Research Article A Fault Diagnosis Approach for Gas Turbine Exhaust Gas Temperature Based on Fuzzy C-Means Clustering and Support Vector Machine

Transcription

1 Mathematical Problems i Egieerig Volume 215, Article ID 24267, 11 pages Research Article A Fault Diagosis Approach for Gas Turbie Exhaust Gas Temperature Based o Fuzzy C-Meas Clusterig ad Support Vector Machie Zhi-tao Wag, 1 Nig-bo Zhao, 1 Wei-yig Wag, 1,2 Rui Tag, 2 ad Shu-yig Li 1 1 College of Power ad Eergy Egieerig, Harbi Egieerig Uiversity, Harbi 151, Chia 2 Harbi Marie Boiler & Turbie Research Istitute, Harbi 1578, Chia Correspodece should be addressed to Nig-bo Zhao; zhaoigbo314@126com Received 16 October 214; Revised 26 November 214; Accepted 27 November 214 Academic Editor: Erico Zio Copyright 215 Zhi-tao Wag et al This is a ope access article distributed uder the Creative Commos Attributio Licese, which permits urestricted use, distributio, ad reproductio i ay medium, provided the origial work is properly cited As a importat gas path performace parameter of gas turbie, exhaust gas temperature (EGT) ca represet the thermal health coditio of gas turbie I order to moitor ad diagose the EGT effectively, a fusio approach based o fuzzy C-meas (FCM) clusterig algorithm ad support vector machie (SVM) classificatio model is proposed i this paper Cosiderig the distributio characteristics of gas turbie EGT, FCM clusterig algorithm is used to realize clusterig aalysis ad obtai the state patter, o the basis of which the preclassificatio of EGT is completed The, SVM multiclassificatio model is desiged to carry out the state patter recogitio ad fault diagosis As a example, the historical moitorig data of EGT from a idustrial gas turbie is aalyzed ad used to verify the performace of the fusio fault diagosis approach preseted i this paper The results show that this approach ca make full use of the usupervised feature extractio ability of FCM clusterig algorithm ad the sample classificatio geeralizatio properties of SVM multiclassificatio model, which offers a effective way to realize the olie coditio recogitio ad fault diagosis of gas turbie EGT 1 Itroductio With the developmet of high efficiecy ad clea eergy, gas turbie plays a icreasigly sigificat role i differet domais, such as aviatio ad marie propulsio systems, electric power statios, ad atural gas trasportatio petroleum [1] With the icreasig demad of security operatio for gas turbie, the traditioal regular maiteace techology has bee uable to fully keep up with the actual demad ad gas turbie health maagemet techology has gradually become oe of the most problems cocered by researchers ad users i recet years [2] I order to guarateethegasturbietoruefficietlyuderthesafe reliable coditio, may sesors are ofte used to moitor the health state of gas turbie i the practical applicatio Massive amouts of data gathered by these sesors are easy to make difficulties i data aalysis ad affect the maiteace decisio Therefore, the choices of appropriate moitor parameters, sigal processig methods ad data miig techiques are very importat to realize the health maagemet of gas turbie Exhaust gas temperature (EGT) is a importat gas path performace parameter of gas turbie, which ca represet the thermal health coditio of gas turbie [3, 4] Cosiderig the characteristics of differet gas path performace moitorig parameters, the multiple liear regressio models for aalyzig the relatioship betwee EGT ad other parameters were established by Sog et al [5] Their results showed that there were strog liear correlatios betwee differet gas path performace parameters ad all the low turbie outlet pressure, high rotatioal speed, high pressure compressor outlet temperature, low rotatioal speed, ad high pressure compressor outlet pressure could be reflected through the chage of EGT Yilmaz [6] also foud the similar results by aalyzig the relatioship betwee EGT ad other egie operatioal parameters at two differet power settigs, icludig maximum cotiuous ad take-off, i the CFM56-7B turbofa egie Hece, EGT is ofte used

2 2 Mathematical Problems i Egieerig as a importat parameter to evaluate the health state of gas turbie ad determie the maiteace policy [7] I the past half cetury, differet methods have bee developed to moitor ad diagose the EGT of gas turbie Wag ad Yag [8] aalyzedmayfaultsofpg6551b idustrial gas turbie, such as turbie ablatio, combustio compoet, ad fuel system failure They foud that the uiformity of EGT could effectively reflect the feature of above fault Che et al [9] proposed a geeral regressio eural etwork (GRNN) approach to costruct a autodetectio etwork for EGT sesors, o the basis that they also studied the optimizig desig of etwork ad error cotrollig ad developed the method of threshold for sesor detectio Based o the advatage of artificial eural etworks (ANN), Muthurama et al [1] developed a autoassociative eural etwork approach to detect combustor-related damage by moitorig EGT Błachio ad Pawlak [11] established a oliear observer ad chose EGT as the importat parameter to evaluate the health state of turbie blades Korczewski [12 14]aalyzed the chage rules of EGT for a aval gas turbie egie uder steady ad usteady operatio coductio i detail Ad they proposed a effective approach for detectig ad evaluatig the failures of the flow sectio ad supply system of gas turbie by usig EGT Keyo et al [15] developed a itelliget system for detectio of EGT aomalies i gas turbies by usig the strog oliear mappig ability of ANN Cosiderig the characteristics of gas turbie operatio cotrol based o thermocouple measured exhaust temperatures, Xia et al [16] discussed the applicatio of Fiber-Bragg-gratig-based sesigtechologyitheegtmeasurigofgasturbie Their results demostrated that the fiber sesig method was more valuable for the moitorig ad fault diagosis of gas turbie because it could well reflect the chagig of EGT I order to icrease the operatioal availability of idustrial gas turbies, Yag et al [17] presetedageeralizatioof multidimesioal liear regressio to facilitate multisesor fault detectio ad sigal recostructio through the use of aalytical optimizatio Güle et al [18] discussedthe relatioship betwee EGT ad other gas turbie performace parameters, o the basis of which a importat diagostic parameter amed profile factor that was the ratio of the maximum exhaust thermocouple ad the average of all exhaust temperature thermocouples was used to evaluate the performace of combustor ad the whole gas turbie i their paper From the reviews discussed above, it is oted that the average EGT is ofte used to evaluate the health state of gas turbie i most researches However, it may be more importat ad valuable to extract the relatioship amog differet EGT sesors i order to realize coditio moitor ad fault diagosis of gas turbie effectively Although maystudieshavepresetedtheeffectsofegtdistributio characteristics o health state of gas turbie or its hot sectios ad may aalyses were discussed i detail, there was still a lack of systematic research i the area of olie automatic idetificatio ad fault diagosis for gas turbie EGT Besides, EGT ca be affected by may ucertai factors i the practical applicatios, which make it difficult to realize fault diagosis quickly by usig traditioal model-drive approach Therefore, it is very useful to develop a data-drive approach based o artificial itelligece techology i order to improve fault diagosis accuracy The fault diagosis of gas turbie EGT based o datadrive approach essetially is the cluster ad classificatio of fault iformatio I the cocrete implemetatio process, the traiig samples icludig ormal ad fault iformatio eed to be obtaied firstly The the fault diagosis model based o artificial itelligece algorithm ca be established ad traied by usig traiig samples The feature iformatio collected from sesors will be iputted to the well traied fault diagosis model ad we ca get the diagostic results fially Obviously, the establishmet of fault iformatio features space ad desig of artificial itelligece algorithm are two key steps to realize accurate fault diagosis of gas turbie EGT As metioed above, may specific state patters or fault types of gas turbie EGT caot be determied directly based o experiece i the practical applicatios due to the effects of may ucertai factors Besides, eough prior kowledge icludig specific fault types is idispesable for the supervised artificial itelligece algorithms (such as ANN [19]adSVM[2]) Cosiderig the distributio characteristics of gas turbie EGT ad the deficiecies of preset literatures, a fusio approach based o FCM clusterig algorithm ad SVM classificatio model (FCM-SVM) is proposed i this paper Firstly, FCM clusterig algorithm is used to realize clusterig aalysis ad obtai the state patters of EGT, which meas that the preclassificatio of EGT is completed The, SVM multiclassificatio model is desiged to carry out the state patter recogitio ad fault diagosis of EGT As a example, the historical moitorig data of EGT from a idustrial gas turbie is aalyzed to verify the effectiveess of the FCM- SVM approach fially The rest of this paper is orgaized as follows I Sectio 2, the distributio characteristics of gas turbie EGT are described briefly Sectio 3 itroduces the basic theory of FCM clusterig algorithm ad SVM classificatio model i detail The fusio fault diagosis approach which combies FCM clusterig with SVM is discussed i Sectio 4 Applicatio examples ad discussio are icluded i Sectio 5 Fially, Sectio 6 presets some coclusios 2 Sigal Feature of Gas Turbie EGT As metioed above, it is very importat to choose the appropriate measured parameters to moitor ad diagose the health state of gas turbie Gas path ad vibratio parameters are two mai types i the practical applicatios [2, 21] Theoretically, as the most importat gas path performace parameter for gas turbie, the outlet temperature of combustor chamber ca ot oly affect the overall performace of egie, but also directly determie the ultimate stregth of turbie blade For example, the creep life of hot chael compoets ca reduce the order of magitude whe the outlet temperature of combustor chamber icreases 5 C[4], which may cause major fault ad icur great maiteace costs However, the outlet temperature of combustor chamber

3 Mathematical Problems i Egieerig 3 Temperature ( C) EGT (ormal) Average EGT Miimum EGT Maximum EGT 3 33 Figure 1: EGT profile of gas turbie with ormal coditio EGT of a idustrial gas turbie that has 12 EGT sesors From Figures 1 ad 2, it is easy to see that the EGT profiles of ormal ad fault coditio are differet although the average EGT are the same This meas that the fault iformatio is icorrect or icomplete if oly the average EGT is used to moitor the health state of gas turbie Besides, Figures 1 ad 2 also show that all the sesors should give similar outputs whe gas turbie operates i ormal coditio If the compoet of gas turbie is failure, differet temperatures will be observed Therefore, the uiformity of EGT ca more effectively reflect the health state of gas turbie, especially for steady state coditio I order to quatitatively describe the uiformity of EGT, Mao [22] preseted three idexes which ca be calculated by the followig fuctios Assume that there are sesors ad their origial outputs are T i, i=1,,,respectively: H 1 =T 1 T, H 2 =T 1 T 1, H 3 =T 1 T 2, (1) Temperature ( C) where T i is the trasform value of T i ad T 1 >T 2 > > T 2 >T 1 >T Accordig to above idexes, it is obvious that all the values of H 1, H 2,adH 3 are smaller whe gas turbie operates i ormal coditio If a sesor fails, it usually causes H 1 or three idexes to icrease But the values of H 1 ad H 2 or all three idexes ca icrease due to hot sectios failure However, it is worth poitig out that the above idexes oly cosider 4 EGT ad others are igored Besides, it is difficult to diagose the specific cause of failure whe the sesor is fault because the above idexes igore the adjacet iformatio betwee differet sesors Therefore, there are some limitatios to evaluate the uiformity of EGT oly usig the above three idexes I order to solve this problem, all the measured EGT are used to realize cluster ad fault diagosis i this paper EGT (fault) Average EGT Miimum EGT Maximum EGT Figure 2: EGT profile of gas turbie with fault coditio isusuallysohighthatitcaotbemeasureddirectlybyusig covetioal sesors Accordig to the well-defied Brayto thermodyamic cycle, there is a cosistet relatioship betwee the outlet temperature of combustor chamber ad EGT Therefore, EGT, as a measured parameter, is ofte used for gas egie cotrol, coditio moitorig, fault diagosis, ad maiteace decisios Compared with the average EGT, EGT profile ca cotai more iformatio about the health state of gas turbie Figures 1 ad 2 show two EGT profiles with the same average 3 Basic Theory 31 Fuzzy C-Meas Clusterig Algorithm As a usupervised machie learig method, FCM clusterig algorithm was improved by Bezdek [23] i1981iordertosolvethe hard clusterig problem by usig fuzzy set theory I the FCM clusterig algorithm, membership degree fuctio is used to idicate the extet to which each data poit belogs to each cluster, ad this iformatio is also used to update the values of cluster ceters [24] Based o the cocept of fuzzy C-partitio, FCM clusterig algorithm has bee applied successfully i a wide variety of applicatios, such as image segmetatio [25], data miig [26], thermal system moitorig [27], ad fault diagosis [28] For the sample set X={x 1,x 2,,x },theobjectoffcm clusterig algorithm is to divide the sample set ito c groups

4 4 Mathematical Problems i Egieerig ad obtai the cluster ceters by miimizig the followig dissimilarity fuctio [29]: c mi J m (U, V) = u m ik d2 ik, (2) i=1 k=1 where m is the fuzzy weightig parameter varyig i the rage [1, ] Thebiggerthem, the more fuzzy the fial cluster result U is fuzzy partitio matrix, V is cluster ceter matrix, ad ad c are the umber of samples ad cluster ceters, respectively u ik is the fuzzy membership degree of the kth sample i the ith cluster ad it should be meetig the followig three costraits [3]: s=s+1 Start Give iterative threshold ad iitialize fuzzy partitio matrix radomly U () Calculate the cluster ceter matrix V Compute Euclidea distace u ik [, 1], 1 i c, 1 k, Update fuzzy partitio matrix U (s+1) c i=1 k=1 u ik =1, u ik (, ), 1 k, 1 i c For the distace d ik betwee kth sample x k ad the cetre of ith cluster V i, it ca be calculated by usig Euclidea distace as follows: m d ik = x k V i = (x kj V ij ) 2, j=1 (3) 1 i c, 1 k (4) No U (s+1) U (s) ε Yes Output cluster result Figure 3: Calculatio flowchart of fuzzy C-meas clusterig algorithm φ(x) V ij cabecalculatedbyutilizigthefollowigformulatio: V ij = k=1 um ik x kj, 1 i c, 1 j m (5) k=1 um ik I essece, fuzzy cluster is performed through a iterative optimizatio by updatig fuzzy membership degree [29]: u (s+1) ik = 1 c j=1 (d(s) ik /d(s) jk )2/(m 1), 1 i c, 1 k, (6) where s is the iterative step Whe the below requiremet is met, we ca stop iteratio ad obtai the cluster result: U(s+1) U (s) ε, (7) where ε is the iterative threshold i the rage [, 1] Basedoabovemethod,theclusterprocessofFCM clusterig algorithm is virtually to determie the fuzzy membership degree ad cluster ceters through cotiuous iteratio, which is show i Figure 3 32 Support Vector Machie Classificatio Model Compared with the covetioal classifiers, support vector machie (SVM), developed by Vapik [31], ca effectively solve the classificatio problem by implemetig the structure risk Origial space Higher dimesioal feature space Figure 4: Classificatio of two classes usig SVM miimizatio based o statistical learig theory Nowadays, SVM has bee widely ad successfully applied to detectio ad diagosis of machie coditios due to its high accuracy ad good geeralizatio for a smaller umber of samples [32, 33] SVM is iitially used to deal with biary classificatio problems Its core idea is to trasform the sample data from origial space to a higher dimesioal feature space through some oliear mappig fuctios ad the fid the optimal separatig hyperplae i this feature space to realize liear classificatio Figure 4 shows the classificatio priciple based o SVM for the oliear classificatio problem For the oliear traiig sample data set icludig two classes {x i,y i }, i = 1,2,,, x R, y { 1,+1}, istheumberofsamplestheoliearmappigfuctio φ(x) will be used to trasform the sample data from origial space to a higher dimesioal feature space ad the optimal

5 Mathematical Problems i Egieerig 5 separatig hyperplae ca be costructed to maximize the margi betwee the two classes by the followig liear fuctio: f (x) = [ω φ(x)] +b, (8) where ω is the ormal vector of optimal separatig hyperplae ad b is a scalar I essece, the solutio of optimal separatig hyperplae is the correspodig costrait optimizatio problem: mi ζ i i=1 1 2 ω 2 +C st y i [ω φ (x i )+b]+ζ i 1, 1 i ζ i, 1 i, where C is the pealty factor that ca realize the trade-off betwee empirical risk ad cofidece iterval ζ i is slack factor Combiig the method of Lagrage multipliers, the above covex optimizatio problem ca be simplified ito the dual quadratic optimizatio problem: max L (a) = i i=1a 1 a 2 i a j y i y j φ(x i ) φ(x j ) i,j=1 st a i, 1 i i=1 a i y i =, where a is Lagragia multiplier The, the oliear decisio fuctio is described as f (x) = sig ( i,j=1 (9) (1) a i y i (φ (x i ) φ(x j )) + b) (11) I order to calculate the value of φ(x i ) φ(x j ),thekerel fuctio K(x i,x j ) is used ad the above fuctio ca be expressed as f (x) = sig ( i,j=1 a i y i K (x i,x j ) +b) (12) For the SVM, there are may kids of kerel fuctio, such as liear kerel, polyomial kerel, polyomial kerel, ad radial basis fuctio (RBF) kerel Compared with other kerel fuctios, the RBF kerel ca obtai the higher classificatio accuracy i may practical applicatios [34] Therefore, the RBF kerel is used i this study As previously metioed, SVM is iitially desiged for biary classificatio However, there are ofte may faults i the practical applicatios, which mea that it is ecessary to develop a method to deal with a multiclassificatio problem Curretly, differet methods have bee developed for the multiclassificatio based o SVM, such as oe-agaist-oe, oe-agaist-all, ad directed acyclic graph (DAG) Accordig to the compariso results obtaied by Hsu ad Li [35], the oe-agaist-oe method is more suitable for practical use tha other methods For the sample set icludig c class, c(c 1)/2 SVM classifiers ca be costructed by usig oeagaist-oe method ad every SVM classifier is traied 4 Fusio Fault Diagosis Modelig of EGT Based o FCM-SVM Approach As the supervised artificial itelligece method, eough prior kowledge icludig specific fault types is ecessary for SVM classificatio model However, the fault types of gas turbie EGT caot be determied directly based o experiece i the practical applicatios due to the effects of may ucertai factors I order to achieve the automatic moitor ad diagosis of EGT effectively, a fusio approach based o FCM clusterig algorithm ad SVM classificatio model (FCM-SVM) is proposed Firstly, FCM clusterig algorithm is used to realize clusterig aalysis ad obtai the state patters, which meas that the preclassificatio of EGT is completed The, SVM multiclassificatio model is desiged adusedtocarryouttheoliestatepatterrecogitioad fault diagosis of gas turbie EGT Figure 5 shows the fusio fault diagosis framework of gas turbie EGT based o FCM-SVM approach The detailed modelig processes are as follows Step 1 Geerated sufficiet EGT samples from the historical database ad the essetial preprocessig upo EGT data are carried out before data aalysis, such as supplemetary data, elimiatig oise ad outliers Step 2 Accordig to the cluster process which is show i Figure 3, FCM clusterig algorithm is used to obtai the iitial clusterig results of gas turbie EGT Step 3 Cluster validity idex λ(c) is used to evaluate the validity of clusterig ad determie the umber of clusters The c is optimum whe λ(c) reaches its maximum value: λ (c) = c i=1 ( k=1 um ik ) V i x 2 / (c 1) c i=1 k=1 um ik x k V i 2 / ( c), (13) x= c i=1 k=1 um ik x k (14) Step 4 After obtaiig the optimal clusterig results, the fault diagosis sample set icludig specific fault types ca be established Step 5 SVM multiclassificatio model will be desiged based o oe-agaist-oe method ad traied by usig fault diagosis sample set Step 6 The measured EGT obtaied from real gas turbie are preprocessed ad iputted to the well traied SVM multiclassificatio model The we ca get the fial diagostic results Step 7 The measured EGT also are stored ito the historical database ad used for later aalysis

6 6 Mathematical Problems i Egieerig Historical database EGT state patter extractio based o FCM clusterig algorithm Data preprocessed Iitial clusterig umber Clusterig aalysis based o FCM clusterig algorithm Adjust clusterig umber Evaluate cluster validity idex No Optimal clusterig? Yes Olie measured EGT EGT fault diagosis based o SVM multiclassificatio model Establish fault diagosis sample set SVM classifier 1 SVM classifier 2 SVM classifier N Comprehesive evaluatio Data preprocessed Well traied SVM multiclassificatio model Output result Figure 5: Fusio fault diagosis framework of gas turbie EGT based o FCM-SVM approach 5 Case Study ad Discussios I order to demostrate the effectiveess of FCM-SVM approach itroduced i this paper, the historical moitorig data of EGT from oe idustrial sigle shaft gas turbie will be aalyzed as a case study i this sectio 51 Sample Data As a idustrial sigle shaft gas turbie, Taurus7 is made i solar turbies icorporated ad used for power geeratio 12 thermocouple temperature sesors are used to measure the EGT ad the average EGT is about 55 C whe gas turbie operates i a ormal state Figure 6 shows the chagig curves of 12 EGT varyig with time uder ormal ruig state of gas turbie Ad the EGT profiles ca be see i Figure 7 FromFigures6 ad 7, it is clear that there is sigificat differece betwee the measured outputs of differet thermocouple temperature sesors at the same time eve whe the gas turbie is ruig i a ormal state Therefore, much feature iformatio will be igored which ca decrease the fault diagosis accuracy ifolytheaverageegtisusedtoevaluateadaalysethe health state of EGT Cosiderig the operatig coditios of gas turbie, 49-group data icludig 4 classes are take to establish the origial sample set 47 samples are selected radomly as traiig samples ad the remaiig 2 samples are selected as testig samples 52 Optimal Clusterig of EGT Based o FCM Clusterig Algorithm FortheFCMclusterigalgorithm,itisveryimportat to determie a appropriate umber of clusters, which is calledclustervalidityproblemithisstudy,thecluster umber is decided automatically by usig the itroduced cluster validity idex which is show i (13) Cosiderig the computatio complexity ad accuracy, the scope of the

7 Mathematical Problems i Egieerig Temperature ( C) Cluster validity idex Average EGT EGT 1 EGT 2 EGT 3 EGT 4 EGT 5 EGT 6 Time (mi) EGT 7 EGT 8 EGT 9 EGT 1 EGT 11 EGT 12 Figure 6: Real-time measured EGT curve of gas turbie with ormal coditio The umber of clusters Figure 8: The effect of the umber of clusters o cluster validity idex λ(c) H 2 ( C) 8 Temperature ( C) F1 class F2 class H 1 ( C) F3 class F4 class Figure 9: Graphic clusterig result by usig H 1 ad H Measured EGT Average EGT 27 3 Miimum EGT Maximum EGT Figure 7: Real-time measured EGT profile of gas turbie with ormal coditio umber of clusters is commoly [2, ] i practical process ad is the umber of samples For the 47 traiig samples show i Table 1, the FCM clusterig algorithm ca stepwise iterate from 2 to 21 clusters Figure 8 shows the chagig treds of cluster validity idex λ(c) as a fuctio of the umber of clusters Based o the result show i Figure 8,it is clear that λ(c) icreases at first ad the drops dow with the icreases of the umber of clusters Ad λ(c) ca reach its maximum value whe the umber of clusters is 4, which is i agreemet with the real samples class Therefore, it is cocluded that the FCM clusterig algorithm is suitable for optimal clusterig of gas turbie EGT Cosiderig the high dimesio characteristics of samples, it is difficult to realize graphical aalysis directly I this paper, three temperature uiformity idexes described by Mao [22] are used to further aalyze ad evaluate the cluster results of gas turbie EGT by usig graphic approach Figures 9 11 show the cluster results of gas turbie EGT based o FCM clusterig algorithm It may be clearly observed i Figures 9 11 that all the three temperature uiformity idexes of F1 class are relatively small ( C H 1 3 C, C H 2 3 C, ad C H 3 3 C) This meas that 12 thermocoupletemperaturesesorsgivethealmostsameoutputs,

8 8 Mathematical Problems i Egieerig Table 1: The origial sample set of gas turbie EGT Number T1/ C T2/ C T3/ C T4/ C T5/ C T6/ C T7/ C T8/ C T9/ C T1/ C T11/ C T12/ C H 3 ( C) 8 6 H 3 ( C) H 1 ( C) H 2 ( C) F1 class F2 class F3 class F4 class F1 class F2 class F3 class F4 class Figure 1: Graphic clusterig result by usig H 1 ad H 3 Figure 11: Graphic clusterig result by usig H 2 ad H 3 which belogs to ormal state Compared with F1 class, F2 class has the followig characteristics: 4 C H 1 7 C, 3 C H 2 45 C, ad 3 C H 3 45 C The actual experimetal results show that the fudametal reaso for this pheomeo is turbie blade wear which ca cause a differece of ethalpy drop betwee differet turbie blade passages For the F3 class, all the three temperature uiformity idexes are very large (14 C H 1, 11 C H 2,ad11 C H 3 ) due to the effects of gas turbie load rejectio I additio, a careful ispectio of Figures 9 ad 1 reveals that the temperature uiformity idex H 1 is sigificatly larger tha the other two idexes (5 C H 1, C H 2 3 C, ad C H 3 3 C) It meas that oe of the 12 thermocouple temperature sesors is fault which ca result i a smaller output 53 EGT Fault Diagosis Based o SVM Classificatio Model Accordig to the optimal clusterig results, the fault diagosis traiig sample set icludig fault types ca be established, which is show i Table 2 The o this basis, we ca develop 6 SVM classifiers based o oe-agaistoe method Table 3 shows the fault diagosis performace of SVM multiclassificatio model for the traiig samples From Table 3, it is clear that the fault diagosis accuracy rate

9 Mathematical Problems i Egieerig 9 Table 2: Fault diagosis traiig sample set of EGT Type T1/ C T2/ C T3/ C T4/ C T5/ C T6/ C T7/ C T8/ C T9/ C T1/ C T11/ C T12/ C F F F F Table 3: Fault diagosis performace of SVM multiclassificatio model for traiig samples Type Number of traiig samples Number of accurate diagoses Accuracy rate/% F F F F of traied SVM multiclassificatio model is 1% for the traiig samples, which meas that the SVM multiclassificatio model has bee well traied for fault diagosis of gas turbie EGT ThethetestigsamplesshowiTable 1 areusedto further demostrate the effectiveess of SVM multiclassificatio model Table 4 shows the compariso betwee actual results ad fault diagosis results by usig the well traied SVM multiclassificatio model for testig samples Based o the results show i Table 4, it is demostrated that the well traied SVM multiclassificatio model ca effectively diagose the fault of gas turbie EGT with a 95% accuracy rate for the testig samples Besides, the reaso of misclassificatio is that the sample data is obtaied whe the turbie blade wear or corrosio is ot severe I order to compare with other models, backpropagatio (BP) eural etwork model is also employed to make the same fault diagosis ad the results are also listed i Table 4Thecomparativeaalysis shows that SVM classificatio model ca improve the fault diagosis accuracy of gas turbie EGT sigificatly compared with BP eural etwork model All these idicate that SVM is more suitable for fault diagosis of gas turbie EGT 6 Coclusios Cosiderig the distributio characteristics of gas turbie EGT ad its effect o the health state of gas turbie, a fusio approach based o FCM clusterig algorithm ad SVM classificatio model (FCM-SVM) is proposed ad successfully appliedtoaidustrialgasturbieithispaperithe aalysis preseted i this study, it is demostrated that FCM- SVMbasedapproachcamakefulluseoftheusupervised feature extractio ability of FCM clusterig algorithm ad the sample classificatio geeralizatio properties of SVM multiclassificatio model, which offers a effective way to realize the olie coditio recogitio ad fault diagosis of gas turbie EGT I the cocrete implemetatio process, the itroduced FCM clusterig algorithm is a good alterative to achieve automatic idetificatio of the fault types of gas turbie EGT I other words, it is effective to overcome the ifluece of experiece judgmet o fault types Besides, the itroductio of SVM multiclassificatio model has a great potetial to improve the fault diagosis performace of gas turbie EGT It is worth oticig that the study of this paper is oly focused o researchig the artificial itelligece approach for the coditio recogitio ad fault diagosis of gas turbie EGT but igores the effects of may other parameters such as ilet temperature of gas

10 1 Mathematical Problems i Egieerig Table 4: The compariso results of differet fault diagosis models for testig samples Number SVM1 SVM2 SVM3 SVM4 SVM5 SVM6 SVM model BP model Actual results 1 F1 F1 F1 F2 F2 F3 F1 F1 F1 2 F1 F1 F1 F2 F2 F3 F1 F1 F1 3 F1 F1 F1 F2 F2 F3 F1 F1 F1 4 F1 F1 F1 F2 F2 F3 F1 F2 F1 5 F1 F1 F1 F2 F2 F3 F1 F1 F1 6 F2 F1 F1 F2 F2 F3 F2 F2 F2 7 F1 F1 F1 F2 F2 F3 F1 F1 F2 8 F2 F1 F1 F2 F2 F3 F2 F2 F2 9 F2 F1 F1 F2 F2 F3 F2 F2 F2 1 F2 F1 F1 F2 F2 F3 F2 F1 F2 11 F2 F3 F1 F3 F2 F3 F3 F3 F3 12 F2 F3 F1 F3 F2 F3 F3 F3 F3 13 F2 F3 F1 F3 F2 F3 F3 F3 F3 14 F2 F3 F1 F3 F2 F3 F3 F3 F3 15 F2 F3 F1 F3 F2 F3 F3 F3 F3 16 F1 F3 F4 F3 F4 F4 F4 F4 F4 17 F1 F3 F4 F3 F4 F4 F4 F4 F4 18 F1 F3 F4 F3 F4 F4 F4 F4 F4 19 F1 F3 F4 F3 F4 F4 F4 F4 F4 2 F1 F3 F4 F3 F4 F4 F4 F4 F4 turbie Therefore, more studies ad improvemet about the applicatio of this approach are eeded further Coflict of Iterests The authors declare that there is o coflict of iterests regardig the publicatio of this paper Refereces [1]WYWag,ZQXu,RTag,SYLi,adWWu, Fault detectio ad diagosis for gas turbies based o a kerelized iformatio etropy model, The Scietific World Joural, vol 214,ArticleID617162,13pages,214 [2] A J Volpoi, Gas turbie egie health maagemet: past, preset, ad future treds, Joural of Egieerig for Gas Turbies ad Power,vol136,o5,ArticleID5121,214 [3] T Palmé, F Liard, ad D Therkor, Similarity based modelig for turbie exit temperature spread moitorig o gas turbies, i Proceedigs of the ASME Turbo Expo 213: Turbie Techical Coferece ad Expositio (GT 13), AmericaSocietyof Mechaical Egieers, Sa Atoio, Tex, USA, Jue 213 [4] P E Patrick Hamilto ad D Ha, Exhaust gas temperature capabilitiesowisystem1software, Product Update, vol25, o 1, pp 88 89, 25 [5] YXSog,KXZhag,adYSShi, Researchoaeroegie performace parameters forecast based o multiple liear regressio forecastig method, Joural of Aerospace Power,vol 24, o 2, pp , 29 [6] I Yilmaz, Evaluatio of the relatioship betwee exhaust gas temperature ad operatioal parameters i CFM56-7B egies, Proceedigs of the Istitutio of Mechaical Egieers Part G: Joural of Aerospace Egieerig, vol223, o 4, pp , 29 [7]EMHeadLTSog, AalysisofEGTadmeasuresto icrease the EGT margi, Aviatio Egieerig & Maiteace, vol 6, pp 2 21, 1999 [8] X F Wag ad J M Yag, Aalysis ad treatmet of larger exhaust dispersity fault for PG6551B gas turbie, Gas Turbie Techology,vol17,o2,pp58 61,24 [9] J Che, Y H Wag, ad S L Weg, Applicatio of geeral regressio eural etwork i fault detectio of exhaust temperature sesors o gas turbies, Proceedigs of the Chiese Society of Electrical Egieerig,vol29,o32,pp92 97,29 [1] S Muthurama, J Twiddle, M Sigh, ad N Coolly, Coditio moitorig of SSE gas turbies usig artificial eural etworks, Isight: No-Destructive Testig ad Coditio Moitorig,vol54,o8,pp ,212 [11] J Błachio ad W I Pawlak, Damageability of gas turbie blades-evaluatio of exhaust gas temperature i frot of the turbie usig a o-liear observer, i Advaces i Gas Turbie Techology, chapter 19, pp , 211 [12] Z Korczewski, Exhaust gas temperature measuremets i diagostic examiatio of aval gas turbie egies, Polish Maritime Research,vol18,o2,pp37 43,211 [13] Z Korczewski, Exhaust gas temperature measuremets i diagostic examiatio of aval gas turbie egies part II: usteady processes, Polish Maritime Research, vol 18, o 3, pp 37 42, 211 [14] Z Korczewski, Exhaust gas temperature measuremets i diagostic examiatio of aval gas turbie egies, Polish Maritime Research,vol18,o4,pp49 53,211 [15] ADKeyo,VMCatterso,adSDJMcArthur, Developmet of a itelliget system for detectio of exhaust gas temperature aomalies i gas turbies, Isight: No-Destructive

11 Mathematical Problems i Egieerig 11 Testig ad Coditio Moitorig, vol52,o8,pp , 21 [16] H Xia, D Byrd, S Dekate, ad B Lee, High-desity fiber optical sesor ad istrumetatio for gas turbie operatio coditio moitorig, Joural of Sesors, vol213,articleid 26738, 1 pages, 213 [17] ZYag,BWKLig,adCBigham, Faultdetectioad sigal recostructio for icreasig operatioal availability of idustrial gas turbies, Measuremet, vol 46, o 6, pp , 213 [18] S C Güle, P R Griffi, ad S Paolucci, Real-time o-lie performace diagostics of heavy-duty idustrial gas turbies, Joural of Egieerig for Gas Turbies ad Power,vol124,o 4, pp , 22 [19] D H Seo, T S Roh, ad D W Choi, Defect diagostics of gas turbie egie usig hybrid SVM-ANN with module system i off-desig coditio, Joural of Mechaical Sciece ad Techology,vol23,o3,pp ,29 [2] Y Hao, J G Su, G Q Yag, ad J Bai, The applicatio of support vector machies to gas turbie performace diagosis, Chiese Joural of Aeroautics,vol18,o1,pp15 19,25 [21] W X Wag ad W Z A, Egie vibratio fault diagosis research based o fuzzy clusterig method of gas turbie, Gas Turbie Techology,vol26,o3,pp44 47,213 [22] H J Mao, Aalysis of exhaust temperature moitor ad protectio fuctio for gas turbie, Huadia Techology, vol 31,o8,pp11 15,29 [23] J C Bezdek, Patter Recogitio with Fuzzy Objective Fuctio Algorithms, Kluwer Academic, New York, NY, USA, 1981 [24] S Wikaisuksakul, A multi-objective geetic algorithm with fuzzy c-meas for automatic data clusterig, Applied Soft Computig,vol24,pp ,214 [25] W Cai, S Che, ad D Zhag, Fast ad robust fuzzy c-meas clusterig algorithms icorporatig local iformatio for image segmetatio, Patter Recogitio, vol4,o3,pp , 27 [26] S N Omkar, S Suresh, T R Raghavedra, ad V Mai, Acoustic emissio sigal classificatio usig fuzzy C-meas clusterig, i Proceedigs of the 9th Iteratioal Coferece o Neural Iformatio Processig,vol4,pp ,22 [27] HZhao,PHWag,JQia,ZSu,adXPeg, Modeligfor target-value of boiler moitorig parameters based o fuzzy C- meas clusterig algorithm, Proceedigs of the Chiese Society of Electrical Egieerig,vol31,o32,pp16 22,211 [28] CXu,PZhag,GRe,adJFu, Egiewearfaultdiagosis based o improved semi-supervised fuzzy c-meas clusterig, Joural of Mechaical Egieerig,vol47,o17,pp55 6,211 [29]HBSahu,SSMahapatra,adDCPaigrahi, Fuzzy c-meas clusterig approach for classificatio of Idia coal seams with respect to their spotaeous combustio susceptibility, Fuel Processig Techology,vol14,pp ,212 [3] N R Pal ad J C Bezdek, O cluster validity for the fuzzy c- meas model, IEEE Trasactios o Fuzzy Systems, vol 3, o 3, pp , 1995 [31] V N Vapik, The Nature of Statistical Learig Theory,Spriger, 1995 [32] Q Hu, Z He, Z Zhag, ad Y Zi, Fault diagosis of rotatig machiery based o improved wavelet package trasform ad SVMs esemble, Mechaical Systems ad Sigal Processig,vol 21, o 2, pp , 27 [33] A Widodo ad B S Yag, Support vector machie i machie coditio moitorig ad fault diagosis, Mechaical Systems ad Sigal Processig,vol21,o6,pp ,27 [34] J Yag, Y Zhag, ad Y Zhu, Itelliget fault diagosis of rollig elemet bearig based o SVMs ad fractal dimesio, Mechaical Systems ad Sigal Processig, vol21,o5,pp , 27 [35] C W Hsu ad C J Li, A compariso of methods for multiclass support vector machies, IEEE Trasactios o Neural Networks, vol 13, o 2, pp , 22

12 Advaces i Operatios Research Advaces i Decisio Scieces Joural of Applied Mathematics Algebra Joural of Probability ad Statistics The Scietific World Joural Iteratioal Joural of Differetial Equatios Submit your mauscripts at Iteratioal Joural of Advaces i Combiatorics Mathematical Physics Joural of Complex Aalysis Iteratioal Joural of Mathematics ad Mathematical Scieces Mathematical Problems i Egieerig Joural of Mathematics Discrete Mathematics Joural of Discrete Dyamics i Nature ad Society Joural of Fuctio Spaces Abstract ad Applied Aalysis Iteratioal Joural of Joural of Stochastic Aalysis Optimizatio