Anoop Verma

Andrew Kusiake-mail: andrew-kusiak@uiowa.edu

Department of Mechanical and

Industrial Engineering,

3131 Seamans Center,

The University of Iowa,

Iowa City, IA 52242–1527

Fault Monitoring of WindTurbine Generator Brushes:A Data-Mining ApproachComponents of wind turbines are subjected to asymmetric loads caused by variable windconditions. Carbon brushes are critical components of the wind turbine generator.Adequately maintaining and detecting abnormalities in the carbon brushes early is essen-tial for proper turbine performance. In this paper, data-mining algorithms are appliedfor early prediction of carbon brush faults. Predicting generator brush faults early ena-bles timely maintenance or replacement of brushes. The results discussed in this paperare based on analyzing generator brush faults that occurred on 27 wind turbines. Thedatasets used to analyze faults were collected from the supervisory control and data ac-quisition (SCADA) systems installed at the wind turbines. Twenty-four data-mining mod-els are constructed to predict faults up to 12 h before the actual fault occurs. To increasethe prediction accuracy of the models discussed, a data balancing approach is used.Four data-mining algorithms were studied to evaluate the quality of the models for pre-dicting generator brush faults. Among the selected data-mining algorithms, the boostingtree algorithm provided the best prediction results. Research limitations attributed to theavailable datasets are discussed. [DOI: 10.1115/1.4005624]

Keywords: wind turbine, data-mining, prediction, Tomek links, generator brush wear,boosting tree, random forest

1 Introduction

Growing usage of wind turbines has made maintenance a prior-ity, and diagnosing potential faults is crucial to maintaining theturbine’s desired availability. Turbine generators are importantcomponents of wind turbines. Harsh environments, variable windspeeds, high humidity, make the turbine generators’ componentssusceptible to damage [1]. In particular, malfunctions of the gen-erator brushes lead to decreased performance of wind turbines.The common problems associated with generator brushes includedusting and damaged slip rings, which lead to arcing and flashover[1,2]. To mitigate brush malfunctions, different brush designshave been introduced, including brushless generators. Althoughthese modified designs minimize the problems caused by the car-bon plate brushes, planned maintenance and replacement remainsan issue on current day generators. Traditional maintenanceapproaches, including scheduled and corrective maintenance havedisadvantages in terms of (1) excessive cost, (2) time lost, and (3)loss of an actual component. Condition-based maintenance is arelatively new concept in wind industry aiming to minimize sys-tem damage by continuously monitoring the component’s condi-tion. Indeed, state-of-the-art research in maintenance has beendiscussed in the literature [3–7].

Supervisory control and data acquisitions (SCADA) systems in-stalled at wind turbines record performance parameters such aswind speed, output power, generator speed, and so on. In addition,SCADA systems record the systems’ potential faults expressed asstatus codes. Analyzing the operational parameters and statuscodes of a wind turbine can be useful for identifying the relation-ship between wind turbine performance parameters and actualfaults [8]. Further, data-mining techniques can be used to monitorperformance of wind turbines [9–11]. Other applications of data-mining in wind energy include: (1) predicting and optimizing

power [12–15]; (2) modeling abnormal behavior [16,17]; and (3)forecasting wind speed [18,19].

The research reported in this paper uses data-mining algorithmsto predict faults of a generator brush. Analysis was done based ondata obtained from 27 wind turbines. Data recorded at 10 minintervals over 4 months constitutes the available dataset for thisresearch.

This paper is organized as follows. The background of the faultdiscussed in this paper is presented in Sec. 2. The description of adataset and the data sampling techniques is provided in Sec. 3.The results of the analysis are presented in Sec. 4. Section 5 con-cludes the research and presents the future research directions.

2 Fault Description

The fault considered in this paper is attributed to the excessivewear of generator slip ring brushes, which is one of the most com-mon faults occurring in wind turbines. This fault leads to unneces-sary downtime because wind turbines are taken out of generationso that brushes can be replaced or cleaned. The turbine monitoringsystem, that is, the brush limit switch, reports excessive wear tomaintenance personnel. In the absence of timely maintenance,however, the turbine needs to be shutdown. Figure 1 illustrates thecause and effect diagram of “generator brush worn” faults. Designimperfections, turbine vibration, aerodynamic asymmetry, andgenerator overspeed are considered the prime causes of “generatorbrush worn” faults. Figure 1 clearly illustrates that the design isthe main cause of such faults; however, other factors such as tur-bine vibration and aerodynamic imbalance also need to be consid-ered. Research in component design has resulted in brushlessgenerators where the drawbacks associated with the generatorbrushes have been eliminated. For existing wind farms, however,the cost of brushless solutions cannot be justified. While the costto replace or maintain a worn out carbon brush cannot be elimi-nated, the cost corresponding to the lost operation can be mini-mized by predicting “generator brush worn” faults in advance. Atypical carbon brush lasts for 6 months under normal operation;

Contributed by the Solar Energy Division of ASME for publication in theJOURNAL OF SOLAR ENERGY ENGINEERING. Manuscript received January 6, 2011; finalmanuscript received December 6, 2011; published online February 27, 2012. Assoc.Editor: Spyros Voutsinas.

Journal of Solar Energy Engineering MAY 2012, Vol. 134 / 021001-1Copyright VC 2012 by ASME

Downloaded 12 Mar 2012 to 128.255.30.115. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

however, the presence of abnormalities in the system operationcan reduce the brush life to a factor of 30% and thus demand earlyreplacement. Figure 2 displays two carbon brushes that wore outone month before the anticipated time. In Sec. 2.1, “generatorbrush worn” faults are analyzed.

These brushes in Fig. 2 wore out a month prior to the antici-pated lifespan.

2.1 Fault Analysis Based Upon Status Data. The switch atthe end of the brush reports the information to the SCADA sys-tem, where the information is stored in the form of status codes.Both status and operational parameters are used in this paper. Sta-tus data is event triggered whereas, operational data is time trig-gered. Based on the data considered for this research, the delaybetween acknowledging the status code and maintenance actioncan be as long 168 h; thereafter, the wind turbine will shutdown.In the present research, data from 27 wind turbines were analyzed.The distribution of faults across 27 wind turbines is shown inFig. 3. The faultiest turbine, Turbine 14, is considered for furtheranalysis.

SCADA status data from Turbine 14 indicates that the“generator brush worn” fault occurred more than 100 times in the

month during which data was collected. To identify the statusesassociated with “generator brush fault,” 1 s snapshot files were an-alyzed. Snapshot files are automatically generated operationaldata files whenever some critical fault occurs in the turbine. Snap-shot files contain data at 1 s increments for the period of 7 minpreceding and 3 min after the fault occurrence. Wind turbine sta-tus data was mapped with the operational data and the statusesrecorded within the past 7 min and within the next 3 min of the“generator brush worn” fault being recorded. On average, morethan 30 different statuses were associated with this fault. Table 1illustrates the various status codes associated with the “generatorbrush worn” fault. The label “before” in Table 1 signifies the sta-tus causing “generator brush worn” faults, whereas the label“after” indicates the statuses generated due to the underlying faultitself. The analysis done here can be helpful to maintenance oper-ators in deciding between the need for actual maintenance orto fix the logic controller. Predicting the fault early, however,is required in either case. Figure 4 shows the power curve of aturbine (here Turbine 14) affected by the brush fault overthree consecutive days, Apr. 14, 2010—Apr. 16, 2010. The power

Fig. 1 Cause and effect diagram of the “generator brush worn” fault

Fig. 2 Samples of worn out carbon brushesFig. 3 Distribution of generator brush worn out faults across27 wind turbines

021001-2 / Vol. 134, MAY 2012 Transactions of the ASME

Downloaded 12 Mar 2012 to 128.255.30.115. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

curve had already started to deteriorate, but the turbine is stilloperational and producing power (Fig. 4(a)). The fault manifesteditself for a time period; therefore, when the deterioration wassignificant, the turbine was shutdown (Figs. 4(b) and 4(c)). It isalso possible that other faults could be responsible for powercurve deterioration; however, the impact of the “generator brushworn” fault prevailed in the two-day time period underconsideration.

3 Dataset Description and Preprocessing

The data available for the current research was collected bySCADA systems at 27 wind turbines recorded at 10 min intervals.In addition, the SCADA status data was used to label the faultconditions. Four months of the data collected from Mar. 1, 2010to July 31, 2010 was analyzed. The description of both data typesis provided next.

3.1 SCADA Operational Data. The SCADA system recordsdata on more than 100 parameters averaged over 10 min intervals(10-min data). Typical parameters recorded by SCADA can bebroadly categorized into (1) wind parameters, such as wind speedand wind deviations; (2) performance parameters, such as poweroutput, rotor speed, and blade pitch angle; (3) vibration parame-ters, such as tower acceleration and drive train acceleration; and(4) temperature parameters, such as bearing temperature and gear-box temperature.

3.2 SCADA Status Data. SCADA also records fault and sta-tus information whenever they occur. Status indicates a poten-tially emerging fault. For a typical large-scale turbine, more than400 different status codes are generated. A status can be triggereddue to internal factors (e.g., component temperature or vibrations)or external factors (e.g., wind speed). Depending on the severityof the occurring problem, status codes are identified as category 1to category 4. Category 1–3 statuses usually become fault in thefuture, whereas category 4 statuses represent a system update. The“generator brush worn” fault is labeled as a category 2 status.

3.3 Data Labeling. To label the output data, the status datacontaining the fault information was merged with the operationaldata. The output (Oi� 1) was labeled as fault or normal at timei� 1 based upon following logic

Oi�1 ¼fault; if tþ i�1o � tf � tþ io; i 2 10; 20; 30;:::f gnormal; otherwise

� �

(1)

In Eq. (1), tf is the time at which the fault is recorded in the statusdata, whereas tþ i� 1o and tþ io are the time interval in opera-tional data. The output is labeled as fault at time i� 1, if the faultwas recorded in between current time tþ i� 1o and the next timestamp tþ io, otherwise the output is normal.

Table 1 Analysis of “generator brush worn” fault based on the status data

Before fault Fault After fault

Status Code Description Status Code Description118 Emergency stop nacelle / hub63 Safety chain136 Limit switch 0 deg-rotorblade angle unclear

221 External power limitation125 Pitch overrun 90 deg 137 Limit switch 90 deg-rotorblade defective201 PLC fault 24V-power supply 164 Maintenance switch pitch62 PLC fault collection 288 Response signal interrupt battery driven101 Motor protection fuses 119 Timeout pitch controller208 No activity CAN-Bus CCU 285 Timeout CAN communication to hub1 Programstart PLC 120 Communication fault pitch controller87 Restart time delay 310 Fault reset141 Rotor CCU collective faults 145 Blade pitch control deviation142 Line CCU collective faults 251 Temp. sensor error (warning)296 Malfunction diverter 340 Pitch control deviation axis 2189 Manual operation yaw 341 Pitch control deviation axis 3

343 Blade angle not plausible axis 3344 Pitch malfunction 2 or 3 blades179 External controller deactivated102 Motor protection top cabinet

Fig. 4 Power curve of wind turbine during generator brushfault (Turbine 14): (a) During fault emergence; (b) one day afterthe fault; and (c) two days after the fault

Journal of Solar Energy Engineering MAY 2012, Vol. 134 / 021001-3

Downloaded 12 Mar 2012 to 128.255.30.115. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

3.4 Evaluation Metric. Selecting a prediction algorithm foractual testing was based upon its performance on the test data.The metrics of Eqs. (2) through (5) were used to measure perform-ance of the algorithms.

Accuracy ¼ TFPþ TNP

TFPþ TFMþ TNMþ TNP(2)

Senstivity ¼ TFP

TFPþ TFM(3)

Specificity ¼ TNP

TNMþ TNP(4)

The f-metric is also used due to class imbalance of the dataset [19].

f measure ¼ 2�recall�precision

recallþ precision(5)

In Eqs. (2)–(4), TFP is the total number of faults predicted, andTNP is the total normal cases predicted. TFM represents the totalfault cases misclassified as normal, whereas TNP is the total nor-mal cases misclassified as fault. The objective here is to maximizethe sensitivity, while keeping the desired level of accuracy. In Eq.(5), recall is equivalent to sensitivity, whereas precision is the ra-tio of the number of fault instances identified to the total numberof identified cases, both fault and normal.

3.5 Parameter Selection. To avoid the curse of dimensional-ity and improve the prediction accuracy, irrelevant parameterswere removed from the analysis. A combination of data-miningtechniques and well-established statistic measures were used toselect parameters. The SCADA system records more than 100 pa-rameters to monitor the wind turbine performance, including fixedvalues. Using domain knowledge, the initial 100 dimensional datawas reduced to 50 dimensions. A description of current dataset isprovided in Table 2.

Three known parameter selection approaches—namely chi-square statistic, boosting tree, and wrapper algorithm with geneticsearch—were employed to identify the best subset of parametersfor the prediction. Chi-square is a statistical approach to selecting

parameters, whereas boosting tree uses a gradient boostingmachine approach to rank the parameters. Wrapper methodologyfor selecting features uses the learning algorithm as a black box torank/score subsets of features according to their predictive power.In the present research, a genetic algorithm based rankingapproach was used. Table 3 describes the parameters selectedbased upon the three approaches, chi-square statistic, boostingtree, and wrapper methodology, respectively. Only the top 10 per-forming parameters were selected for each case. The role of theselected parameters in classifying faults and normal cases can beexplained by calculating simple averages. Table 4 lists the 10selected parameters and their comparison for both faults and nor-mal cases. The parameter selection approach has reduced 50 pa-rameters to 14.

3.6 Data Sampling. The class imbalance problem is preva-lent in wind turbines where the fault cases are rare compared withnormal cases (that is, when the turbine is fully operational). Con-sidering the huge amount of data, for a typical fault (e.g., rotorimbalance, blade angle asymmetry) the ratio of normal to faultinstances can be as large as 1000:1. Class imbalance is consideredone of the most critical problems in machine learning and data-mining based applications, and has gained attention from theresearchers worldwide [20–22]. Possible recommended solutionsto balance the data include: (1) over-sampling of the minorityclass; (2) undersampling of the majority class; and (3) cost sensi-tive classification. Irrespective of their widespread applications,these solutions have certain limitations. For example, over-sampling might create minority class data points very close to ma-jority data points thereby making the classification task more chal-lenging; undersampling might cause over fitting; and the costsensitive class classification requires cost information for

Table 2 Description of the dataset

No. Data type Description

1 Data frequency 10 min2 Fault description Generator brush worn3 Range of fault instances 53–3394 Range of normal instances 253–69565 Prediction time stamps 30 min, 60 min,…, 720 min6 Maximum prediction length 12 h7 Maximum number of input parameters 50

Table 3 Parameter selection based on different data-mining algorithms

Chi-square statistic Boosting tree algorithm Wrapper (genetic search)

Parameter Rank/Score Parameter Rank/Score Parameter Rank/Score

Temperature, shaft bearing 74.2 Nacelle revolution 1.00 Voltage phase A-C 1Nacelle revolution 69.9 Hydraulic pressure 0.98 Generator speed 2Torque (actual value) 55.4 Drive train acceleration 0.95 Temperature, generator2 3Drive train acceleration 53.5 Generator/gearbox speed 0.90 Temperature bearing A 4Temperature, shaft bearing 51.6 Generator speed 0.89 Wind speed 5Temperature, generator1 51.3 Temperature, shaft bearing 0.89 Temperature bearing B 6Voltage phase A-C 49.7 Temperature, gearbox bearing 0.86 Generator/gearbox speed 7Temperature, generator cooling air 44.8 Rotor speed 0.83 Operating state 8Wind speed 44.7 Temperature, bearing A 0.82 Wind deviation, 10 s 9Temperature control box axis 1 41.6 Voltage phase A-C 0.80 Temperature, shaft bearing 10

Table 4 The importance of selected parameters inclassification

Parameter Fault NormalDeviation

(%)

Nacelle revolution 1.4047 1.0253 37.00Drive train acceleration 37.6203 32.1777 16.91Temperature, shaft bearing 19.1223 16.8320 13.61Torque, actual value 36.0871 31.8459 13.32Generator speed 1072.9571 955.6942 12.27Wind speed 7.1663 6.4841 10.52Temperature, generator cooling air 58.2032 53.3175 9.16Temperature, generator 2 61.8917 57.1011 8.39Temperature, generator 1 62.3152 57.5166 8.34Temperature, bearing B 46.7021 43.5128 7.33Temperature, bearing A 30.6385 29.6817 3.22Voltage phase A 331.9429 332.5581 0.18Voltage phase C 331.3905 331.9967 0.18Voltage phase B 332.1161 332.7191 0.18

021001-4 / Vol. 134, MAY 2012 Transactions of the ASME

Downloaded 12 Mar 2012 to 128.255.30.115. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

misclassified cases. Another method of data sampling called datacleaning aims to remove redundant and noisy data points from themajority class. Common techniques include neighborhood prepro-cessing, Tomek links, and so on. In this paper, a combination ofTomek links and random forest algorithm based data samplingapproaches were used.

3.6.1 Tomek Links Based Data Sampling. In general, datafrom the majority class can be categorized into four types: noise,borderline, redundant, and safe. Removing noisy, borderline, andredundant data from the majority class can improve the predictionaccuracy by making the output class more distinguishable. In thispaper’s research, the Tomek links based data sampling method wasused. Tomek links use Euclidean distance information of input datapoints to identify borderline and noisy data. For example let usassume pi and pj to be two data points that belong to the minorityclass, and assume ni to be the data point representing the majorityclass. The distance between pi and pj is assumed to be d(pi,pj),whereas the distance between ni and pi is assumed to be d(pi,ni).The data point ni will be Tomek link to data point pi, ifd(pi,ni)< d(pi,pj). All identified majority class forming the Tomeklinks are removed iteratively. The process is repeated until theTomek links are identified. In this research, the logic illustrated inFig. 5 was used for data sampling. Figure 6 describes the applica-tion of the logic on one such dataset (e.g., dataset tþ 3). For theparticular dataset, the algorithm ran for 50 iterations and the datapoints from the majority class were removed. Comparing just theminimum distance interclass data points and minimum distance outof class data points made the process computationally efficient.

3.6.2 Random Forest Based Data Sampling. The applicationof Tomek links reduces the class imbalance only up to a certainlimit because it depends on a distance function defined for thedataset. To further reduce the class imbalance, a heuristic based

upon random forest algorithm was developed [23]. Random forestis an ensemble of unpruned classification or regression trees,trained from bootstrap samples of the training data [24]. It uses arandom subset of features in the tree induction process and classi-fies the output class based upon the majority of the decisionobtained through individual trees in a random forest. The mainemphasis is to reduce the negative instances (majority class) thatoverlap or close with positive instances (minority class). Theextent of overlap or closeness of instances is indicated by the Eu-clidean distance. Eliminating such negative instances will reducefalse negative classification, which can improve the predictionperformance. The input of the heuristic will be the cleaned datasetobtained after applying the Tomek links. The initial step of theheuristic is similar to the Tomek link based data sampling method;however, rather than the distance metric an actual classification isperformed (Fig. 7). The number of subsets for the majority classdataset depends upon the ratio of majority and minority classinstances. The training data for each subset are obtained by merg-ing all remaining majority class subsets and minority class data,and the algorithm is tested on individual majority class subsetdata that is not used in training. Because the test data containsonly one output class, a dummy output class, which is a positive,is added to the dataset to convert it into a two-class classification.Any misclassified instance potentially resembles the overlapbetween fault and normal class output and is therefore eliminated.The finally merged dataset is used to train the boosting tree algo-rithm, which is tested against the blind dataset. The process isrepeated for all time-stamped data in the dataset.

For this paper, the number of trees was set to 100. The classifica-tion was made through a majority vote, which took the decisions ofall trees into consideration. The output obtained using the afore-mentioned data sampling approaches was compared with the origi-nal data (Fig. 8). Figure 8(a) describes the graphical distribution ofboth fault and normal class of the 24 datasets that were extracted

Fig. 5 Identifying Tomek links

Journal of Solar Energy Engineering MAY 2012, Vol. 134 / 021001-5

Downloaded 12 Mar 2012 to 128.255.30.115. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

Fig. 6 Iterative sampling of dataset t 1 21 using Tomek links

Fig. 7 Process flow for random forest based data sampling approach

021001-6 / Vol. 134, MAY 2012 Transactions of the ASME

Downloaded 12 Mar 2012 to 128.255.30.115. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

from the original data. On average, more than 90% of the databelonged to the majority class, which indicates the dataset is highlyimbalanced. The initial imbalance in the output class was reduced;that is, the ratio of normal to fault instances became 80%:20%(Fig. 8(b)). Applying random forest based data sampling reducedthe class imbalance ratio to 65%:35% (Fig. 8(c)). For some of thetime-stamped dataset (e.g., tþ 36, tþ 57, tþ 66), the output classbecame completely balanced.

In Sec. 3.7, an approach to select a suitable data-mining algo-rithm is illustrated.

3.7 Algorithm Selection. Four well-known data-miningalgorithms, namely, multilayered perceptron (MLP), boostingtree, k-NN (k¼ 10), and support vector machine (SVM) were usedto build a prediction model at the tþ 21 time stamp. The best per-forming algorithm was used to perform predictions at 24 time

stamps. A description of the dataset for algorithm selection isshown in Tables 2 and 5. Table 6 describes the results obtainedusing various data-mining algorithms. Compared with other data-mining algorithms, boosting tree algorithms produced the best

Fig. 8 Distribution of fault and normal class: (a) original dataset; (b) dataset after Tomek linksbased sampling; and (c) dataset after Tomek links and random forest based sampling

Table 5 Dataset description for algorithm selection

Data setStart time

stampEnd time

stamp Description

tþ 21_overall 3/1/2010 12:00 AM 3/31/2010 11:20 PM 492observations

tþ 21_train 3/1/2010 12:00 AM 3/21/2010 09:30 PM 328observations

tþ 21_test 3/20/2010 09:60 PM 3/31/2010 11:20 PM 164observations

Journal of Solar Energy Engineering MAY 2012, Vol. 134 / 021001-7

Downloaded 12 Mar 2012 to 128.255.30.115. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

f_measure for fault class and overall best accuracy. It is thereforeconsidered best to build prediction models at all 24 time stamps.Poor sensitivity obtained using the remaining data-mining algo-rithms depicts their inability to classify the minority (fault) classinstances, whereas the boosting tree algorithm appears to beinsensitive to class imbalance by producing good sensitivity.Table 7 describes the specific details of data results obtained usingboosting tree algorithms at the tþ 21 time stamp. Results on thetest dataset indicated that out of 37 fault instances, the boostingtree algorithm was correctly able to predict 31 instances, whereasin the case of normal instances, 101 were correctly predicted outof 127. In Table 7, the column percentage represents the positive(false output), and negative (normal output) predictive values,whereas, the row values represent sensitivity and specificity.Figure 9 illustrates convergence of the boosted tree. The higherthe number of tree, the lower is the average multinomial deviance.Average multinomial deviance is a measure of goodness of fit

(estimated in terms of log likelihood of the model of interest andsaturated model). In the proposed research, the optimal number oftrees was found to be 110 with a maximum tree size being 3.

4 Discussion of the Computational Results

In this section, the results produced by the boosting tree algo-rithm are presented. The boosting tree algorithm is an ensemblelearning algorithm that combines many weak classifiers to producea power committee. In this approach, the output of a particularinstance is computed by averaging the output of several classifiers[25–27]. For the current analysis, the learning rate was set to 0.1.Two-thirds of the dataset was used to train the boosting tree,whereas the remaining one-third was used for testing. Table 8describes the results of the analysis on 24 different datasets. Theresults of three cases, (1) the original dataset; (2) the sampled data-set based upon Tomek links only; and (3) the sampled datasetusing Tomek links and random forest algorithm, are displayed inTable 8. Prediction accuracy on the cleaned dataset using bothTomek links and random forest algorithm was found in the range82.1%–97.1% for all time stamps, whereas accuracy in predictingfault cases was found in the range 59.2%–100%. The significantimprovement in accuracy indicates the effectiveness of data sam-pling methods.

Figure 10 compares the performance of all three cases in termsof the relative f_measure. The f_measure obtained using Tomeklinks and random forest based data sampling was always found tobe better than the other two cases. In addition, it is important tomention that the misclassification costs for both fault and normalclass output were kept the same; that is, equal to 1. The reasonableamount of accuracy for all 24 datasets indicates that the boostingtree algorithm is able to learn efficiently even in the case of classimbalance.

The results presented in this paper offer early prediction ofemerging faults. This allows operators to schedule maintenanceand minimize operations and maintenance cost. In addition, thepotential of collateral and severe faults is reduced.

Table 6 Prediction results at the t 1 21 time stamp

Accuracy Specificity Sensitivity f_measure

k-NN(k¼ 10) 80.8 82.8 61.5 0.372MLP 78.5 91.3 50.0 0.590SVM 78.7 78.7 0.00 0.000Boosting tree 80.4 79.5 83.7 0.659

Table 7 Results obtained from the boosting tree algorithm

ObservedPredictednormal

Predictedfault

Rowtotal

Number Normal 101 26 127Column percentage 94.39% 45.61%Row percentage 79.53%a 20.47%Total percentage 61.59% 15.85% 77.44%Number Fault 6 31 37Column percentage 5.61% 54.39%Row percentage 16.22% 83.78%b

Total percentage 3.66% 18.90% 22.56%Count All groups 107 57 164Total percent 65.24% 34.76%

aspecificitybsensitivity

Fig. 9 Boosted tree plot of prediction error function for datasett 1 21

Table 8 Comparison of the results obtained through differentdata sampling methods

Original Tomek Tomekþ random forest

Time stamp Acca Spea Sena Acca Spea Sena Acca Spea Sena

tþ 3 74.7 74.6 78.1 79.5 79.7 77.2 87.7 100 80.0tþ 6 76.3 76.7 69.7 76.7 77.0 74.7 94.2 100 88.8tþ 9 75.8 77.7 64.9 77.9 77.2 82.5 97.1 100 94.7tþ 12 73.9 74.6 66.6 74.5 76.4 65.3 83.4 98.5 82.8tþ 15 74.3 74.0 77.5 76.1 76.7 73.4 95.9 96.0 95.8tþ 18 75.0 75.5 70.2 69.1 69.0 69.7 94.3 95.2 93.7tþ 21 74.2 75.7 60.0 80.4 79.5 83.7 91.3 95.0 88.4tþ 24 75.0 78.3 46.6 77.2 79.0 70.5 80.0 92.8 71.4tþ 27 81.1 83.3 55.0 73.2 74.1 65.3 86.3 92.3 84.0tþ 30 72.5 73.9 63.3 70.8 74.7 58.0 87.5 91.6 86.2tþ 33 76.3 82.3 41.6 76.0 78.7 64.0 88.2 90.6 81.8tþ 36 83.1 95.7 28.1 72.2 69.8 76.6 86.2 89.4 84.6tþ 39 83.4 92.8 29.1 77.7 86.7 50.0 84.8 87.5 83.9tþ 42 88.7 97.7 37.5 76.8 76.0 79.1 83.1 86.6 83.3tþ 45 75.7 81.6 53.3 75.4 80.2 55.0 88.2 86.6 89.4tþ 48 83.8 90.7 40.0 82.2 89.4 70.8 87.4 84.4 89.0tþ 51 76.0 84.1 22.2 77.2 80.0 70.5 84.1 83.3 84.4tþ 54 67.9 69.3 60.8 74.0 76.9 66.6 82.2 83.3 81.8tþ 57 68.7 79.2 45.0 84.1 81.4 90.0 84.0 83.3 84.2tþ 60 74.7 79.2 54.5 74.1 80.9 56.2 84.2 81.2 86.3tþ 63 70.9 74.1 58.3 76.3 72.9 91.6 77.6 77.7 77.6tþ 66 63.0 64.5 53.3 55.0 54.1 57.1 77.0 69.5 78.4tþ 69 67.9 72.6 50.0 76.1 82.1 66.6 72.3 60.8 83.3tþ 72 60.6 61.6 53.8 64.0 64.5 63.1 82.1 59.2 92.9

aAcc¼Accuracy; Sen¼Sensitivity; Spe¼Specificity

021001-8 / Vol. 134, MAY 2012 Transactions of the ASME

Downloaded 12 Mar 2012 to 128.255.30.115. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

5 Conclusions and Future Research

This paper presented a methodology to predict the “generatorbrush worn” fault in wind turbines. Information from the SCADAoperational data and status data were used to generate datasets formaking predictions. Analyzing wind turbine status data indicatedthat other than poor brush design, the main contributing factorscausing worn generator brushes were drive train acceleration, gen-erator speed, torque, and aerodynamic asymmetry. The impor-tance of these parameters was validated by different parameterselection approaches. The computational results reported in thepaper demonstrated that in most cases, faults can be predictedwith a reasonable accuracy 12 h before they occur. Applying datasampling techniques improved the average prediction accuracy by12% compared to the original data. Further improvements in pre-diction accuracy can be obtained by finding an optimal number ofsubsets from the majority class data.

Due to the data sharing practice, only low frequency (10 min)data was available for the current research. Future work will bedirected toward testing the proposed model on high frequency data(e.g., 10 s or higher). Extending the prediction horizon to 24 h andbeyond is a topic for future research.

Acknowledgment

The research reported in the paper has been supported by fund-ing from the Iowa Energy Center, Grant No. 07-01.

References[1] Hatch, C., 2004, “Improved Wind Turbine Condition Monitoring Using Accel-

eration Enveloping,” Orbit, pp. 58–61.[2] Kusiak, A., and Verma, A., 2010, “The Future of Wind Turbine Diagnostics,”

Wind Syst. Mag., 2(8), pp. 66–71.[3] Ding, Y., Byon, E., Park, C., Tang, J., Lu, Y., and Wang, X., 2007, “Dynamic

Data-Driven Fault Diagnosis of Wind Turbine Systems,” Proceedings of the7th International Conference on Computational Science, Part I: ICCS, 4487,pp. 1197–1204.

[4] Ribrant, J., 2005, “Reliability Performance and Maintenance—A Survey ofFailures in Wind Power Systems,” Master Thesis, KTH School of Electrical En-gineering, Sweden.

[5] Kusiak, A., and Li, W., 2010, “Virtual Models for Prediction of Wind TurbineParameters,” IEEE Trans. Energy Convers., 25(1), pp. 245–252.

[6] Pan, F., Converse, T., Ahn, D., Salvetti, F., and Donato, G., 2009, “FeatureSelection for Ranking Using Boosted Trees,” CIKM’09, Nov. 2–6, Hong Kong,China, pp. 1–9.

[7] Kusiak, A., Zheng, H.-Y., and Song, Z., 2009, “Short-Term Prediction of WindFarm Power: A Data-Mining Approach,” IEEE Trans. Energy Convers., 24(1),pp. 125–136.

[8] Kusiak, A., Verma, A., 2011, “Prediction of Status Patterns of Wind Turbines:A Data-Mining Approach,” Transactions of the ASME: J. Sol. Energy Eng.,133(1), pp. 011008-1–011008-10.

[9] Tavner, P. J., Xiang, J., and Spinato, F., 2007, “Reliability Analysis for WindTurbines,” Wind Energy, 10(1), pp. 1–18.

[10] Rao, B. K. N, 1996, Handbook of Condition Monitoring, Oxford, Elsevier.[11] Kusiak, A., and Verma, A., 2010, “A Data-Driven Approach for Monitoring

Blade Pitch Faults in Wind Turbines,” IEEE Trans. Sustainable Energy, 2(1),pp. 87–96.

[12] Huyer, S. A., Simms, D., and Robinson, M. C., 1996, “Unsteady AerodynamicsAssociated With a Horizontal-Axis Wind Turbine,” AIAA J., 34(7), pp.1410–1419.

[13] Landberg, L, 1999, “Short-Term Prediction of the Power Production FromWind Farms,” J. Wind Eng. Ind. Aerodyn., 80(1), pp. 207–220.

[14] Caselitz, P., Giebhardt, J., and Mevenkamp, M., 1997, “Application of Condi-tion Monitoring Systems in Wind Energy Convertors,” Proceedings of theEWEC, Dublin, pp. 1–4.

[15] Rodriguez, L., Garcia, E., Morant, F., Correcher, A., and Quiles, E., 2008,“Application of Latent Nestling Method Using Colored Petri Nets for the FaultDiagnosis in the Wind Turbine Subsets,” Proceedings of 2008 IEEE Interna-tional Conference Emerging Technologies and Factory Automation, pp.767–773.

[16] Echavarria, E., Tomiyama, T., and Bussel, G., 2007, “Fault Diagnosis ApproachBased on a Model-Based Reasoner and a Functional Designer for a WindTurbine: An Approach Towards Self-Maintenance,” J. Phys. Conf. Ser., 75(1),pp. 1–9.

[17] Zaher, A., and McArthur, S., 2007, “A Multi-Agent Fault Detection System forWind Turbine Defect Recognition and Diagnosis,” Proceedings of 2007 IEEELausanne POWERTECH, pp. 22–27.

[18] Li, Q, Wang, Y., and Bryant, S. H, 2009, “A novel method for mining highlyimbalanced high-throughput screening data in PubChem,” Bioinformatics,25(24), pp. 3310–3316.

[19] Tarek, H., Ehab, F., and Magdy, M., 2009, “One Day Ahead Predictionof Wind Speed and Direction,” IEEE Trans. Energy Convers., 23(1), pp.191–201.

[20] Cleofas, L., Valdovinos, R. M., Garcia, V., and Alejo, R., 2009, “Use of Ensem-ble Based on GA for Imbalance Problem,” Lect. Notes Comput. Sci., 5552, pp.547–554.

[21] Hripcsak, G., and Wilcox, A., 2002, “Reference Standards, Judges, ComparisonSubjects: Roles for Experts in Evaluating System Performance,” J. Am. Med.Inform. Assoc., 9, pp. 1–15.

[22] Weiss, G. M., 2004, “Mining With Rarity: A Unifying Framework,” SIGKDDExplorations, 6, pp. 7–10.

[23] Barandela, R., Sanchez, J. S., Garcia, V., and Rangel, E., 2003, “Strategiesfor Learning in Class Imbalance Problems,” Pattern Recogn., 36, pp.849–851.

[24] Gu, J., 2007, “Random Forest Based Imbalanced Data Cleaning and Classi-fication,” PAKDD, pp. 1–7.

[25] Breiman, L., 2001, “Random Forest” Mach. Learn., 45, pp. 5–32.[26] Margineantu, D., 2000, “On Class Probability Estimates and Cost-Sensitive

Evaluation of Classifiers,” Workshop on Cost-Sensitive Learning at the 17thInternational Conference on Machine Learning.

[27] Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P, “SMOTE:Synthetic Minority Over-Sampling Technique,” J. Artif. Intell. Res., 16, pp.321–357.

Fig. 10 The relative contribution of f_measure corresponding to different data samplingtechniques

Journal of Solar Energy Engineering MAY 2012, Vol. 134 / 021001-9

Downloaded 12 Mar 2012 to 128.255.30.115. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm