimated bearing temperatures. The individuation of the precise location of the damage is corroborated by the analysis of vibration signals collected by the industrial Turbine Condition Monitoring (TCM) system: the statistical novelty between healthy and faulty wind turbines is identified through Principal Component Analysis (PCA) of a set of features.
The above study indeed solves by integration of multiple data sources an issue that is quite common in SCADA-based wind turbine condition monitoring. Actually, the main bearing temperature is often selected as the target temperature to monitor, typically based on data-driven considerations. Yet, there is a physical reason why the temperature of this component is quite responsive to incoming faults: the main bearing is a large component, which rotates relatively slowly, and it is, therefore, reasonable that it releases much heat in a way that can be easily captured (together with its anomalies) by data-driven algorithms. In fact, in [17], it is shown that by monitoring the temperature of the main bearing, it is possible to diagnose a stator fault. Finally, it is worth noting that a few studies [18] approach the diagnosis of bearing faults through the analysis of tower sound and vibration, without knowing the transfer function between the bearings and the tower. Statistical analysis techniques are used for distinguishing features of faulty and healthy vibration signals.

2. Gearbox Failure

In [19], the testing is based on Condition Monitoring System (CMS) data from 10 WTs to detect the common failures in the gearbox HS module. The signal correlation with RMS values was found to be good for detecting progressive failures such as HS bearing pitting or shaft cracks at least one month in advance but was not suitable for detecting gear tooth fracture. Using correlation and an extreme vibration model, the peak value did a better job detecting gear tooth fracture. As the extreme vibration model does not rely on historical data, it can be used for new WTs or WTs with missing CMS history. The “delta RMS” plot gives insight into the severity of the failure. One of the limitations of this model is that changes in RMS vibrations are sensitive only to high shaft revolutions and therefore can only be used for high-speed modules of the gearbox, with higher shaft revolutions than other modules.
Compared to the traditional gradient-based training algorithm widely used in the single-hidden layer feed-forward neural network, Extreme Learning Methods (ELMs) can randomly choose the input weights and hidden biases and need not be tuned in the training process. Therefore, the ELM algorithm can dramatically reduce the learning time. The drawbacks of the traditional gradient-based training algorithms, such as overtraining, high computational time, and trapping at local minima, can all be overcome by the ELM algorithm, as it randomly chooses the input weights and hidden biases and needs not to be tuned in the training process [20].
In [21], a Deep Belief Network (DBN) is used to merge in a purely data-driven way the measurements collected by four vibration sensors attached on the casing of the gearbox from low-speed to high-speed stages of a wind turbine gearbox, for which accelerated lifetime tests are conducted in the laboratory. The Wiener model is employed to describe the process of gearbox degradation and to predict the remaining useful lifetime. In [22], accelerated lifetime tests are performed as well. A method for signal de-noising is proposed, which is based on complete ensemble empirical mode decomposition with adaptive noise and kernel principal component analysis. Multi-sensor fusion is performed using kernel principal component analysis and Hotelling statistics, and the estimation of the remaining useful lifetime is optimized through the fruit fly algorithm.
In [23], real-world data sets from three Suzlon wind turbines (two healthy and one faulty) are analyzed. Vibrations are measured at the pinion gears with a sampling rate of 97.656 (kHz) and a recording time of 6 s. The Signal Intensity Estimator (SIE) method and the principal component analysis of the statistical features of the vibration signals are employed for estimating the remaining useful lifetime. This work indicates that it is possible to extract meaningful prognosis information from highly modulated real-world data, such as those originating from wind turbine gears. The SIE method is also employed in [24].
In [4], the LSTM prediction model is implemented to indicate abnormal behavior in the gearbox by monitoring the gearbox bearing rise in temperature. It should be noticed that SCADA data are much more used for bearing condition monitoring, i.e., as opposed to gearbox data. This occurs because the heat released by bearings is easier to use as a target to monitor, while the precise location of the gearbox fault requires specific measurements, which are collected through accelerometers that are placed ad hoc. SCADA data are employed in [25] for a long-term fatigue life assessment based on a three-stage gearbox multibody dynamic model. The main result is that the most vulnerable part of the gearbox is the sun gears, which are mostly stressed at wind speeds higher than 10 m/s. In [26], a method for prognosis based on SCADA data is formulated, which employs Gaussian process and principal component analysis. A fleet of 24 faulty wind turbines is selected for validating the model. The detection rate is 79% and 76% component-wise, where the most important involved components are the gearbox and the generator. In [27], the health status of wind turbine main components (gearbox and generator, mainly) is assessed through a regression model based on an Extreme Learning Machine (ELM) strategy. Internal temperatures are simulated by using as input variables other internal temperatures, environmental variables, and working parameters of the machine. The health status of the component is assessed by performing a linear regression between simulated and measured target variables once per day and analyzing how much the slope deviates from the unity. The prognosis is formulated by analyzing the time evolution of such estimated slopes. High-frequency SCADA data are employed in [28], where a normal-behavior model for the gearbox oil temperature is set up and a one-class Support Vector Machine (SVM) classifier is employed for setting a threshold for anomaly detection. A sensitivity study on the data averaging time is performed, and it is shown that the trade-off is non-trivial, in the sense that, the higher the frequency, and the higher the information but at the same time the higher the noise. In [29], the measured one-phase stator current of a wind turbine is processed in order to extract information on the health status of the gearbox through the adaptive neuro-fuzzy inference system (ANFIS) and particle filtering (PF) approaches.

3. Generator Failure

Condition monitoring of wind turbine generators is a fundamental task given that this component ranks in the top three regarding failure rates and downtime [30][31]. The diagnosis of generator faults has more or less the same balance between pros and cons as the other rotating elements. On the one hand, TCM systems recording vibrations at the sub-component are costly, and inspections based on voltages and current analysis are even more costly [32][33]. On the other hand, SCADA data have a much lower cost but their diagnostic and prognostic capabilities are questionable, especially regarding electrical faults. Nevertheless, some interesting attempts at using SCADA data for generator diagnosis and prognosis are being developed. The method described in [34] does not require any additional hardware beyond the SCADA system for determining WT generator failure. The authors propose a method to predict the remaining useful life (RUL) of generators using the Anomaly Operation Index (AOI), which determines performance degradation in runtime. SCADA monitors the run-time operation condition of the wind turbine, such as temperature, speed, and power. Such information may be leveraged to support the generator’s prognosis. This method proposes an autoregressive integrated moving average (ARIMA)-based statistical model to conduct online prognostics and a time series analysis-based RUL estimation method to provide accurate RUL prediction.
The SCADA system is the only new hardware needed for the method proposed in [33] to determine WT generator failure. The Anomaly Operation Index (AOI), which measures performance degradation in runtime, is used by the authors to offer a method for estimating the remaining usable life (RUL) of generators. SCADA keeps track of the wind turbine’s operational parameters during operation, including temperature, speed, and power. The generator’s prediction may be strengthened with the use of such information. This approach suggests using a time series analysis-based RUL estimation method in combination with an autoregressive integrated moving average (ARIMA)-based statistical model to undertake online prognostication. In order to forecast performance, AOI is analyzed using historical failure data from the past. The normal and anomaly can be determined at runtime with the use of sophisticated data mining techniques such as DBScan and the SVM algorithm. This makes this experiment a more effective and cost-effective technique of failure detection. In [35], a series of phenomena related to generator incoming faults is individuated through SCADA data analysis. These include miscorrelation between the rotational speed and active or reactive power, anomalous heating, and anomalies related to the shaft torque. In [36], the Mahalanobis distance between appropriately selected features is employed to diagnose generator bearing
The peculiarity of the study in [37] is the attempt at diagnosing an electrical generator fault using SCADA data. A normal behavior model for the power and for the voltage and current of each phase is constructed through a support vector regression, where the input variables are working parameters (such as the blade pitch and rotational speed) and generator temperatures. The features are pre-processed through PCA, and a threshold for alarm raising is identified from the statistical properties of the residuals between measurements and model estimates. In that work, it is shown that the alarm raising occurs two weeks before a real-world electrical fault of a wind turbine generator, and this anticipates the alarm log book collected by the SCADA control system.
Similarly, in [33][36], an anomaly operation index is formulated based on the number of anomalous points in the feature space with respect to normal behavior. A one-class support vector machine is employed for anomaly identification, and the AOI is de-trended by employing a moving average. An Autoregressive Integrated Moving Average (ARIMA) method allows forecasting the remaining useful life, and the results show that the model successfully forecasts failures of the generator while providing a 21-day lead time for the operators to plan the necessary maintenance action.

4. Blade Pitch System Failure

Attention has been growing in the literature regarding the assessment of the health status of the blade pitch systems because evidence is being collected about the decisive role of blade pitch degradation on wind turbine performance worsening. In [38], an a priori knowledge-based adaptive neuro-fuzzy inference system is employed with the aim to achieve automated detection of significant pitch faults. The method is tested on variable speed, variable pitch wind turbines, and on variable pitch, fixed-speed wind turbines; 49 GB of SCADA data from several companies are analyzed by the authors. Furthermore, in [38], an interesting discussion is conducted on the pros and cons of the two methods for controlling the blade pitch of a wind turbine, which are hydraulic and electrical. While each blade is controlled by an electric servo-motor connected to a gearbox that lowers the motor speed to apply torque to the blades, in the case of hydraulic pitch control, actuators in the rotor hubs provide torque directly or via mechanical linkages. The advantages of the former type are the simplicity and the high torque that can be exerted. This is fundamental if one takes into account that such a mechanism is also responsible for stopping the wind turbine in the case of gusts. The diffusion of electrical pitch control is growing, but at present, the majority of wind turbines have hydraulic blade pitch control.
Several studies have recently been devoted to the investigation of the long-term health state of hydraulic vs. electrical blade pitch, based on SCADA data analysis. In [39], it is shown that the aging of electrical blade pitch motors leads to performance worsening over time, which is quite limited. In [40][41], through comparative test case analysis, it is shown that the aging of hydraulic blade pitch actuators likely leads to performance worsening over time, which can also be severe. This is due to pressure losses, which lead to the fact that the wind turbine operates at a non-optimal working point and, in turn, also in the full aerodynamic load regime where less power can be extracted for a given rotational speed. Intelligent predictive maintenance strategies should therefore be developed for optimal management of the blade pitch health, which is an overlooked topic in wind energy practice and literature.
The aging of the blade pitch systems and their health state prognosis are also addressed in the recent study [42], where several indicators are formulated: the behavior of the power coefficient, the power fluctuations above the rated speed (the higher the fluctuations, the more degraded the blade pitch system), overheating, and failure rates. The diagnosis of electrical blade pitch faults through SCADA data analysis is pursued, for example, in [43], where an optimized relevance vector machine regression is set up for the blade pitch motor power upon feature selection through the random forest algorithm. In total, 38 pitch system fault cases are analyzed, which provides an interesting overview: nine encoder failures, seven pitch controller failures, seven electric motor failures, eight slip ring failures, three limit switch failures, two backup battery failures, and two stud failures.

5. Yaw Failure

The yaw mechanism of wind turbines is quite delicate because the yaw motion needs to counteract the large inertia of the rotor in order to achieve the best possible orientation with respect to incoming wind that has rapid fluctuations. The yaw movement is typically achieved through yaw motors that undergo alternating stress and might suffer from mechanical damages, such as tooth face abrasion, gearbox failure, yaw bearing failure, and brake actuator failure.
Vibration analysis techniques have been employed for detecting slewing bearing damages, for example, in [44][45]. In [46], a method based on circular domain resampling and piecewise aggregate approximation is formulated and validated through a highly accelerated life test. It is shown that the incoming fault can be identified through the statistical features of the processed signal. In [12], acoustical damage detection of a wind turbine yaw system is proposed. A real-world experiment is proposed: a microphone is mounted inside the nacelle of a 1.5 MW WT sited in China. The collected measurements have a frequency of 64,000 Hz. The sound pressure levels are extracted from the raw signals, and a data discretization method based on a self-organizing map and information gain rate are employed. Finally, a Bayesian Network diagnostic model is used to detect the incoming fault. In [47], several types of faults are simulated and diagnosed using a data-driven method based on a benchmark model of wind turbine component’s functioning. It follows the construction of robust residual generators using the observer-based residual generation technique, and one of the diagnosed faults regards the yaw actuator.
The yaw system of a wind turbine might be affected by systematic error (also known as zero-point shift), which can be relevantly non-vanishing if the wind vane sensor is incorrectly aligned with the rotor shaft due to wind vane defects, incorrect installation or maintenance, or the aging of the machine. Numerous research papers have been devoted to the individuation of such a type of fault through SCADA data analysis, despite it being non-trivial to formulate reliable and general algorithms. In [48], (Jing, 2020), the power curve is analyzed, which is the relationship between the wind speed measured by the nacelle anemometer and the extracted power. The rationale for analyzing the power curve for individuating a systematic yaw error is the expectation that an underperformance should be visible. However, this task is challenging due to the multivariate dependence of the wind turbine power on environmental conditions and working parameters, and adequate data-mining methods are required. Furthermore, in [48], for example, the power curve is studied per interval of yaw error. In [49], a similar approach is employed, but the power curve is analyzed through a different model, which is a least-square B-spline approximation. In [50], a multivariate data-driven power curve model is employed, in the form of Gaussian process regression that takes as input the rotational speed and the blade pitch. The systematic yaw error is individuated from the mismatch between the measured power and model estimate. In a yaw error case study performed in [51], two methodologies—Gaussian process and IEC binned power curve—are used to predict the anomaly. In [52], the power curve is also analyzed with a non-trivial data rejection algorithm. The idea of diagnosing the systematic yaw error by observing under-performance is shown in [53]. The peculiarity of that study is that the data are labelled, in the sense that a utility-scale wind turbine installed at a research facility was controllable by the authors, who imposed yaw offsets and therefore had at their disposal ground truth associated with the observed behavior.
The limitations of the above-cited studies about the systematic yaw error are the lack of validation, in the sense that it is unclear if one or more data-driven algorithms work accurately for most wind turbine models available on the market. It is desirable to formulate a comprehensive approach, similar to what has been done in [42] for the blade pitch health status, based on the observation of several manifestations, such as under-performance, augmented tower vibrations, heating, and anomalous blade loads.

6. Underperformance and Power Coefficient

How effectively a WT turns wind energy into electricity is shown by the power coefficient (Cp). Researchers attempted to create an adaptive neuro-fuzzy inference system (ANFIS) in Table 1 to calculate the power coefficient of the WT. Applications for ANFIS include forecasting, managing, diagnosing, and classifying. This method combines a neural network with the Takagi–Sugeno fuzzy inference system.
Table 1. Statistical properties of wind turbine data [54].
Wind Turbine Parameters Average Value Max. Value Min. Value
TSR 0.6640 30.758 0.4119
Pitch angle (degree) 0 5 −5
Power coefficient 0.1707 0.4868 −5.4984
In an ideal situation, it would be anticipated that all wind energy would be transformed into power (electricity); however, in reality, this is not feasible for a variety of clear reasons: 53% of the wind energy input is the maximum amount of energy a wind turbine can output. In the experiment described in [54], the optimal result was observed when the input was a 6-bell-shaped membership function. The Gaussian method also provides a close approximation of the optimal solution. The neuro-fuzzy system using a hybrid learning algorithm develops a fuzzy rule to obtain a minimum error. The model’s accuracy is dependent upon the training and test dataset provided to the algorithm. Hence, close attention should be given to the input values. High errors may lead to overfitting of the model. ANFIS is an adaptive and fast-speed operation. Other hybrid learning systems should also be adapted for a comparative study.
The decrease of the power coefficient with respect to the normal behavior, which is typically established through data-driven analyses, can be employed for individuating faults that have the peculiarity of resulting in noticeable underperformance. Mechanical damage to rotating elements is typically characterized by negligible under-performance, but this is not the case, for example, of systematic errors affecting wind turbine operation, such as the systematic yaw error. Actually, the decrease of the observed power coefficient is targeted for the diagnosis of the faults in [40][55].

7. Anomaly Detection

SCADA data comprise measures such as active and reactive power, generator current and voltages, wind speed, generator shaft speed, generator, gearbox, and nacelle temperatures, among others. The data are normally recorded at 10-minute intervals to reduce the sent data bandwidth and storage. The performing of statistical analysis on various trends within the data can determine when the turbine enters a time of sub-optimal performance or if there is a fault in the component of the system. Lily Hu et al. proposed in the paper [56] a way to derive features from SCADA data based on domain knowledge. These extra features are based on three factors: (1) knowledge of the physical quantities the SCADA sensors measure; (2) time series behavior of the sensor measurements; and (3) statistical features, see Table 2.
Table 2. Example Features from knowledge of WTs [56].
Average of Front and Rear Bearing temp
Rotor temp
Stator temp
Difference Between Max. and min. wind speed
Max. and average wind speed
Min. and average wind speed
Front and rear bearing temperatures
Nacelle ambient temperature
Generator temperature and nacelle temperature
Ratio of Average power to available power (from wind, technical reasons, force, external reasons)
This enables higher classification scores and improved detection of faults while using fewer features—an improvement in the F1 score of almost 20% while using a similar number of features. This method allows the freedom to decide the number of features to select for the machine learning algorithm best suited for the selected parameters. It is a very smart and efficient manner of analyzing the data [56]. The blade pitch angle curve describes the nonlinear relationship between the pitch angle and hub height wind speed and can be used for the detection of faults. An SVM is an improved version of an artificial neural network (ANN) and is widely used for classification- and regression-related problems. The binning method is a benchmark data reduction approach for the wind industries, but its application is generally limited to the power curve; its use is seen in [57] to calculate the blade pitch curve. In [58], the performance of wind turbines is monitored through data-driven models for power, rotor speed, and blade pitch curves, having the wind speed as input. A multivariate outlier detection approach based on k-means clustering and the Mahalanobis distance is applied. In [59][60], a similar approach is formulated for operation curves that do not employ the wind speed as input: namely, the rotor speed-power and the blade pitch-power curves, which are modelled through a support vector regression with Gaussian kernel.

This entry is adapted from the peer-reviewed paper 10.3390/en16041654

References

  1. Bearing Damage and Failure Analysis. Available online: https://www.skf.com/binaries/pub12/Images/0901d1968064c148-Bearing-failures---14219_2-EN_tcm_12-297619.pdf (accessed on 20 January 2023).
  2. Herp, J.; Ramezani, M.H.; Bach-Andersen, M.; Pedersen, N.L.; Nadimi, E.S. Bayesian state prediction of wind turbine bearing failure. Renew. Energy 2018, 116, 164–172.
  3. Fu, J.; Chu, J.; Guo, P.; Chen, Z. Condition Monitoring of Wind Turbine Gearbox Bearing Based on Deep Learning Model. IEEE Access 2019, 7, 57078–57087.
  4. Feng, B.; Zhang, D.; Si, Y.; Tian, X.; Qian, P. A condition monitoring method of wind turbines based on Long Short-Term Memory neural network. In Proceedings of the 2019 25th International Conference on Automation and Computing (ICAC), Lancaster, UK, 5–7 September 2019; pp. 1–4.
  5. Encalada-Dávila, Á.; Puruncajas, B.; Tutivén, C.; Vidal, Y. Wind Turbine Main Bearing Fault Prognosis Based Solely on SCADA Data. Sensors 2021, 21, 2228.
  6. Guo, Y.; Sheng, S.; Phillips, C.; Keller, J.; Veers, P.; Williams, L. A methodology for reliability assessment and prognosis of bearing axial cracking in wind turbine gearboxes. Renew. Sustain. Energy Rev. 2020, 127, 109888.
  7. Corley, B.; Koukoura, S.; Carroll, J.; McDonald, A. Combination of Thermal Modelling and Machine Learning Approaches for Fault Detection in Wind Turbine Gearboxes. Energies 2021, 14, 1375.
  8. Maatallah, H.; Fuente, M.J.; Ouni, K. Condition monitoring of wind turbine bearings progressive degradation using principal component analysis. In Proceedings of the 2020 Fifteenth International Conference on Ecological Vehicles and Renewable Energies (EVER), Monte Carlo, Monaco, 10–12 September 2020; pp. 1–6.
  9. Teng, W.; Zhang, X.; Liu, Y.; Kusiak, A.; Ma, Z. Prognosis of the Remaining Useful Life of Bearings in a Wind Turbine Gearbox. Energies 2017, 10, 32.
  10. Kordestani, M.; Rezamand, M.; Carriveau, R.; Ting, D.S.K.; Saif, M. Failure Diagnosis of Wind Turbine Bearing Using Feature Extraction and a Neuro-Fuzzy Inference System (ANFIS). In Advances in Computational Intelligence IWANN 2019; Lecture Notes in Computer Science; Rojas, I., Joya, G., Catala, A., Eds.; Springer: Cham, Switzerland, 2019; Volume 11506.
  11. Shi, X.; Li, W.; Gao, Q.; Guo, H. Research on Fault Classification of Wind Turbine Based on IMF Kurtosis and PSO-SOM-LVQ. In Proceedings of the 2017 IEEE 2nd Information Technology, Networking, Electronic and Automation Control Conference (ITNEC), Chengdu, China, 15–17 December 2017; pp. 191–196.
  12. Chen, X.; Xu, W.; Liu, Y.; Islam, M.R. Bearing Corrosion Failure Diagnosis of Doubly Fed Induction Generator in Wind Turbines Based on Stator Current Analysis. IEEE Trans. Ind. Electron. 2020, 67, 3419–3430.
  13. Wang, J.; Liang, Y.; Zheng, Y.; Gao, R.X.; Zhang, F. An integrated fault diagnosis and prognosis approach for predictive maintenance of wind turbine bearing with limited samples. Renew. Energy 2020, 145, 642–650.
  14. Elasha, F.; Shanbr, S.; Li, X.; Mba, D. Prognosis of a Wind Turbine Gearbox Bearing Using Supervised Machine Learning. Sensors 2019, 19, 3092.
  15. Nejad, A.R.; Odgaard, P.F.; Gao, Z.; Moan, T. A prognostic method for fault detection in wind turbine drivetrains. Eng. Fail. Anal. 2014, 42, 324–336.
  16. Natili, F.; Daga, A.P.; Castellani, F.; Garibaldi, L. Multi-Scale Wind Turbine Bearings Supervision Techniques Using Industrial SCADA and Vibration Data. Appl. Sci. 2021, 11, 6785.
  17. Xiao, X.; Liu, J.; Liu, D.; Tang, Y.; Zhang, F. Condition Monitoring of Wind Turbine Main Bearing Based on Multivariate Time Series Forecasting. Energies 2022, 15, 1951.
  18. Mollasalehi, E.; Wood, D.; Sun, Q. Indicative Fault Diagnosis of Wind Turbine Generator Bearings Using Tower Sound and Vibration. Energies 2017, 10, 1853.
  19. Igba, J.; Alemzadeh, K.; Durugbo, C.; Eiriksson, E.T. Analysing RMS and peak values of vibration signals for condition monitoring of wind turbine gearboxes. Renew. Energy 2016, 91, 90–106.
  20. Qian, P.; Ma, X.; Wang, Y. Condition monitoring of wind turbines based on extreme learning machine. In Proceedings of the 21st International Conference on Automation and Computing, Glasgow, UK, 11–12 September 2015; pp. 1–6.
  21. Pan, Y.; Hong, R.; Chen, J.; Wu, W. A hybrid DBN-SOM-PF-based prognostic approach of remaining useful life for wind turbine gearbox. Renew. Energy 2020, 152, 138–154.
  22. Pan, Y.; Hong, R.; Chen, J.; Singh, J.; Jia, X. Performance degradation assessment of a wind turbine gearbox based on multi-sensor data fusion. Mech. Mach. Theory 2019, 137, 509–526.
  23. Elforjani, M. Diagnosis and prognosis of real world wind turbine gears. Renew. Energy 2020, 147, 1676–1693.
  24. Elforjani, M.; Shanbr, S.; Bechhoefer, E. Detection of faulty high speed wind turbine bearing using signal intensity estimator technique. Wind Energy 2018, 21, 53–69.
  25. Qiu, Y.; Chen, L.; Feng, Y.; Xu, Y. An Approach of Quantifying Gear Fatigue Life for Wind Turbine Gearboxes Using Supervisory Control and Data Acquisition Data. Energies 2017, 10, 1084.
  26. Chen, N.; Yu, R.; Chen, Y.; Xie, H. Hierarchical method for wind turbine prognosis using SCADA data. IET Renew. Power Gener. 2017, 11, 403–410.
  27. Marti-Puig, P.; Blanco, M.A.; Serra-Serra, M.; Solé-Casals, J. Wind Turbine Prognosis Models Based on SCADA Data and Extreme Learning Machines. Appl. Sci. 2021, 11, 590.
  28. Verma, A.; Zappalá, D.; Sheng, S.; Watson, S.J. Wind turbine gearbox fault prognosis using high-frequency SCADA data. In Journal of Physics: Conference Series; IOP Publishing: Bristol, UK, 2022; Volume 2265, p. 032067.
  29. Cheng, F.; Qu, L.; Qiao, W. Fault prognosis and remaining useful life prediction of wind turbine gearboxes using current signal analysis. IEEE Trans. Sustain. Energy 2017, 9, 157–167.
  30. Carroll, J.; McDonald, A.; McMillan, D. Failure rate, repair time and unscheduled O&M cost analysis of offshore wind turbines. Wind Energy 2016, 19, 1107–1119.
  31. Castellani, F.; Pandit, R.; Natili, F.; Belcastro, F.; Astolfi, D. Advanced Methods for Wind Turbine Performance Analysis Based on SCADA Data and CFD Simulations. Energies 2023, 16, 1081.
  32. Artigao, E.; Honrubia-Escribano, A.; Gómez-Lázaro, E. In-service wind turbine DFIG diagnosis using current signature analysis. IEEE Trans. Ind. Electron. 2019, 67, 2262–2271.
  33. Artigao, E.; Sapena-Bano, A.; Honrubia-Escribano, A.; Martinez-Roman, J.; Puche-Panadero, R.; Gómez-Lázaro, E. Long-term operational data analysis of an in-service wind turbine DFIG. IEEE Access 2019, 7, 17896–17906.
  34. Zhao, Y.; Li, D.; Dong, A.; Lin, J.; Kang, D.; Shang, L. Fault prognosis of wind turbine generator using SCADA data. In Proceedings of the 2016 North American Power Symposium (NAPS), Denver, CO, USA, 18–20 September 2016; pp. 1–6.
  35. Zhao, Y.; Li, D.; Dong, A.; Kang, D.; Lv, Q.; Shang, L. Fault prediction and diagnosis of wind turbine generators using SCADA data. Energies 2017, 10, 1210.
  36. Jin, X.; Xu, Z.; Qiao, W. Condition monitoring of wind turbine generators using SCADA data analysis. IEEE Trans. Sustain. Energy 2020, 12, 202–210.
  37. Castellani, F.; Astolfi, D.; Natili, F. SCADA Data Analysis Methods for Diagnosis of Electrical Faults to Wind Turbine Generators. Appl. Sci. 2021, 11, 3307.
  38. Chen, B.; Matthews, P.C.; Tavner, P.J. Automated on-line fault prognosis for wind turbine pitch systems using supervisory control and data acquisition. IET Renew. Power Gener. 2015, 9, 503–513.
  39. Astolfi, D.; Castellani, F.; Lombardi, A.; Terzi, L. Data-driven wind turbine aging models. Electr. Power Syst. Res. 2021, 201, 107495.
  40. Astolfi, D.; Castellani, F.; Terzi, L. An operation data-based method for the diagnosis of zero-point shift of wind turbines yaw angle. J. Sol. Energy Eng. 2020, 142, 024501.
  41. Astolfi, D.; Pandit, R.; Celesti, L.; Vedovelli, M.; Lombardi, A.; Terzi, L. Data-Driven Assessment of Wind Turbine Performance Decline with Age and Interpretation Based on Comparative Test Case Analysis. Sensors 2022, 22, 3180.
  42. Wei, L.; Qian, Z.; Zareipour, H.; Zhang, F. Comprehensive aging assessment of pitch systems combining SCADA and failure data. IET Renew. Power Gener. 2022, 16, 198–210.
  43. Wei, L.; Qian, Z.; Zareipour, H. Wind turbine pitch system condition monitoring and fault detection based on optimized relevance vector machine regression. IEEE Trans. Sustain. Energy 2019, 11, 2326–2336.
  44. Niu, M.Z.; Chen, J.; Feng, Y.; Wang, H. Fault diagnosis of wind power yawing slewing bearing based on support vector machine. J. Nanjing Univ. Technol. 2014, 36, 117–122.
  45. Žvokelj, M.; Zupan, S.; Prebil, I. EEMD-based multiscale ICA method for slewing bearing fault detection and diagnosis. J. Sound Vib. 2016, 370, 394–423.
  46. Pan, Y.; Hong, R.; Chen, J.; Qin, Z.; Feng, Y. Incipient fault detection of wind turbine large-size slewing bearing based on circular domain. Measurement 2019, 137, 130–142.
  47. Yin, S.; Wang, G.; Karimi, H.R. Data-driven design of robust fault detection system for wind turbines. Mechatronics 2014, 24, 298–306.
  48. Jing, B.; Qian, Z.; Pei, Y.; Zhang, L.; Yang, T. Improving wind turbine efficiency through detection and calibration of yaw misalignment. Renew. Energy 2020, 160, 1217–1227.
  49. Bao, Y.; Yang, Q. A data-mining compensation approach for yaw misalignment on wind turbine. IEEE Trans. Ind. Inform. 2021, 17, 8154–8164.
  50. Pandit, R.; Infield, D.; Dodwell, T. Operational variables for improving industrial wind turbine yaw misalignment early fault detection capabilities using data-driven techniques. IEEE Trans. Instrum. Meas. 2021, 70, 1–8.
  51. Pandit, R.K.; Infield, D. SCADA-based wind turbine anomaly detection using Gaussian process models for wind turbine condition monitoring purposes. IET Renew. Power Gener. 2018, 12, 1249–1255.
  52. Qu, C.; Lin, Z.; Chen, P.; Liu, J.; Chen, Z.; Xie, Z. An improved data-driven methodology and field-test verification of yaw misalignment calibration on wind turbines. Energy Convers. Manag. 2022, 266, 115786.
  53. Gao, L.; Hong, J. Data-driven yaw misalignment correction for utility-scale wind turbines. J. Renew. Sustain. Energy 2021, 13, 063302.
  54. Asghar, A.; Liu, X. Estimation of wind turbine power coefficient by adaptive neuro-fuzzy methodology. Neurocomputing 2017, 238, 227–233.
  55. Yang, J.; Wang, L.; Song, D.; Huang, C.; Huang, L.; Wang, J. Incorporating environmental impacts into zero-point shifting diagnosis of wind turbines yaw angle. Energy 2022, 238, 121762.
  56. Hu, R.L.; Leahy, K.; Konstantakopoulos, I.C.; Auslander, D.M.; Spanos, C.J.; Agogino, A.M. Using Domain Knowledge Features for Wind Turbine Diagnostics. In Proceedings of the 2016 15th IEEE International Conference on Machine Learning and Applications (ICMLA), Anaheim, CA, USA, 18–20 December 2016; pp. 300–307.
  57. Pandit, R.K.; Infield, D. Comparative assessments of binned and support vector regression-based blade pitch curve of a wind turbine for the purpose of condition monitoring. Int. J. Energy Environ. Eng. 2019, 10, 181–188.
  58. Kusiak, A.; Verma, A. Monitoring Wind Farms With Performance Curves. IEEE Trans. Sustain. Energy 2013, 4, 192–199.
  59. Pandit, R.; Infield, D. Comparative Analysis of Binning and Support Vector Regression for Wind Turbine Rotor Speed Based Power Curve Use in Condition Monitoring. In Proceedings of the IEEE 53rd International Universities Power Engineering Conference (UPEC), Glasgow, UK, 4–7 September 2018; pp. 1–6.
  60. Pandit, R.; Infield, D. Comparative study of binning and gaussian process based rotor curves of a wind turbine for the purpose of condition monitoring. In Proceedings of the 3rd International Conference on Offshore Renewable Energy, Glasgow, UK, 28–29 August 2018.
More
© Text is available under the terms and conditions of the Creative Commons Attribution (CC BY) license; additional terms may apply. By using this site, you agree to the Terms and Conditions and Privacy Policy.
This entry is offline, you can click here to edit this entry!
Video Production Service
Feedback