Grey Clustering Analysis Based Classifier for Steam Turbine-Generator Fault Diagnosis

Transcription

1 Grey Clusterng Analyss Based Classfer for Steam Turbne-Generator Fault Dagnoss Whe-Mn Ln, Chen-Hsen Wu, Cha-Hung Ln, and Chh-Hsung Su Abstract Ths paper proposes a method for steam turbnegenerator fault dagnoss usng grey clusterng analyss (GCA). Accordng to the feld records, dagnostc nformaton can be provded to montor mechancal condton by the spectrum of the vbraton sgnal. Frequency-based features are computed by fast Fourer transformaton (FFT), the frequency ranges are <.4f, f, 2f, 3f, and >3f. The maxmum and mnmum values of power spectrum ndcate mechancal vbraton fault at a partcular frequency, and frequency patterns are appled to dagnose faults. For numercal tests wth practcal fled records, test results were conducted to show the proposed method demonstrates computatonal effcency and hgh accuracy. Indexng Terms Steam Turbne-Generator, Grey Clusterng Analyss (GCA), Fast Fourer Transformaton (FFT). I. Introducton Turbne generator s a major devce of the thermal plants to convert heat energy of steam nto electrcal energy through mechancal energy. The generator fault not only damages the generator tself, but also causes outages and loss of profts. Wth hgh-temperature, hgh-pressure and factors such as thermal fatgues, many components may go wrong, whch wll not only lead to great economy loss, but sometmes a threat to socal securty. Independent Power Producers (IPPs) need to provde a hgh qualty servce to retan ther customers. So the fault dagnoss becomes an mportant research subject more and more. It s necessary to detect generator faults and take mmedate actons to cut the loss. Wth the complextes, couplng effects and some uncertan factors on turbne generator's structure, the fault dagnoss s dffcult to detect by theoretcal analyss or mathematcal model. Varous artfcal ntellgent (AI) technques have been proposed for fault dagnoss, such as the artfcal neural networks (ANNs) []-[3], fuzzy logc (FL) theorem [4], fuzzy neural networks (FNN) [5] and expert systems [6]. In ths research, the major fault dagnoss scheme s based on the vbraton feature of rotary machnes [7]-[8]. Combnng the AI theorem and vbraton features to mprove the dagnostc accuracy, some theores were proposed ncludng the ANNs based method [3], wavelet based technques [9] and FL based scheme [], and provde promsng results. Wavelet neural network (WNN) has been appled on pattern recognton, such as classfyng voces and Whe-Mn Ln and Chen-Hsen Wu are wth the Department of Electrcal Engneerng, Natonal Sun Yat-Sen Unversty, Kaohsung 8424, Tawan (e-mal:wmln@ee.nsysu.edu.tw). Cha-Hung Ln and Chh-Hsung Su are wth the Department of Electrcal Engneerng, Kao-Yuan Unversty, Lu-Chu Hsang, Kaohsung 82, Tawan (e-mal: eechl53@eductes.edu.tw). mages, wth successful results []. The applcaton of fault dagnoss on turbne generator appeared n [9], where the WNN must consder many parameters' ntalzaton whch s vtal to the success of the complcated network. However, the learnng tme s too long to be appled on lne. The fuzzy based method strongly depends on experenced experts. The nference rules and defuzzfer must be contnuously revsed and mantaned. The ANNs are used the error back propagaton to adjust weghted parameters, and acheve the nonlnear mappng relaton to the desred dagnostc results. However, the local mnmum problem, slower leanng speed and the weghts nterferences among dfferent fault patterns are ts major drawbacks. The grey theory provdes the applcatons of clusterng analyss, relatonal analyss, predcaton, and decson for the grey system [2]. The so-called grey means that system nformaton s ncomplete, unclear, and uncertan. It s a useful method to deal wth the problems of lmted, defcent, and or no rules avalable for data processng. Its analyss makes use of mnor data and does not demand strct statstcal procedures and nference rules. In ths paper, a dagnostc scheme based on GCA s appled to steam turbne-generator fault dagnoss. GCA has a functon of mathematcal operaton for processng numercal data wthout adjustng any parameter. For numercal tests, the proposed method demonstrates computatonal effcency, easy mplementaton, and hgh accuracy for practcal tests. II. Problem Descrpton Power plants convert the fuel energy nto mechancal energy, and then converted nto electrcal energy. Heat engnes convertng steam or gas energy nto mechancal energy, are classfed as steam engne or steam turbne by usng ether pston or rotary desgn. In the thermal power plants, steam turbnes are generally employed. To ncrease the effcency, the temperature and pressure are gradually rased to hgh levels [3]. Water flows through steel ppes desgned to wthstand hgh steam pressures. The chemcal fuel such as coal, gas, or atomzed ol s burned n the furnace. Steam s produced n the boler. Turbne-generator s major components of the thermal plant and conssts of three parts: the turbne, generator and excter. The turbne can also be dvded nto the hgh-pressure (HP), ntermedate-pressure (IP), and low-pressure (LP) parts. All of these sectons are grdled by the bearngs that provde most of the dagnostc nformaton [4]. The generator fault can be classfed three types ncludng electrcal fault, mechancal vbraton fault, and coolng system fault. Frst two are the common ncpent

2 Ol-membrane Oscllaton <.4f f 2f 3f > 3f Imbalance <.4f f 2f 3f No Orderlness > 3f <.4f f 2f 3f > 3f Normal Condton <.4f f 2f 3f > 3f Frequency (f s the rotor frequency) Fg.. Power spectra of typcal fault type n the frequency doman faults n the generator. Electrcal fault has rotor exctaton short crcut, stator wndng grounded fault, and stator wndng short crcut nvolvng three-phase fault or lne-to-lne fault. Rotor exctaton short crcut causes unbalance magnetc pull that acts on the rotor and stator and causes vbraton. When a short crcut occurs at the stator wndng, or between a stator wndng and ground, the protecton should quckly trp the man crcut breaker to dsconnect the machne from the rest of the system and dsconnect the feld wndng from the excter. The protecton s provded by a percentagedfferental relay for wndng short-crcut faults. Mechancal vbraton fault has mbalance, no orderlness, ol-membrane oscllaton, msalgnment, and rotor crack. The montor nstrument such as shaft poston meter, shaft eccentrcty meter, speed meter, dfference expanson meter, and shaft vbraton meter, are used to detect the condton of turbne generator durng operaton. The dsplacement meter and accelerometer are set at the rotor and the bearng poston. The shaft vbraton s regulated wth restrcton to shut down the operaton automatcally usng the trp system. For fault dagnoss, the vbraton waveform analyss, frequency analyss, phase analyss, precesson analyss, and probablty densty analyss are carred out to take all possble measures [5]. Vbraton s an mportant ndcator for montorng mechancal condton and avodng defects development. Vbraton sgnals are collected from data acquston Weght Value F ( f ) F2 ( f ) F F 2 a b c Vbraton Frequency Fg. 2. Boundary values of whten-weght functons system. The tme-doman sgnals are extracted nto frequency-doman features by usng FFT. Wth the frequency-doman features, dagnostc analyss s used to detected faults n characterstc frequences assocated wth varous fault types. Accordng to the feld records, dagnostc nformaton can be provded to montor abnormal condton by the spectrum of the vbraton sgnal. The frequency ranges are <.4f, f, 2f, 3f, and >3f as shown n Fgure. Wthn each frequency band, the upper and lower values of power spectrum ndcate symptom ncludng ol-membrane, mbalance, no orderlness, and normal condton [4]. Therefore, a GCA based classfer s appled to mechancal vbraton fault dagnoss. Usng advantages of lmted, defcent, and no rules avalable for data processng. III. Steam Turbne-Generator Fault Dagnostc Procedure A. Grey Clusterng Analyss (GCA) (b-a)=(c-b)=b 5% F (f )+F 2 (f )= If a steam turbne-generator operates under the normal condton, vbraton condtons are usually small and constant. When a fault occurs or gradually grows, the vbraton sgnals have morphology changes. The vbraton phenomena are collected from data acquston system. The spectrum vares wth dfferent condtons ncludng normal condton, ol-membrane oscllaton, mbalance, and no orderlness, and power spectra are selected n the frequency range <.4f, f, 2f, 3f, and >3f. Varous condtons occupy dfferent lower and upper lmts as shown n Fgure. Dagnostc nformaton can be extracted from the spectrum of the vbraton sgnal. Varous ranges of frequency spectra are used to dentfy the faults, but t s dffcult to descrbe the uncertan boundares n the numercal data such as the lower and upper values. Therefore, GCA s ntroduced to develop the dagnostc procedure. Lke the Fuzzy approach, dfferent ranges of frequency spectra are descrbed by the whten-weght functon. The common ones are trangular and trapezodal functons. The trapezodal functons are used as shown n Fgure 2. Parameter b s the lower/upper value of vbraton frequency f. Parameters a and c are boundary values, whch can be determned [6] (b-a)=(c-b)=b 5% () F (f ) and F 2 (f ) are the whten-weght functons whch typcally fall wthn [,]. Weghted values F (f ) and

3 F 4 F 3 F <.4 f F f F 4 F F 2 F 3 F F F 3 F f F 4 F F 2 F f F 4 F 2 F F > 3 f Fg. 3. Whten-weght functons for vbraton frequences <.4f, f, 2f, 3f, and >3f. F 2 (f ) are complements. Whten-weght functons for pre-selected frequences are defned by lower and upper values as shown n Fgure 3. For unknown nput vector F=[f, f 2, f 3,, f,, f n ], =, 2, 3,, n, n s the number of the vbraton frequency (n=5 n our study). Whten-weghted values for each vbraton frequency (VF) n the dfferent ranges are evaluated by VF = [ F f ) F ( f ) L F ( f ) L F ( f )] ( 2 k K (2) where k=, 2, 3,, K, K s the number of whten-weght functons for each vbraton frequency (K=4 n our study). Then m grey clusters for each vbraton frequency are evaluated by j G = [ g ] = VFW Preprocessng Pattern Recognton Vbraton Sgnal Data Acquston System Feature Extracton (FFT Method) Frequency Spectrum Feature Selecton ( <.4f, f, 2f, 3f, >3f) Fault Dagnoss (GCA) Produced New Boundary Values New Practcal Record? Yes Whten-Weght Functons Tunng Fg. 4. The flow chart of fault dagnostc procedure G = [ F f ) F ( f ) L F ( f ) L F ( f )] ( 2 k K w w2 L w j L w m w2 w22 L w2 j L w2m M M O M O M wk wk 2 L wkj L wkm M M O M O M wk wk2 L wkj L wkm (3) K g j = Fk ( f ) wkj, j=, 2, 3,, m (4) k= where g j s the grey cluster, m s the number of fault types (m=4 n our study). For m classes classfcaton, matrx W could be expressed as weghtng factor w kj [,]. If F k (f ) belongs to class j, where the weghted factor w kj equals one, and the rest of the factors are zero, t can be represented by, w kj =, k Class k Class Compute the fnal grey grade FG j by j, j=, 2, 3,, m (5) j n FG j = g j, j=, 2, 3,, m (6) n = σ = [ FG FG2 FG3 L FG j L FGm ] (7) Then fnd the maxmum value, FG max =max{fg, FG 2, FG 3,, FG j,, FG m }. The maxmum value FG max ndcates the fault type. In ths paper, the major classes nclude: FG :Ol-membrane Oscllaton, FG 2 :Imbalance, FG 3 :No Orderlness, and FG 4 :Normal Condton.

4 Table. Tested data of steam-turbne generator sets Generator Number <.4f f 2f 3f >3f Note: f (Hz) s the frequency of the generator rotor. B. Fault Dagnostc Procedure The GCA based classfer s developed to work wth the exstng dagnoss equpment. The FDP conssts of two stages: (a) preprocessng stage and (b) pattern recognton stage, as shown n Fgure 4. Vbraton sgnals from data acquston system are collected and transferred to computer. Features n the frequency doman are extracted by FFT method. The ncpent faults can be revealed by vbraton frequency spectrum. Abnormal vbraton nformaton can be detected at pattern recognton stage by contnuously montorng. If new practcal records are produced new boundary values, lower/upper boundares of whten-weght functons can be adjusted contnually. The decson boundares can be easly modfed by equaton (). Its analyss makes use of mnor data, and also has a functon of mathematcal operaton for processng numercal data wthout any nference rule. IV. Test Results The proposed fault dagnostc procedure was desgned on a PC Pentum-IV 2.4GHz wth 48MB RAM and Matlab software. Usng the practcal records, the proposed procedure provdes hghly confdent results for judgng the faults. To demonstrate the effectveness of the proposed procedure, twenty recorded data from steam-turbne generator sets are tested as shown n Table [4]. There are 6 recorders of ol-membrane Oscllaton (No.~No.6), 6 recorders of mbalance (No.7~No.2), 6 recorders of no orderlness (No.3~No.8), and 2 recorders of normal condton (No.9~No.2). Two study cases are chosen for demonstraton, as detaled below. A. Study Case Each pre-selected frequency (<.4f, f, 2f, 3f, and >3f) has a specfc range wth upper and lower values. The nput data ncludes the fve ampltude values of the Ol-membrane Oscllaton Test Record <.4f f 2f 3f > 3f Frequency (Hz) Fg. 5. Frequency spectrum for the test records of olmembrane oscllaton vbraton spectrum. For example, Fgure 5 shows the frequency spectrum for the test records of Ol-Membrane Oscllaton. Usng the record of generator number, the nput data are as follow <.4f: 3.35, f: 2.5, 2f:.94, 3f: 2.3, >3f: The dagnostc procedures are: Step : Vbraton spectrum: F=[ ] Step 2: Whten-weghted values: VF = [....] VF 2 = [....] VF = [. 3...] VF 4 = [ ] VF = [ ] Step 3: Grey cluster for each vbraton frequency: G = [....] G 2 = [....] G 3 = [....] G 4 = [ ] G = [ ] Step 4: Fnal grey grades: σ = [ FG FG2 FG3 FG4] = [ ] The fault s judged by max(σ)=fg =.958. Maxmum value ndcates the Ol-Membrane Oscllaton fault. For generator number, the maxmum grey grade ndcates the fault, and confrms that the fault s ol-membrane oscllaton. CGA takes.875 seconds CPU Tme) to detect 2 tests, and overall tests results are shown n Fgure 6. Maxmum value ndcates the whte or lght gray and corresponds to the fault. The GCA based classfer promses the results wth % accuracy. B. Study Case 2 The same twenty recorded data are also used to test for four fault types. In practcal measurement, sgnals may be dsturbed by nose such as quantfcaton error. Addng 3% to +3% noses to the orgnal patterns,

5 Fault Type FG FG 2 FG 3 FG Generator Number Note: FG :Ol-membrane Oscllaton, FG 2 :Imbalance, FG 3 :No Orderlness, and FG 4 :Normal Condton. Fg. 6. The colormap for study Case Grey Grade G 5 =[....] Step 4: Fnal grey grades: σ = [ FG FG2 FG3 FG4] = [ ] The fault s judged by max(σ)=fg 2 =.8. Maxmum value ndcates the Imbalance fault. Fault Type FG FG 2 FG 3 Grey Grade Imbalance Test Record Generator Number FG 4 Note: FG :Ol-membrane Oscllaton, FG 2 :Imbalance, FG 3 :No Orderlness, and FG 4 :Normal Condton. Fg. 8. The colormap for study Case 2 (added +% nose) <.4f f 2f 3f > 3f Frequency (Hz) Fg. 7. Frequency spectrum for the test records of Imbalance test results show that the proposed method can work wth nosy background. For example, Fgure 7 shows the frequency spectrum for the test records of Imbalance. Usng the record of generator number 7, the nput data are added +% nose to the vbraton frequences. The dagnostc procedures are: Step : Vbraton spectrum: F=[ ] Step 2: Whten-weghted values: VF = [....] VF 2 = [....] VF 3 = [....] VF 4 = [....] VF 5 = [....] Step 3: Grey cluster for each vbraton frequency: G = [....] G 2 = [....] G 3 = [....] G 4 =[....] The maxmum grey grade.8 ndcates the fault, and confrms that the fault s mbalance. Overall test results for twenty steam-turbne generator sets are shown n Fgure 8. Ths study case confrms that the overall accuraces are also % wth +% nose. V. Dscussons and Concluson The fault dagnostc procedure based on GCA has been presented to mechancal vbraton fault dagnoss for steam turbne-generators. The proposed method could avod the determnaton of the lngustc varables, membershp functons, nference rules, and parameters assgnment, and s easy to mplement n the portable devce. Snce the mathematcal operaton for numercal data wthout adjustng any parameter, GCA requres less computaton tme, and t s a useful method to deal wth the problems of lmted, defcent, and no rules avalable for data processng. Its analyss makes use of mnor data to construct the whten-weght functons and does not demand strct statstcal procedures. Random addng noses to each selected vbraton frequency wth 5% to +5% noses (symbol + means postve nose and symbol means negatve nose). Wth 2 generators, the test shows that proposed method has % detecton accuracy under nosy background wth 5% to +2% noses, and the accuraces decay as postve/negatve noses ncrease (Accuracy < 8%). To develop an assstance tool, the proposed method can be

6 further constructed n the on-lne model, and be ntegrated nto the montorng nstrument. Reference [] W.M. Ln, C.D. Yang, C.H. Ln, and M.T. Tsay, A Fault Classfcaton Method by RBF Neural Network wth OLS Learnng Procedure, IEEE Transactons on Power Delvery, Vol.6, No.4, 2, pp [2] Y. Zhang, X. Dng, Y. Lu, and P. J. Grffn, An Artfcal Neural Network Approach to Transformer Fault Dagnoss, IEEE Trans. on Power Delvery, Vol., No.4, October 996, pp [3] LI, H., Sun, C. X., Lao, R.J., Chen, W.G., Hu, X.S., Improved BP Algorthm n Vbraton Falure Dagnoss of Steam Turbne-Generator Set, Journal of Chongqng Unversty, Vol.22 No.5, I 999. [4] Syed Mofzul Islam, Tony Wu, and Gerard Ledwch, A Novel Fuzzy Logc Approach to Transformer Fault Dagnoss, IEEE Trans. on Delectrcs and Electrcal Insulaton, Vol.7, No.2, Aprl 2, pp [5] Wu,C.Z., Yan, H., and Ma, J.F., Method Research of Nose Dagnoss Based on Fuzzy Neural Network, Proc. of 4th Int. Conf. on Sgnal Processng, Bejng, Chna, 2 6 Oct., 998, pp [6] Heung-Jae Lee, Bok-Shn Ahn, and Young-Moon Park, A Fault Dagnoss Expert System for Dstrbuton Substatons, IEEE Trans. on Power System, Vol.5, No., January 2, pp [7] Umemara, S., et al., A New Health Montorng System for Rotatng Machnery, Mtsybsh Juke Gho, 8(6), pp, 7-78, 98. [8] Mtcheell, J. S., Machnery Analyss and Montorng, Pennwell Publshng Company, 98. [9] LI, H., Sun, C. X., Hu, X.S., Yue, G., Tang, N.F., and Wang, K., Study of Method on Adaptve Wavelets for Vbraton Fault Dagnoss of Steam Turbne-Generator Set, Journal of Electrcal Engneerng (Chna), Vol. I5, No.3, 2. [] LI, H., Sun, C. X., Hu, X.S., YUE, G., and Wang, K., The Fuzzy Inputtng and Outputtng Method n Vbraton Fault Dagnoss of Steam Turbne-Generator Set, Journal of Chongqng Unversty,Vol. 22, No.6, 999. [] Szu H. and Bran T., Neural Network Adaptve Wavelets for Sgnal Representaton and Classfcaton, Optcal Engneerng, pp ,992. [2] Ju-Long Deng, Introducton to Grey System Theory, Journal of Grey System, Vol., No., 989, pp.-24. [3] Za A. Yamayee and Juan L. Bala, Electromechancal Energy Devces and Power Systems, John Wley & Sons, Inc., 994. [4] M.-H. Wang, Applcaton of Extenson Theory to Vbraton Fault Dagnoss of Generator sets, IEE Proc.- Gener. Transm. Dstrb., Vol. 5, No. 4, July 24, pp [5] Masatake Kawada, Koj Yamada, Katsuya Yamashta, and Katsuo Isaka, Fundamental Study on Vbraton Dagnoss for Turbne Generators Usng Wavelet Transform, Power System Conference and Exposton., vol..3, Oct. 24, pp [6] Wen-Yng Chang, Herng-Der Ln, Yaw-Wu Chen, Chen-Hua Tuan, Kuang-Chao Wu, and Jeeng-Horng La, Expert System for Transformer faults Dagnoss, Monthly Journal of Tapower s Engneerng, Vol. 55, July 994, pp Bographes Whe-Mn Ln (M 87) was born n 954. He receved hs B.S. degree n electrcal engneerng from the Natonal Chao-Tung Unversty, Hsn-Chu, Tawan, and the M.S. degree n electrcal engneerng from the Unversty of Connectcut, Storrs, and the Ph.D. degree n electrcal engneerng from the Unversty of Texas at Arlngton n 985. Currently, he s wth Natonal Sun Yat-Sen Unversty, Tawan, where has been snce 99. He was wth Chung-Hwa Insttute for Economc Research, Tawan, as a vstng researcher after hs graduaton. He joned Control Data Crop., Mnneapols, Mnnesota, n 986 and worked wth Control Data Asa, Tape, Tawan n 989. Dr. Ln s a member of Tau Beta P. Hs man nterests are GIS, Dstrbuton system, SCADA, and automatc control system. Chen-Hsen Wu was born n 97. He receved the B.S. degree n electrcal engneerng from the Natonal Tawan Unversty of Scence and Technology, Tape, Tawan, n 994 and M.S. degree n electrcal engneerng from the Natonal Sun Yat-Sen Unversty, Kaohsung, Tawan, n 2. He s currently pursung hs Ph.D. degree at Natonal Sun Yat-Sen Unversty. Hs research nterests nclude energy management systems, dstrbuton automatc system, and harmonc analyss. Cha-Hung Ln was born n 974. He receved the B.S. degree n electrcal engneerng from the Tatung Insttute of Technology, Tape, Tawan, n 998, the M.S. degree n electrcal engneerng from the Natonal Sun Yat-Sen Unversty, Kaohsung, Tawan, n 2, and the Ph.D. degree n electrcal engneerng from Natonal Sun Yat-Sen Unversty n 24. Currently, he s the assstant professor of department of electrcal engneerng, Kao-Yuan Unversty, Lu-Chu Hsang, Kaohsung, Tawan, where has been snce 24. Hs research nterests nclude fault dagnoss n power system, artfcal ntellgent applcaton, and harmonc analyss. Currently, hs area of nterest s bomedcal sgnal process and analyss. Chh-Hsung Su was born n 972. He receved the B.S. degree n electrcal engneerng from the Kao-Yuan Insttute of Technology, Kaohsung, Tawan. Currently, he s wth the Department of Electrcal Engneerng, Kao-Yuan Unversty, Lu-Chu Hsang, Kaohsung, Tawan. Hs research nterests nclude artfcal ntellgent applcaton and electrcal mechansm control.