Applied Soft Computing 13 (2013) 259270

Contents lists available at SciVerse ScienceDirect

Applied Soft Computing

j ourna l ho mepage: www.elsev ier .co

Wind t CAmodels

Meik Sch ichea Department o tr. 4, 2b Department o , Denmc Department o anada

a r t i c l

Article history:Received 9 December 2011Received in revised form 3 May 2012Accepted 8 August 2012Available online 23 August 2012

Keywords:ANFIS modelsCondition monWind turbineSCADA dataNormal behav

d turbference Systems (ANFIS). For this purpose: (1) ANFIS normal behavior models for common SupervisoryControl And Data Acquisition (SCADA) data are developed in order to detect abnormal behavior of thecaptured signals and indicate component malfunctions or faults using the prediction error. 33 differentstandard SCADA signals are used and described, for which 45 normal behavior models are developed.The performance of these models is evaluated in terms of the prediction error standard deviations to

1. Introdu

Conditiotance as thnowadays mcial. Unexpeexcessive dother speciaauxiliary eqdowntime dtors point oto monitor downtime a

CorresponE-mail add

1568-4946/$ http://dx.doi.oitoring

ior models

show the applicability of ANFIS models for monitoring wind turbine SCADA signals. The computationaltime needed for model training is compared to Neural Network (NN) models showing the strength ofANFIS in training speed. (2) For automation of fault diagnosis Fuzzy Interference Systems (FIS) are usedto analyze the prediction errors for fault patterns. The outputs are both the condition of the componentand a possible root cause for the anomaly. The output is generated by the aid of rules that capture theexisting expert knowledge linking observed prediction error patterns to specic faults. The work is basedon continuously measured wind turbine SCADA data from 18 turbines of the 2 MW class covering a periodof 30 months.

The system proposed in this paper shows a novelty approach with regard to the usage of ANFIS modelsin this context and the application of the proposed procedure to a wide range of SCADA signals. Theapplicability of the set up ANFIS models for anomaly detection is proved by the achieved performance ofthe models. In combination with the FIS the prediction errors can provide information about the conditionof the monitored components.

In this paper the condition monitoring system is described. Part two will entirely focus on applicationexamples and further efciency evaluation of the system.

2012 Elsevier B.V. All rights reserved.

ction

n monitoring of wind turbines is of increasing impor-e size and remote locations of wind turbines usedakes the technical availability of the turbine very cru-cted faults, especially of large components, can lead to

owntime offshore due to lack of suitable crane ships orlized vessels. However, also smaller issues and faults ofuipment like pumps or fans can cause expensive turbineue to restricted turbine accessibility. From an opera-f view it is therefore worth increasing the effort spentthe turbine condition in order to reduce unschedulednd thus operational costs.

ding author. Tel.: +49 40533268134.ress: m.schlechtingen@enbw.com (M. Schlechtingen).

The available CMS mostly require high level knowledge aboutthe system to be monitored. However, this knowledge is difcultto access and does often not exist. Physical models of the systemto monitor its condition and predict failures can thus seldom bebuilt with high accuracy due to its complex interaction among sev-eral dynamical subsystems. Moreover the available CMS mainlyfocus on vibrations. Vibration analysis is by far the most prevalentmethod for machine condition monitoring [2]. However, vibrationsensors are not installed on all turbines and components due totheir high costs. This causes a large number of turbines not beingcondition monitored at all or vibration sensors being installed atthe main components only.

On the other hand, there is a large amount of operational(SCADA) data available, which can be used to give an indication ofthe turbine condition. This fact is also stressed by Yang and Jiang [3]who additionally point out that these data are the cheapest resourcefor developing a CMS for wind turbines. The operational data can beeither turbine status information or measurements of signals such

see front matter 2012 Elsevier B.V. All rights reserved.rg/10.1016/j.asoc.2012.08.033urbine condition monitoring based on S. Part 1: System description

lechtingena,, Ilmar Ferreira Santosb, Soane Achf Technical Operation Wind Offshore, EnBW Erneuerbare Energien GmbH, Admiralittsf Mechanical Engineering, Section of Solid Mechanics, Technical University of Denmarkf Mechanical Engineering, Machine Design Section, cole Polytechnique de Montral, C

e i n f o a b s t r a c t

This paper proposes a system for winm/locate /asoc

DA data using normal behavior

c

0459 Hamburg, Germanyark

ine condition monitoring using Adaptive Neuro-Fuzzy Inter-

260 M. Schlechtingen et al. / Applied Soft Computing 13 (2013) 259270

as temperatures, currents or pressures. Using turbine status infor-mation Kusiak and Wenyan [4] and Kusiak and Verma [5] show thatfaults can be predicted 560 min in advance. In order to performpredictive maintenance this prediction period is too short, since itdoes not leaactions. By trends of recant change[6]. Using Net al. [7], Zahthat changesome casessuited to allbefore the chistorical opels capable more input suited, sincsignals meaoutput.

One advwind turbinbehavior is behavior mdecoupled [1] and Sanoped in perhealthy (noof the compto predict acation of ch

This appbility of a sywith eXogewind turbinever, this aselection toof signals anbe avoided.algorithms systems usi(MAS) [1,11behavior met al. [13], wdition of a good perforresearch acgen [10] prodrive train

Instead oa novelty, nonlinear stuning the phase. Jangparametersfaster traindata procesthermore, aincorporatetheir outpu

The moncan furtherfor modelinZaher and Mthe normal gen and San

ReConstruction (FSRC) approach, since this approach additionallyallows for monitoring the signal value magnitude. The differentapproaches mainly concern the question of the used input signalsto model the target signals. This approach is also followed by the

h prnterp

exptaintert ied gecessfrbinees.

resegatioADAes me com

devrateg

mal bd in predurredrted

idens whs and

SCAiffereuoused.ectioped A datf dess forode

ses os and

predion 5eniw th

ed anorkinents are

9.

crip

CMSs ind tule orime. e comCMS he fois brve the operator with enough time to take maintenanceapplying advanced signal analysis methods focused onpresentative signals or combination of signals, signi-s in turbine behavior can be detected at an early stageeural Network (NN) model based approaches Sanz-Bobier et al. [1] as well as Schlechtingen and Santos [8] shows in signal behavior can be detected, days, weeks and in

months in advance. These methods are therefore betterow operators to take measures to improve the conditionomponent nally fails. In the model based approacheserational data is used to develop normal behavior mod-

of predicting a certain output signal, when given one orsignals. For wind turbine signals these attempts are welle many signals can be found to be correlated to othersured simultaneously, e.g. the wind speed or the power

antage of using normal behavior models to monitore signals is that no prior knowledge about the signal

needed. Another important property is that with normalodels the possibility of monitoring the signal is widelyfrom the operational mode as reported by Zaher et al.z-Bobi et al. [7]. The normal behavior models are devel-iods where the turbine components can be consideredrmally operating), usually the period at the beginningonent lifetime. Afterwards, the trained model is used

specic signal where the prediction error gives an indi-anges in signal behavior and thus incipient faults.roach is of large interest in research. In [9] the applica-stem identication approach using an AutoRegressivenous input (ARX) model to monitor the condition of ae generator bearing using SCADA data is shown. How-pproach requires human intervention for parameter

nd a good performing model. Due to the large amountd turbines to be monitored, human intervention should

Most activities take advantage of articial intelligence(learning capabilities) and among the most advancedng this approach is SIMAP [7] and a Multi Agent System]. Both systems use articial NN to set up the normalodels of SCADA data. This line is also followed by Xianghere a NN model based method to monitor the con-

wind turbine generator bearing is developed and themance of NN in this context shown. Additionally, earliertivities by Schlechtingen and Santos [8] and Schlechtin-posed a condition monitoring system for wind turbine

components using NNs.f using NNs, ANFIS is used in this paper, which presentsin this eld of application. ANFIS models can learnignal relations by setting up a set of fuzzy rules andMembership Function (MF) parameters in a training

[14] shows that in comparison to NN models, fewer must be trained in ANFIS models, leading generally toing. The major drawbacks of NNs are their black-boxsing structure and slow convergence speed [15]. Fur-

priori knowledge about the system is difcult to bed. Here ANFIS models have a major advantage, due tot being based on linguistic rules and tuneable MFs.itoring systems developed by the different researchersmore be classied according to the input signals usedg. While Sanz-Bobi et al. [7], Zaher et al. [1] as well ascArthur [11] use autoregressive approaches to set up

behavior models, the research presented by Schlechtin-tos [8] showed that it is advantageous using a Full Signal

researcTo i

a fuzzythe ceran expsurveybe sucone tuturbin

Theinvestibine SCadvanc[11] ar

Thestep st

1. Noropethe

2. Occrepo

3. Thebasetern

Theteen dcontingather

In Sdevelo(SCADA briemodelused m4 focumodelfor thein Sectand a dies hodetect(FIS) wcompoResultSection

2. Des

Thepatterning winschedudowntthat arof the

In tFig. 1) esented in this paper.ret the prediction error of the normal behavior models,ert system that outputs a diagnosis, the condition andy of statement based on rules that were established bys employed by Sanz-Bobi et al. [7]. With regard to thearbox faults in their research, this approach proved toul, since the rules established with fault patterns from

were also applicable to predict the same faults on other

arch presented in this paper is the result of two yearsns carried out to develop a CMS that uses wind tur-

data available to wind turbine operators. Thereby theade by Sanz-Bobi et al. [7] as well as Zaher and McArthurbined and further developed.

elopment of the aforementioned CMS follows the threey below:

ehavior models of the relevant SCADA data were devel-order to monitor and detect anomalies by consideringiction error.

anomalies within the prediction error were related to faults.tied relations were implemented in knowledge dataich are used by FIS to automatically analyze these pat-

output a diagnosis.

DA data used in this research is obtained from eigh-nt operating onshore turbines of the 2MW class, where

operational data from April 2009 to March 2012 were

n 2 of this article the general concept of the systemis described giving details about how the informationa) is processed to nally output condition statements.cription of the advantages of ANFIS models over NN

the given application and a short description of thel setup and structure is supplied in Section 3. Sectionn the performance of the developed normal behavior

describes the identied inputoutput sets. An exampleiction error analysis and the model interactions is given. In Section 6 the concept of anomaly detection is showntion of abnormal signal behavior is given. Section 7 clar-e prediction error together with the information aboutomalies is processed by the Fuzzy Interference Systemng as fuzzy expert. The output is a diagnosis about the

condition and a potential root cause of the anomaly. discussed in Section 8 and conclusions are drawn in

tion of the general CMS concept

developed in this research aims at detecting trends and SCADA data in order to predict possible failures, giv-rbine operators enough time to adapt the maintenance

take further measures to prevent unexpected systemFor this purpose 10 min averaged SCADA data are usedmonly available to operators. The general architecture

developed is shown in Fig. 1.llowing, the function of the different CMS modules (seeiey described.

M. Schlechtingen et al. / Applied Soft Computing 13 (2013) 259270 261

2.1. Trainin

Within trained if a The latter ischange as apreprocesseincludes: (1processing

In the trato allow earone month formed wheoutput of thstandard thwind turbin

2.2. Predict

The predprocessed snormal-behlated and st

2.3. Anoma

In this mtied. This ithresholds output is anquency andanomaly in

2.4. Fuzzy e

Here thecomponentnumber of ito be monit

ile thiagn

osisitionntial

zzy e

hin the plies.

the

del sFig. 1. CMS overview.

g module

the training module the normal behavior model ismodel is not yet available or new training is required.

true if a component is replaced and the signal relations consequence. Before training the model, the data ared according to the methodology proposed in [8] which) a validity check, (2) data range check, (3) missing dataand (4) lag removal.ining module different training levels are implementedly monitoring. A rst model training is performed afterof operational data collection. Further trainings are per-n three, six and nine months of data are available. Thee training module is the trained ANFIS model and theresholds marking the normal operational range of thees using the prediction error.

ion module

Whto be d

diagn cond pote

2.5. Fu

Witgiven anomato givedition.

3. Moiction module is active once a trained model of theignal is available in the model base. The developedavior model is applied and the prediction error calcu-ored.

ly detection module

odule the anomalies in the prediction errors are iden-s done on the basis of the determined normal-behaviorby the training module or expert dened values. The

anomaly matrix containing information about the fre- date of occurrence, as well as the duration of the current

days.

xpert initialization module

FIS structures used for diagnosis of the anomalies and condition statements are initialized with regard to thenputs and outputs as well as their MFs. Each componentored has its own FIS structure.

3.1. Data se

The avaicontain moters, calculset points, tvoltages, etmaximum, ing period ifor mean vastored in thtions e.g. suthis researcthe 10 min pvariations; tative signatime series,

The remTable 1 andare illustrat

In additnon-operate inputs sorely depend on the component or subsystemosed, each FIS structure has the following outputs:

(information about the abnormal behaving signal) (classication in green, yellow and red color code)root cause

xpert application module

this module the initialized FIS structure is evaluated,rediction errors and the information about presentThe output is stored in text format and is visualizedanalyst a comprehensive summery of the turbine con-

etup and structuret description

lable SCADA data sets from the operating wind turbinesre than 150 different signals, ranging from hour coun-ated values, digital indicators of switch positions ando continuous measurements of temperatures, currents,c. For some of the continuous measurements the mean,minimum and standard deviation of the 10 min averag-s available. In this research only normal behavior modelslues were considered for three reasons: (1) The peakse min.max. values can be caused by transient situa-dden changes in wind speed that are not in the scope ofh; (2) stochastic effects in the signals are averaged out ineriod, making the prediction less sensitive to stochastic

(3) modeling the standard deviations requires represen-ls, i.e. other standard deviations or higher resolution

that are not accessible.aining signals considered for the CMS are given in

a schematic of the turbine and of the sensor locationed in Fig. 2.ion to these signals, information about service andional periods is extracted from the corresponding hour

262 M. Schlechtingen et al. / Applied Soft Computing 13 (2013) 259270

Table 1SCADA data signal list used for normal behavior model development.

Name of variable Unit Sensor location (Fig. 2) Short description Normal behavior modeled

Spinner temp. C 1 Spinner temperature (in hub housing) YesHub controller temp. C 1 Pitch controller temperature YesPitch angle 1 Blade pitch angle (avg. over all 3 blades) YesHydraulic oil temp. C 10 Hydraulic oil temperature (used for pitching blades) YesRotor speed rpm 2 Rotor speed (low speed shaft) YesGear bearing temp. (HSS) C 3 High speed shaft bearing temperature YesGear oil temp C 3 Gearbox oil temperature YesGenerator speed rpm 4 Generator speed (high speed shaft) YesGenerator bearing temp.1 C 5 Generator bearing temperature gearbox end YesGenerator bearing temp. 2 C 5 Generator bearing temperature transformer end YesGenerator slip ring temp. C 5 Generator slip ring temperature YesGenerator ph.1 temp. C 5 Generator stator temperature phase 1 YesGenerator ph.2 temp. C 5 Generator stator temperature phase 2 YesGenerator ph.3 temp. C 5 Generator stator temperature phase 3 YesGenerator current ph.1 A 6 Generator current phase 1 YesGenerator current ph.2 A 6 Generator current phase 2 YesGenerator current ph.3 A 6 Generator current phase 3 YesPower output kW 6 Turbine power output YesReactive power kVAr 6 Turbine reactive power consumption YesGrid inverter ph.1 temp. C 6 Inverter temperature grid end YesGrid rotor in erature phase 1 generator end YesGrid rotor in eratuGrid rotor in eratuConverter co ling wConverter ch ke coConverter co trolleTop controll oller tGrid busbar ratureHV transform transfHV transform transfHV transform transfNacelle tem in nacWind speedWind direct n Ambient tem eratu

counters. Taccording t

3.2. Model

Several rels in the of NN are massive paduce instanincompletewide rangeshown thatinputoutpHowever, ittapped in lo

e nus peverter ph.1 temp. C 6 Inverter tempverter ph.2 temp. C 6 Inverter tempverter ph.3 temp. C 6 Inverter tempoling water temp. C 6 Converter coooke coil temp. C 6 Converter chontroller temp. C 6 Converter coner temp. C 6 Turbine contrtemp. C 8 Busbar tempeer ph.1 temp. C 8 High voltage er ph.2 temp. C 8 High voltage er ph.3 temp. C 8 High voltage

p. C 7 Temperature m/s 9 Wind speed ion 9 Wind directiop. C 9 Outdoor temp

his information is used for preprocessing the datao the methodology presented in [8] (Fig. 3).

over thneurontype

esearchers applied NN to set up normal behavior mod-eld of condition monitoring. Some of the key featurestheir high processing speeds which are due to theirrallelism, their proven ability to be trained, to pro-taneous and correct responses from noisy or partially

data, and their ability to generalize information over a [16]. In earlier publications of the authors [8,10] it was

NN are indeed capable of accurately nding an existingut mapping for different wind turbine SCADA signals.

was found that NN have a high likelihood of becomingcal minima. In Fig. 4 three examples of NN performances

427

9

10

8

6

53

1

Fig. 2. Wind turbine schematic: sensor positions.

The variof local migenerally uperformingruns with ashortfall. Fisen. This leacceptable

Fuzzy syuations invwell underfast, solutiois that exist

Fig. 3. Wind turbines.re phase 2 generator end Yesre phase 3 generator end Yesater temperature Yes

il temperature Yesr temperature Yesemperature Yes

Yesormer temperature phase 1 Yesormer temperature phase 2 Yesormer temperature phase 3 Yeselle (housing) of the turbine Yes

YesNo

re No

mber of runs are shown. For this example two hiddenr input signal are used.

ations in performances are up to 20%, stressing the risknima. Furthermore the number of hidden neurons isnknown. Tarrassenko [17] and Raq et al. [18] suggest

several runs with random weight initializations and varying number of hidden neurons to overcome thisnally the network that performs best should be cho-ads to a large number of different trials to obtain ansolution.stems are very useful in two general contexts: (1) in sit-olving highly complex systems whose behaviors are notstood and (2) in situations where an approximate, butn is desired [19]. A further advantage of fuzzy systemsing expert knowledge can be implemented to improve

nacelleHV

transformer

converter

meteorology

auxiliaryequipment

generatorgearbox

rotor system

turbine schematic: components/subsystems in the considered wind

M. Schlechtingen et al. / Applied Soft Computing 13 (2013) 259270 263

0.7

0.75

0.8

0.85

0.9

0.95

1

No

rma

lize

d s

td(e

rro

r)

Fig. 4.

the approxifunctions an

Considering capabilisystem andnetworks hreal-world benets of This gives thknowledge

In ANFISthe FIS is usparameters

In the fomodels are

3.2.1. FIS stThere ar

dani [22] antypes is in tion (Sugenwork the Sin the premfuzzy equatfuzzy ifthetem under cto be more ity of the ouprediction.

3.2.2. MemMember

whose shapIn most fuzused (triangtribution mand linear Mtribution fuwith regardFurthermorin the inpushape.

Normal FSRC model

G

Wind speed (t-lag)

Wind direction (t-lag)

A

Generator current ph.1

ompa

Num conv

famr of h thmberter pn co

Train learin [1uarese thgene4].

Input numThe cbinated antic alAPLSool ox supinpumaypoinenomatic i2 4 6 8 10

Run

Pitch angle

Grid busbar temp.

Conv. choke coil temp.

Exemplary performances of trained NNs over different runs.

mation by tuning, removing or adding of membershipd rules.able work has been carried out to integrate the learn-ty of NN with FIS for deriving the initial rules of a fuzzy

tuning the membership functions [20]. Fuzzy neuralave shown to be very advantageous in dealing withproblems [21]. These neuro-fuzzy systems combine thethese two powerful paradigms into a single capsule.e ability to accommodate both data and existing expertabout the problem under consideration [20].

the advantages of NN are combined with FIS. Therebyed to set up a set of rules whose membership functions

are tuned in a training phase.llowing sections the setup and structure of the ANFISdescribed.

ructuree two common types of fuzzy interference: the Mam-d Sugeno [23]. The main difference between the twothe consequent part, which can be a nonfuzzy equa-o) instead of a fuzzy linguistic value (Mamdani). In thisugeno type is used as it has fuzzy sets only involvedise part. However, the consequent part can be a non-ion. Due to the qualiers on the premise parts, eachn rule can be viewed as a local description of the sys-

Fig. 5. Cmodel.

3.2.3. In a

who isnumberesearcthe nuthis, afbetwee

3.2.4. The

posed least sqdecreabut in time [1

3.2.5. The

tored. a commodela genesion (Glarge ptoolboto the which toring the phautomonsideration [14]. Moreover the Sugeno type is knowncomputationally efcient and has guaranteed continu-tput surface, which makes it well suited for time series

bership functionsship functions can be represented by an arbitrary curvee is dened as a function that suits the application.

zy applications straight-line membership functions areular/trapezoidal). In this paper generalized normal dis-embership functions are employed for the input spaceFs for the output space. The generalized normal dis-

nctions have the advantage of giving a broad exibility to the function shape, depending on their parameters.e these functions assure smoothness of the transitionst space. The free parameters are: location, scale and

for which tinput to maccurate bulute currenthe phases

Hence, aunderstandis the prefe

Since bonal readingimportant, (Full Signalmodeled bydifferent tycorrelated sFig. 5 for th

With thinstance difCross prediction FSRC model

ANFISGenerator current ph.1enerator current ph.2

mbient temp. (t-lag)

ANFIS

rison between a normal FSRC model and the cross prediction FSRC

ber of membership functions/rulesentional FIS, the number of rules is decided by an expertiliar with the system to be modeled [14]. Generally, theMFs can be dened for each input separately. In thise expert for this particular task is not available, and

of membership functions per input is xed to two, asreliminary tests, was found to be a good compromisemputational efciency and model performance.

ing methodning algorithm used is a hybrid learning rule as pro-4], consisting of a combination of gradient decent ands estimation. Not only can this hybrid learning approache dimension of the search space in the gradient method,ral, it will also cut down substantially the convergence

signalsber of input signals differs for each signal to be moni-hoice of relevant input signals is not trivial and requiresion of physical understanding of the system to bed advanced data reduction techniques. In this researchgorithm combined with a partial least squares regres-) [24] is used to detect potential input signals from thef SCADA data. For this purpose the genetic algorithmplied by Leardi [25] is applied. However, GAPLS pointst signals containing most of the target signal features,

not be the best ones to choose from a condition moni-t of view if one considers the physical understanding ofenon at hand. An example for possible difculties withnput signal selections is the generator phase currentshe GAPLS suggests using the generator current ph.1 asodel generator current ph.2. A model like this is veryt by using this inputoutput set, monitoring of the abso-t level is impossible. Instead only differences betweencan be detected. combination of data reduction techniques and theing of the physical process to select suitable input setsrred strategy used in this paper.th, the information about relative changes between sig-s of the same type as well as their absolute level istwo types of models are developed. The rst is a FSRC

ReConstruction) approach, where the target signal is fully reconstructing it through correlated signals of

pes and the second is by fully reconstructing it throughignals of the same type. The difference is illustrated ine generator phase current.e cross prediction FSRC models, asymmetries (forferences in generator phase currents or temperatures of

264 M. Schlechtingen et al. / Applied Soft Computing 13 (2013) 259270

Nacelle temp .Wind speed

Ambient temp .Wind dire ction

Nacelle temp .Wind spee d

Input Outpu t

loped

the three phmal FSRC mchanges in

Howeveing input scondition mbility, as beierror. The rchoice is diexpert knowrequiremenlow predicttion error ogenetic algonent or subfor the 45 m

4. Model p

A good choice of inways the peof the modperformancis more of ingiving an inphysical untion errors centered abSpinner temp .Hub control ler temp .

Pitch angleHydraul ic oil temp .

Rotor s peed

Gear oil temp .Genera tor s peed

Generator ph .1 temp .Generator ph .2 temp .Generator ph .3 temp .

Generator cur rent ph.1Generator cur rent ph.2Generator cur rent ph.3

Power ou tput

Grid rotor in verter ph .1 temp .Grid rotor in verter ph .2 temp .Grid rotor in verter ph .3 temp .

Conver ter coo ling wa ter temp .Con ver ter c hoke coil temp .Conver ter con trol ler temp .

Top con trol ler temp .Grid bu sbar temp .

HV tran sformer ph .1 temp .HV tran sformer ph .2 temp .HV tran sformer ph .3 temp .

Fig. 6. Inputoutput sets of the deveases) can be identied very accurately whereas the nor-odel is used to monitor the absolute level and relativecomparison to other signals.r, the model accuracy is not the only criterion for choos-ignals when developing normal behavior models foronitoring purposes. A further requisite is the fault visi-ng the visibility of the developing fault in the predictionelation between the fault visibility and the input signalfcult to dene and not exactly known, which is whyledge is required to estimate this inuence. A general

t of normal behavior models is that they should have aion error in case of normal behavior and a high predic-therwise. For the given problem the combination of therithm with engineering knowledge about the compo-

system lead to the input-output sets illustrated in Fig. 6odels developed.

erformance

model performance is the direct result of the correctput signals and successful training. There are severalrformance can be measured. Due to the different natureeled signals (power output, currents, temperatures) ae comparison between them is not meaningful. Whatterest is the prediction error variation around its mean,dication of how good the prediction is, by keeping theit. Here the standard deviation is used. Since the predic-after training are close to being normal distributed andout zero this choice is acceptable (compare Fig. 9). The

values are ctraining.

E = xm xp

P = (E)

E is the predicted valudeviation.

In ordermal operatinstance a operationaloperational

4.1. Compa

To stresperformancfor ANFIS aanalog to thonly ve ruThe numbeand output

The train(29,513 10

The comstandard derequired fois due to thSpinner temp .Hub contro ller temp.Pitch ang leHydraul ic oil temp.Rotor spee dGear bear ing temp. (HSS)Gear oil temp .Generator spee dGenerator bearing temp .Generator bearing temp. 2Generator slip ring temp .Generator ph.1 temp.Generator ph.2 temp.Generator ph.3 temp.Generator ph.1 temp. cro ssGenerator ph.2 temp. cro ssGenerator ph.3 temp. cro ssGenerator current ph .1Generator current ph .2Generator current ph .3Generator current ph .1 cro ss.Generator current ph .2 cro ss.Generator current ph .3 cro ss.Power ou tputReactive powe rGrid inver ter ph .1 temp.Grid ro tor inver ter ph .1 temp.Grid ro tor inver ter ph .2 temp.Grid ro tor inver ter ph .3 temp.Grid ro tor inver ter ph .1 temp. cros sGrid ro tor inver ter ph .2 temp. cros sGrid ro tor inver ter ph .3 temp. cros sConver ter coo ling wa ter temp.Conver ter cho ke coil temp.Conver ter con trol ler temp .Top con tro ller temp.Grid bu sbar temp.HV tran sformer ph .1 temp.HV tran sformer ph .2 temp.HV tran sformer ph .3 temp.HV tran sformer ph .1 temp. cro ssHV tran sformer ph .2 temp. cro ssHV tran sformer ph .3 temp. cro ss

models.alculated based on test data sets that were not used for

(1)

(2)

diction errors; xm is the measured values; xp is the pre-es; P is the performance measure; is the standard

to assess the performance information about the nor-ional range of the signals is essential (see Table 2). Formodel performance of 1 C is good when the normal

range is 100 to 100 C, but rather poor if the normal range is 2530 C.

rison between NN and ANFIS models

s the advances of ANFIS models the results of a briefe and training speed comparison is shown in Table 3nd NN training. The chosen NN training procedure ise one described in [8], but with the difference that herens with random weight initializations are performed.r of hidden neurons and the number of MFs, per input

signal is set to two (Table 4).ing is performed with nine month of operational data

min average values).parison shows that the performance in terms of theviation of both approaches is similar. However, the timer training is generally smaller with ANFIS models, whiche necessary trial and error procedure when training NN.

M. Schlechtingen et al. / Applied Soft Computing 13 (2013) 259270 265

Table 2Normal operational ranges of the SCADA data.

Signal Normal range Signal Normal range

Spinner temp. 10 to 45 C Power output 50 to 2100 kWHub controller temp. 5 to 50 C Reactive power 25 to 300 kWArPitch angle (avg. over all blades) 5 to 95 Grid inverter ph.1 temp. 25 to 50 CHydraulic oil temp. 10 to 65 C Grid rotor inverter ph.1 temp. 25 to 55 CRotor speed 0 to 16 rpm Grid rotor inverter ph.2 temp. 25 to 55 CGear bearing temp. (HSS) 30 to 80 C Grid rotor inverter ph.3 temp. 25 to 55 CGear oil temp. 30 to 65 C Converter cooling water temp. 5 to 50 CGenerator speed 0 to 1700 rpm Converter choke coil temp. 5 to 120 CGenerator bearing temp. 1 10 to 85 C Converter controller temp. 5 to 60 CGenerator bearing temp. 2 10 to 85 C Top controller temp. 10 to 50 CGenerator slip ring temp. 5 to 45 C Grid busbar temp. 5 to 60 CGenerator ph.1 temp. 10 to 140 C HV transformer ph.1 temp. 20 to 95 CGenerator ph.2 temp. 10 to 140 C HV transformer ph.2 temp. 20 to 95 CGenerator ph.3 temp. 10 to 140 C HV transformer ph.3 temp. 20 to 95 CGenerator current ph.1 0 to 1700 A Nacelle temp. 5 to 50 CGenerator current ph.2 0 to 1700 A Wind speed 0 to 35 m/sGenerator current ph.3 0 to 1700 A

Jang [14] showed that in ANFIS models fewer parameters must betrained, leading generally to faster training.

The performance of ANFIS models does not vary among differ-ent runs, which is due to the initialization of the MFs, involvinga grid partithat with Nter performnding an other showstraining spe

NN are gput is difcabout the shave a majoterms of rulby using thegrid portionnMF being number of MFs is equamum of 16 to very fastthe model cthe numbeexponentia

Table 3Performance a

SCADA signa

Hydraulic oiGenerator bePower outpuTop controll

Table 4Performance a

SCADA signa

Hydraulic oiGenerator bePower outpuTop controll

4.2. Performance of the set up ANFIS models

Schlechtingen and Santos [8] showed that averaging the predic-tion error decreases the variance of the system and consequentially

es the sensitivity against anomalies. In fact it was shown averaging, upcoming faults and anomalies can be detected

and with a higher certainty (fault patterns are stronger pro-d). Generating average values can be advantageous, becauseveraged signals of several turbines can easily be compared

a single plot and turbine operators are most often interestedrmat

actiouired

erro indied pr

(E)

aver (E)

0

0

0

std

(err

or)

[C

]

7

r) [

]tion algorithm [14,26]. What is important to notice isN and the suggested number of different trials, a bet-ance can be achieved (see Fig. 7) potentially. However,ptimal solution is computational expensive. Fig. 7 fur-

that with ANFIS models a good compromise betweened and model accuracy can be found.ood in identifying non-linear relations, but their out-

ult to back-track (black box model). A priori knowledgeystem is difcult to in cooperate. Here ANFIS modelsr advantage. Not only can a priori knowledge be used ines or shape of MFs, the output can also be back-tracked

information about the rule that red. When using theing method, the number of rules is equal to: nMFni withthe number of MFs and ni the number of inputs. Theinputs is smaller or equal to four and the number ofl to two for the models developed, leading to a maxi-rules. Especially for small numbers of inputs this leads

to train and simple to back-track models. However,omplexity (number of rules) scales exponentially withr of inputs, which causes training speed also increaselly.

nd speed of ANFIS training.

l ANFIS

increasthat byearliernouncedaily awithinin infoessaryare reqdictionfor 144averag

Pavg =

E is theerrors;Elapsed time for training P

l temp. 125.2 s 2.00 Caring temp. 37.3 s 1.30 Ct 93.6 s 45.78 kWer temp. 14.1 s 1.24 C

nd speed of NN training.

l NN

Elapsed time for training P

l temp. 452.3 s 2.06 Caring temp. 373.2 s 1.33 Ct 284.9 s 45.99 kWer temp. 369.3 s 1.57 C

6

std

(err

o

1

1.

1.

1.

1.

std

(err

or)

[C

]

Fig. 7. Performion that can support their day-to-day planning of nec-ns, for which not too much details (such as hourly data)

[27]. In this paper all analysis is based on averaged pre-rs. The averaging period used is one day, i.e. the averagevidual prediction errors is calculated, giving one singleediction error per day.

(3)

aged prediction errors; Pavg is the performance measure is the standard deviation of E.

.7

.8

.9 Grid busba r temp.

.5

8

Pitch angle6

.5

7

1 2 3 4 5 6 7 8 9 10.2

21

22

23

24

Run

Nacelle temp.

NN ANFIS

ance of different NN and ANFIS trainings over the number of runs.

266 M. Schlechtingen et al. / Applied Soft Computing 13 (2013) 259270

Table 5Example model performance in terms of standard deviation of the prediction error after training.

Model P Pavg Unit Model P Pavg Unit

Spinner temp. 1.55 0.63 C Power output 47.00 16.34 kWHub controller temp. 0.55 0.18 C Reactive power 0.26 0.03 kWArPitch angle (avg. over all blades) 0.44 0.08 Grid inverter ph.1 temp. 0.45 0.13 CHydraulic oil temp. 2.59 1.35 C Grid rotor inverter ph.1 temp. 0.42 0.14 CRotor speed 0.21 0.09 rpm Grid rotor inverter ph.2 temp. 0.42 0.15 CGear bearing temp. (HSS) 0.79 0.35 C Grid rotor inverter ph.3 temp. 0.41 0.13 CGear oil temp. 2.09 1.02 C Grid rotor inverter ph.1 temp. cross. 0.39 0.07 CGenerator speed 23.46 9.32 rpm Grid rotor inverter ph.2 temp. cross. 0.48 0.13 CGenerator bearing temp. 1 1.44 0.73 C Grid rotor inverter ph.3 temp. cross. 0.37 0.08 CGenerator bearing temp. 2 1.86 0.75 C Converter cooling water temp. 0.69 0.33 CGenerator slip ring temp. 1.30 0.46 C Converter choke coil temp. 2.95 1.26 CGenerator ph.1 temp. 4.98 1.90 C Converter controller temp. 0.54 0.22 CGenerator ph.2 temp. 4.92 1.87 C Top controller temp. 1.15 0.48 CGenerator ph.3 temp. 4.92 1.86 C Grid busbar temp. 0.57 0.19 CGenerator ph.1 temp. cross. 0.51 0.11 C HV transformer ph.1 temp. 3.59 1.64 CGenerator ph.2 temp. cross. 0.51 0.09 C HV transformer ph.2 temp. 2.88 1.32 CGenerator ph.3 temp. cross. 0.66 0.15 C HV transformer ph.3 temp. 3.25 1.47 CGenerator current ph.1 39.07 13.62 A HV transformer ph.1 temp. cross. 1.59 0.87 CGenerator current ph.2 38.63 13.29 A HV transformer ph.2 temp. cross. 1.39 0.80 CGenerator current ph.3 39.05 13.57 A HV transformer ph.3 temp. cross. 1.35 0.91 CGenerator current ph.1 cross. 2.32 1.45 A Nacelle temp. 1.21 0.60 CGenerator cu ed Generator cu

The perfbine Genereach model

The stanwhen the p

In someachieve a gis not foundtaneously mset up. Althit should beapproach iset al. [1] an

5. Predicti

The interequires sothiness of input signamal. Henceinto accounmodels. Exaoutput or tthese signa

tions the n

visumpllableedicton:

) =

the

colo are sredicolortion etion s lab

8 shoat thrrent ph.2 cross. 2.37 1.72 A Wind sperrent ph.3 cross. 1.60 0.99 A

ormances are listed in Table 5 for a random Wind Tur-ator (WTG) of the eet (the individual performance of

differs from turbine to turbine).dard deviations decrease almost by a factor of three,redictions error is averaged.

cases it may be difcult to identify suitable inputs toood model performance. This is for instance if a signal

to be well correlated to any of the other signals simul-easured and thus an accurate FSRC model cannot be

ough this was not the case for the models developed, mentioned that for those signals an alternative model

to build autoregressive models as proposed by Zaherd Sanz-Bobi et al. [7].

on errors

rpretation of the prediction errors for fault diagnosisme information about the validity or the trustwor-the input signals to the models. Abnormal behavingls, consequentially cause the prediction to be abnor-

it is of utmost importance to take this dependency

predicplot oftime iserror ais avaiThe prequati

Eper (%

Eper islimit.

Thesignalsaged pother cpredicpredicthe axi

Fig.errors t. Some input signals are used by a large number ofmples for excessively used input signals are the powerhe nacelle temperature. In turn that means that oncels behave abnormal, a number of models will give bad

inuences pattern. Futake into aa powerful

Fig. 8. 2D Waterfall plot of the normalized averaged0.20 0.09 m/s

. This behavior is shown in Fig. 8, where a 2D waterfallormalized averaged percentage prediction errors overalized. In this plot the colors indicate the prediction

itude. White areas mark periods where no prediction, e.g. due to missing data or non-operational periods.ion errors are normalized according to the following

E

3(E) 100 (4)

percentage error normalized by determined anomaly

r bar in Fig. 8 is set in a way that normally behavinghown with a light green color. Only if the one day aver-tion error exceeds the anomaly limit (see Section 6)

s become visible. The color range indicates whether therror is positive or negative as well as its amplitude. The

error of each model is shown in horizontal lines fromel on the left.ws that eight further models have abnormal predictione time the nacelle temperature increases. This effect

the analysis of the root cause of the prediction errorrthermore it emphasizes the need of algorithms thatccount the context of the patterns. Here fuzzy logic istool.

prediction errors.

M. Schlechtingen et al. / Applied Soft Computing 13 (2013) 259270 267

-2000

2000

4000

6000

Num

ber

of

Valu

es i

n b

in [-]

Generator curr ent ph.1

400 0

600 0

es i

n b

in [-]

Power output

0

2000

4000

PNum

ber

of

Valu

es in

bin

[-]

6. Anomal

Anomalydata that doing patternoutliers [28the central in this reseathat highliglimit. In 200about anomtechniques neighbor-bIn this papbecause of ntrained moaround zerothe assumpity regions oprobability

Due to tphase of thtistical procby a maximmay be biasselection of

The estiof the standform the updata. Instanthe traineded as anomcalculated a

Predictiontrained m

Predictiontrained m

The chosystem andto suppressof turbines

0 0.5 1 1.5 2 2.5

x 104

-10

-5

0

5

10

Value [ 10 min]

-10 -5 0 5 100

0.1

0.2

0.3

lower bound upper bou nd

ity [

-]

Gen erat or be aring te mp.

Predict ion err or [C]

norma l abnormalabnor mal

Probability density distribution and evolution of the 10 min avg. predictionTG 8).

l data as abnormal. False alarms may cause extra unsched-aintenance visits, or cause the operator to lose faith intem and subsequently ignore indications of real problems.gnosmainboun0 min

shoal b

d then ace anis leafurthold i

0.4

0.5

0.6 -10 0 0 100 200

Prediction error [A]

-200 -10 0 0 100 2000

200 0

Pred iction error [kW]Num

ber

of

Valu

-1 0 1

rediction err or [m/s]

Wind spe ed

-10 -5 0 5 100

200 0

400 0

600 0

Prediction error [C]Num

ber

of

Valu

es in

bin

[-]

Genera tor bearing temp.

Fig. 9. Prediction error distributions after training.

y denition and detection

detection refers to the problem of nding patterns in not conform to expected behavior. These nonconform-s are among others often referred to as anomalies or]. Detecting anomalies in the prediction error is one ofissues of the condition monitoring approach addressedrch. The task is to nd an anomaly detection algorithmhts any unexpected patterns, once they exceed a certain9 Chandola et al. [28] reported a comprehensive reviewaly detection algorithms. A majority of the reportedcan be categorized into classication-based, nearestased, clustering-based and statistical techniques [28].er the focus lies on parametric statistical techniques,ature of the prediction errors coming from successfully

dels that are usually normally distributed with a mean. Statistical anomaly detection techniques are based ontion that normal data instances occur in high probabil-f a stochastic model, while anomalies occur in the lowregions of the stochastic model [28].he availability of the prediction error in the traininge ANFIS models, the parameters of the underlying sta-ess can be estimated. In the current CMS this is doneum likelihood estimator, knowing that the parameters

Pre

dic

tio

n e

rror

[C

]P

rob

abili

ty d

en

s

Fig. 10. error (W

normauled mthe sysMisdiational lower tion (1Fig. 10abnormtion an

Whvarianturn th

To thresh

sity [

-]ed. Fig. 9 shows the prediction error distributions for a models after training.mated parameters are stored and used for estimationard thresholds for anomaly detection. These thresholdsper and lower bounds of the normal range of the SCADAces that have a low probability of being generated from

model, based on the applied test statistic, are classi-alies [28]. For this purpose the standard thresholds areccording to the following probabilistic assumptions:

errors that have a probability of being generated by theodel larger or equal than 0.01% are considered normal

errors that have a probability of being generated by theodel smaller than 0.01% are considered an anomaly.

ice of the probability inuences the sensitivity of the can be adjusted if required. Here 0.01% was chosen

false anomaly classication. Due to the large amountmonitored later, the approach taken must not classify

-5 0

0.1

0.2

0.3

Pro

ba

bili

ty d

en

a

0 -4

-2

0

2

4

Pre

dic

tion e

rror

[C

]

Fig. 11. Probaerror (WTG 8)is can result in replacement of the wrong part and addi-tenance to correct such errors [29]. The upper andd is calculated with regard to the original data resolu-

average) and the averaged values (one day averages).ws an example of the boundary between normal andehavior, both in terms of the probability density func-e prediction error in time domain.veraged over one day, the predictions have a reducedd the upper and lower bound move closer to zero. Inds to less alarm limit violations as visible in Fig. 11.er reduce the risk of false classications, an additionals set, that requires at least three values violating the

Gene rator bea ring te mp.-4 -3 -2 -1 0 1 2 3 4 5

lower bound upper bound

Predict ion err or [C]

normal abnor malbnor mal

50 100 150 200

Value [days]

bility density distribution and evolution of the one day avg. prediction.

268 M. Schlechtingen et al. / Applied Soft Computing 13 (2013) 259270

probability threshold within a week, before the anomaly is analyzedby the fuzzy expert system.

7. Fuzzy expert system

The Fuzization moanalyze pata potential In total ninwhich a FISare (see Fig

Meteorol Rotor sys Gearbox Generato Converter Transform Auxiliary Turbine p Nacelle

7.1. FIS inp

Only onethe predictiThe numbeof models tponent or swhen settinedge is impwhen rulesinput signabehavior mponent or spassed to thonly contaipresent. ThUsing the mvalues excestatements

7.2. FIS mem

Consideization modfunctions bangular andvariation ofdependencyproposed a in inductionwith triangucombinatiopriated for to trapezoiddened, wh

It provedify two sepFIS input ting the clashigh and veallowing clathe expert

0

0.2

0.4

0.6

0.8

13*lb lb ub 3*ub

normalhigh very highlowvery low

sh

ip

12. M

0

0.2

0.4

0.6

0.8

1

Mem

bers

hip

13. M

ns i2 and

FIS . Thi

, ruleoftwrmapper

stanntiden

ind

master threshold table, the four bounds can be denedlly. This is usually done after fault occurrence, or after deter-on of the critical prediction error levels. Manual adoption is

since if for instance a model is very accurate, this consequen-lso leads to tight standard upper and lower bounds. Hencediction error may have a high membership in the MF verylthough the prediction error level is still considered to becal. In this case the thresholds can be adopted in the masterion threshold le to match the real fault progression. Thisure is required, since the real bounds are unknown for mostnents and signals before fault occurrence. The thresholdsd as master thresholds are valid for all turbines and replaceividual ones. It is therefore possible to identify similar pat-n the whole eet, since it is expected that the general faultteristics are similar for turbines of the same type.

outputs of the FIS are diverse with regard to their MFs. One ofee outputs are initialized for each FIS equally. This concernstialization of the output condition, visualized in Fig. 14.

MF gray is reserved for periods, in which no diagnosis is pos-ue to missing data, whereas green, yellow and red indicatedition status.

n: component/subsystem working as expected; condition; state okzy expert system consists of the fuzzy expert initial-dule and the fuzzy expert application module whichterns to classify the component conditions and outputroot cause if a matching rule is found in the rule base.e components or subsystems were dened, for each of

structure is developed. The components or subsystems. 3):

ogytem

r

erequipmenterformance

uts

FIS is developed per component or subsystem, henceon error of several ANFIS models is processed by the FIS.r of inputs to the FIS therefore depends on the numberhat exist to monitor the signals behavior of the com-ubsystem. The relevant signals for the FIS are denedg up the FIS structure the rst time, i.e. when the knowl-lemented. However, the list of inputs is extendable,

implemented at a later stage require more inputs. Thels are selected by considering the corresponding normalodel inputs and the physical understanding of the com-ubsystem. For each input, a modied prediction error ise input space of the FIS. The modied prediction error

ns the true prediction error value when an anomaly ise prediction error is set to a small number otherwise.odied prediction error has the advantage that singleeding the anomaly limits do not cause false condition.

bership functions

ring which inputs are used, the fuzzy expert initial-ule initializes the FIS with regard to the membershipelonging to each modied prediction error input. Tri-

trapezoidal MFs are used in this research. A linear the membership values is reasonable as long as no other

is known or desired. In 2008 Rodrgues and Arkkio [30]diagnosis method for detection of stator winding faults

motors based on fuzzy logic. The system was testedlar, trapezoidal and Gaussian MFs. It was found that the

n of triangular and trapezoidal MFs is the most appro-fault diagnosis in induction motors [30]. In comparisonal MFs, triangular MFs require one less parameter to beich is why triangular MFs are preferred in this research.

useful in expert knowledge implementation to spec-arate FIS inputs per relevant prediction error. The rstype is with ve membership functions (MFs) allow-sication into the categories: very low, low, normal,ry high. The second FIS input type is with three MFsssication into the categories: low, normal, high, givingfull freedom to implement the rules. The membership

Me

mb

er

Fig.

Fig.

functioFigs. 1

Theformedinputsfewer slow, no the u

Thefor ideThese ues are6.

In amanuaminatiuseful,tially athe prehigh, auncriticonditprocedcompodenethe indterns icharac

Thethe thrthe ini

Thesible dthe con

GreegoodInput variable: Anomaly error (5 MF`s)

lb = lower

boundub = upper

bound

embership functions initialized for each input: FIS input type 1.

normal highlow

Input var iable: Ano maly error (3 MF`s)

lb ub

lb = lower bou ndub = upper boun d

embership functions initialized for each input: FIS input type 2.

nitialized for each of these inputs are visualized in 13.structures are initialized each time an analysis is per-s has the advantage that modications of the structures and membership functions position and shape lead toare conicts. The denition of the statements, very low,l, high and very high is given by the standard thresholds

and lower bounds according to the 0.01% probability.dard thresholds do only serve as a rst approximationcation and classication of the component condition.itions are equal for all signals and turbines, but the val-ividual, due to the anomaly denition given in Section

M. Schlechtingen et al. / Applied Soft Computing 13 (2013) 259270 269

0

0.2

0.4

0.6

0.8

1grey re d yello w gree n

Output var iable: Cond itio n

Mem

bers

hip

Fig. 14. Initialization of the membership functions of the variable condition.

Yellow: service shall be scheduled for component/subsystem toimprove the condition or investigate the issue; condition bad;state warning

Red: service shall be performed at component/subsystem toimprove the condition; condition very bad; state alarm

The other two outputs contain information about the diagnosisand the potential root cause. This information is individual to eachFIS and is part of the knowledge that needs to be supplied by anexpert. The setup is shown in Figs. 15 and 16.

7.3. FIS rules

The MFs are linked via rules that allow the FIS to output thediagnosis, condition and potential root cause. The rules are imple-mented by an expert manually. The concept of the rule base isthat the base is constantly extended to allow the FIS to iden-tify the various fault types accurately. The rule base is updatedafter fault occurrence and the validity of the already existing rulesis checked.

0

0.2

0.4

0.6

0.8

1

Mem

bers

hip

Fig. 15. S

0

0.2

0.4

0.6

0.8

1

Mem

bers

hip

Fig. 16. Schem

memberships according to their MFs and selects the rule that ismost applicable. The consecutives are then given according to thisrule. All rules currently implemented employ the and method,taking the minimum value of the MF evaluation for decision on therule applica

generic ru specic ru

The geneno specic the specicoperator thon the otheof impleme

(1) If (Anomerror (3temp. defect)

(2) If (Anom(Anoma(Diagno(Pot. ro

The ruleday average

As a basevaluating ble rule is uhigh value igiving a hig

ults

deveifferbilit

ter toces.

into turb

of ththe mfunch thg.

focuce tcreafalse peratte reor th For each set of inputs, the FIS evaluates the input

diag . 1 diag. 2 diag. 3 .. . diag. n

Output variable: Diagnosis

chematic of the membership function of the variables diagnosis.

root 1 root 2 root 3 ... root n

8. Res

Theety of dapplicaare fasformantakingels permodesany of or malresearctrainin

Theto reduonly infewer bine oaccurabasis fOutput variab le: Pot. root cause

atic of the membership function of the variables potential root cause.

structure seexpected to

In the fudrastically vant predic1 is used forwhereas tydifferentiatIf for instanby a speciing temperbility. The rules can be classied into two categories:

lesles

ric rules are used to highlight present anomalies, in caserule applies. This rule type gives no information about

condition or potential root cause, but highlights to theat an anomaly is present in the data. The specic rulesr hand provide this level of information. Two examplesnted specic rules are:

aly error (5 MFs) Spinner temp.==high) & (Anomaly MFs) Nacelle temp.==ok) then (Diagnosis = Spinnerhigh) (Condition = yellow) (Pot. root cause = Sensor

aly error (5 MFs) Generator ph.1 temp. cross==high) &ly error (3 MFs) Generator ph.3 temp. cross==low) thensis = Generator Ph.1 temp. to high) (Condition = yellow)ot cause = Generator phase insulation damage)

s are implemented to automate the analysis of the one prediction errors.

is for certainty statements of the diagnosis, the result ofthe output values through the MFs for the best applica-sed. This value can be seen as a matching indicator. Andicates that the rule represents well the input pattern,h certainty in the diagnosis.

and discussion

loped ANFIS models show good performances at a vari-ent SCADA data (see Table 5), which proves their generaly in this context. Moreover it is shown that these models

train than NN models by achieving similar or better per- The differences in training speed are signicant, whenaccount that the current development counts 45 mod-ine which require training. This is, because the failuree turbines are not known beforehand. An anomaly inonitored SCADA signals may indicate a potential fault

tion. For the eighteen turbines accessible during thisis leads to 810 different ANFIS models, which require

s in this work is on one day average predictions in orderhe uctuations in the prediction error. This does notse the fault visibility, but it also is expected to lead toalarms. False alarms are of major concern for wind tur-ors, due to the large number of operating systems. Anpresentation of the signal normal behavior builds thee subsequent analysis of occurring patterns by the FISt up. The overall performance levels achieved here are

allow early anomaly detection.zzy expert system the overall number of rules can be

reduced by specifying two separate FIS inputs per rele-tion error as proposed in this research. The FIS input type

the signals directly linked to the component monitored,pe 2 may be used for all other inputs where the neion of the prediction error magnitude is not necessary.ce the gearbox bearing condition shall be determinedc rule, it is useful using type 1 for the gearbox bear-ature to be able to classify the condition according to

270 M. Schlechtingen et al. / Applied Soft Computing 13 (2013) 259270

the prediction error magnitude (high and very high). For all otherinputs, the ne differentiation may not be necessary, making FIStype 2 for those inputs more useful. This reduces the number ofpossible combinations within the rule.

9. Conclusions

In this paper a method to monitor wind turbine SCADA data, vianormal behavior models and fuzzy logic is proposed. The method-ology gives the possibility of mining large amount of SCADA dataavailable to wind turbine operators in a systematic manner, searchfor anomalies and use this information for condition statements. Itallows not only monitoring of large components, but also auxiliaryequipment that currently available wind turbine CMS do not cover.

Using fuzzy logic the existing expert knowledge inanomaly/prediction error pattern interpretation and root causediagnosis can be implemented in an intuitive manner. This givesthe possibility for automated fault diagnosis once specic rules areimplemented.

However, the applicability of the proposed method and there-with the acvariety of dcriterion is posed methlimitation cexpert knowto highlightstatements

Future rsurements,analysis in tmenting rufor larger ddifferent ty

In part presented bdeveloped proposed mally with rethat the FISvalid for althe developknowledge.

References

[1] A. Zahertion thro574593.

[2] R.B. RandSussex, U

[3] W. Yang, J. Jiang, Wind turbine condition monitoring and reliability analysis bySCADA information, IEEE (2011) 18721875.

[4] A. Kusiak, L. Wenyan, The prediction and diagnosis of wind turbine faults,Renewable Energy 36 (2011) 1623.

[5] A. Kusiak, A. Verma, A data-driven approach for monitoring blade pitch faultsin wind turbines, IEEE Transactions on Sustainable Energy 2 (1) (2011) 8796.

[6] T.W. Verbruggen, Wind Turbine Operation & Maintenance based on ConditionMonitoring, ECN-C-03-047, 2003.

[7] M.A. Sanz-Bobi, J. del Pico, M.C. Garcia, SIMAP: Intelligent System for PredictiveMaintenance Application to the health condition monitoring of a windturbinegearbox, Computers in Industry 57 (2006) 552568.

[8] M. Schlechtingen, I.F. Santos, Comparative analysis of neural networkand regression based condition monitoring approaches for wind turbinefault detection, Mechanical Systems and Signal Processing 25 (5) (2011)18491875.

[9] W.G. Garlick, R. Dixon, S.J. Watson, A model-based approach to wind turbinecondition monitoring using SCADA data, in: 20th International Conferenceon Systems Engineering, 810 September, Coventry University, Coventry, UK,2009.

[10] M. Sttmann (Schlechtingen), Master Thesis: Condition Monitoring in WindTurbines A Drive Train Monitoring System, 2010.

[11] A.S. Zaher, S.D.J. McArthur, A multi-agent fault detection system for wind tur-bine defect recognition and diagnosis, in: Proc. IEEE Lausanne POWERTECH,2007, pp. 2227.

[13] J. Xiang, S. Watson, Y. Liu, Smart monitoring of wind turbines using neuralnetworks, Sustainability in Energy and Buildings Part 1 (2009) 18.

R. Jangions o. Gae, t Comhand

ce on posi

arasseord, U. Ra

appli Ross, sex, Uim, Nir app113. Yen,lth m1 (4) (. Mamstic sySugenper Saeardi,ressioy SystLearddels.liR. Jangrningk, USA

Wigghanding Su. Hye

nitori06) 18.J. Rodng fuzhieved accuracies rely on the availability of a broadifferent SCADA data to set up the ANFIS models. Thisusually fullled, which allows application of the pro-od to both existing older and new turbines. A furtheroncerns the availability of expert knowledge. In case noledge is available the methodology can only be used

present anomalies in the SCADA data and only general about the condition and the diagnosis can be made.esearch will focus on the implementation of other mea-

such as high resolution vibrational data or oil samplehe CMS. This will open even more possibilities in imple-les that capture the expert knowledge and will allowiversity of the diagnoses. Furthermore turbines from ape and brand will be included in the research.two of the article the results of a eld test will bey giving examples that show the applicability of the

system based on real faults. It will be shown that theethod is capable of treating every turbine individu-gard to its characteristic. Additionally it will be shown

structures proposed allow generalizing rules that arel turbines of the same type. This drastically reducesment effort that is required to implement the expert

, S.D.J. McArthur, D.G. Ineld, Online wind turbine fault detec-ugh automated SCADA data analysis, Wind Energy 12 (2009)

all, Vibration-based Condition Monitoring, Wiley, Chichester, WestK, 2011, ISBN 9780470747858.

[14] J.S.act

[15] X.ZSof

[16] K. CforCom

[17] L. TOxf

[18] M.Ying

[19] T.J.Sus

[20] J. Kthe130

[21] G.GheaB 3

[22] E.Hgui

[23] M. Up

[24] R. Lregtor

[25] R. mo

[26] J.S.LeaYor

[27] E.J.[28] V. C

put[29] R.W

mo(20

[30] R.Vusi, ANFIS: adaptive-network-based fuzzy interference system, Trans-n Systems, Man and Cybernetics 23 (3) (1993) 665685.S.J. Ovaska, Soft computing methods in motor fault diagnosis, Appliedputing 1 (2001) 7381.rashekhara, A. Chukwujekwu Okafor, Y.P. Jiang, Estimation of contactcomposite plates using impact-induced strain and neural networks,te Part B 29B (1998) 363370.nko, Guide to Neural Computing Applications, Elsevier, Jordan Hill,SA, 1998, ISBN 0340705892.q, G. Bugmann, D.J. Easterbrook, Neural network design for engineer-cations, Computers and Structures 79 (2001) 15411552.Fuzzy Logic with Engineering Applications, Wiley, Chichester, WestK, 2010, ISBN 9780470743768.. Kasabov, HyFIS: adaptive neuro-fuzzy interferance systems andlication to nonlinear dynamical systems, Neural Networks 12 (1999)19.

P. Meesas, An effective neuro-fuzzy paradigm for machinery conditiononitoring, IEEE Transactions on Systems, Man and Cybernetics Part2001) 523536.dani, Application of fuzzy logic to approximate reasoning using lin-nthesis, Transactions on Computers 23 (12) (1977) 11821191.o, Industrial Applications of Fuzzy Control, Elsevier Science Publishers,ddle River, New York, USA, 1985.

A.L. Gonzlez, Genetic algorithms applied to feature selection in PLSn: how and when to use them, Chemometrics and Intelligent Labora-ems 41 (1998) 195207.i The Pls-genetic Algorithm Toolbox for Matlab. http://www.fe.ku.dk/GAPLS, Neuro-Fuzzy and Soft Computing: A Computational Approach to

and Machine Intelligence, Prentice Hall, Upper Saddle River, New, 1997, ISBN 978-0132610667.

elinkhuizen, et al., Conmow Final Report, ECN-E07-044, 2007.ola, A. Banerjee, V. Kumar, Anomaly detection: a survey, ACM Com-rveys 41 (3) (2009), 15:115:58.rs, J.G. McGrowan, K.L. Sullivan, J.F. Manwell, B.C. Syrett, Conditionng and prognosis of utility scale wind turbines, Energy Materials 1 (3)7203.rgues, A. Arkkio, Detection of stator winding fault in induction motorzy logic, Applied Soft Computing 8 (2008) 11121120.

Wind turbine condition monitoring based on SCADA data using normal behavior models. Part 1: System description1 Introduction2 Description of the general CMS concept2.1 Training module2.2 Prediction module2.3 Anomaly detection module2.4 Fuzzy expert initialization module2.5 Fuzzy expert application module

3 Model setup and structure3.1 Data set description3.2 Model type3.2.1 FIS structure3.2.2 Membership functions3.2.3 Number of membership functions/rules3.2.4 Training method3.2.5 Input signals

4 Model performance4.1 Comparison between NN and ANFIS models4.2 Performance of the set up ANFIS models

5 Prediction errors6 Anomaly definition and detection7 Fuzzy expert system7.1 FIS inputs7.2 FIS membership functions7.3 FIS rules

8 Results and discussion9 ConclusionsReferences