Application of Self Organizing Map Approach to Partial Discharge Pattern Recognition of Cast-Resin Current Transformers

Transcription

1 Applcaton of Self Organzng Map Approach to Partal Dscharge Pattern Recognton of Cast-Resn Current Transformers WEN-YEAU CHANG HONG-TZER YANG * Department of Electrcal Engneerng * Department of Electrcal Engneerng St. John s Unversty * Natonal Cheng Kung Unversty No. 499, Sec. 4, Tam Kng Road, Tamsu, Tape 251, Tawan * No. 1, Unversty Road, Tanan Cty 701, Tawan TAIWAN changwy@mal.sju.edu.tw * htyang@mal.ncku.edu.tw Abstract: Partal dscharge (PD) measurement and recognton s a sgnfcant tool for potental falure dagnoss of a power transformer. Ths paper proposes the applcaton of self organzng map (SOM) approach to recognze partal dscharge patterns of cast-resn current transformer (CRCT). The PD patterns are measured by usng a commercal PD detector. A set of features, used as operators, for each PD pattern s extracted through statstcal schemes. The proposed SOM classfer has the advantages of hgh robustness to ambguous patterns and s useful n recognzng the PD patterns of electrcal transformers. To verfy the effectveness of the proposed method, the classfer was verfed on 250 sets of feld-test PD patterns of CRCTs. The test results show that the proposed approach may acheve qute satsfactory recognton of PD patterns. Key-Words: Cast-resn current transformer, Partal dscharge, Pattern recognton, Self organzng map 1 Introducton Partal dscharge measurement and pattern recognton are mportant tools for mprovng the relablty of hgh-voltage nsulaton systems. The pattern recognton of PD ams at dentfyng potental nsulaton defects from the measured data. The potental defects can then be used for estmatng the rsk of nsulaton falure of the hgh-voltage equpment [1]. In the presence of a suffcently strong electrc feld, a sudden local dsplacement of electrons and ons wll lead to a PD f there exsts a defect n an nsulator [2]. A PD event that occurs n the epoxy resn nsulator of hgh-voltage equpment would have harmful effects on nsulaton that may fnally cause servce falure. A defect n hgh-voltage equpment, resultng n PD, wll have a correspondng partcular pattern. Therefore, pattern recognton of PD s sgnfcant for nsulaton condton evaluaton of hgh-voltage equpment. Thanks to physcal understandng of PD made substantal progress n the last decade, t can now be exploted to support nterpretaton of nsulaton defects [1]. Recently, several methods have been employed for the pattern recognton of PD, ncludng neural networks [3], expert systems, fuzzy classfcaton, and wavelet analyss methods. The applcaton of neural networks to pattern recognton and system dentfcaton has become a major trend n the fault dagnoss. Neural networks has been appled for spatal varablty dentfcaton of greenhouse [4], and PD pattern recognton of current transformers [5], and PD montorng technque of gas nsulated substaton [6]. Although the speed of neural networks allows real-tme operaton wth comparable accuracy, the tranng process of multlayer neural networks s often very slow, and the tranng data must be suffcent and compatble. The recognton of PD pattern and the estmaton of nsulaton performance are relatvely complcated, a task whch s often completed by experenced experts. Several expert systems for the dagnostcs of nsulaton systems have been developed [7]. The expert system method acqures the knowledge of human expertse to buld knowledge base. However, t needs to buld and mantan the base wth efforts. The thrd method s the fuzzy clusterng algorthm [8]. The fuzzy c-means clusterng algorthm s one of the most popular fuzzy clusterng algorthms [9]. Fuzzy c-means clusterng algorthm has been appled for pattern recognton for PD of CRCT [10]. Another method s the wavelet analyss method, whch has been used to carry out tme-frequency ISSN: Issue 3, Volume 3, March 2008

2 analyss n fault dagnoss [11] and de-nosng [12]. Wavelet analyss method has also been appled to dentfy the PD characterstcs by decomposton of acoustc emsson sgnals [13] and PD sgnal denosng [14-16]. In ths paper, a novel SOM based pattern recognton technque for the PD dentfcaton of CRCT s proposed wth more effectveness and robustness than the conventonal pattern recognton methods. Ths paper s organzed as follows. Creaton of the PD pattern dataset s descrbed n Secton 2. The development of the algorthm of feature extracton s descrbed n Secton 3. The prncples of SOM and the operaton flowchart of the proposed pattern recognton scheme are gven subsequently. The expermental results and the analyss usng 250 sets of feld-test PD patterns from fve artfcal defect types of CRCTs are presented n Secton 5. From the test results, the effectveness of the proposed scheme to mprove the recognton accuracy has been demonstrated. The paper s concluded n the last Secton. Fg. 1 VH on the hgh-voltage sde of CRCT 2 PD Pattern Dataset Creaton In order to nvestgate the PD features and to verfy the classfcaton capabltes of the SOM for dfferent PD types commonly occurrng n CRCTs, a PD dataset s needed. The PD dataset was collected from laboratory tests on a seres of model CRCTs. The materal and process used to manufacture the model CRCTs were exactly the same as that of makng a feld CRCT. The specfcatons of model CRCTs are shown n Table 1. Fve types of expermental models wth artfcal defects embedded were made to produce fve common PD events n the CRCTs. The fve PD actvtes nclude (a) normal PD actvty (NM) n standard CRCT, (b) nternal cavty dscharge (VH) caused by an ar cavty nsde the epoxy resn nsulator on the hgh-voltage sde, as shown n Fg. 1, (c) nternal cavty dscharge (VL) caused by two cavtes nsde the epoxy resn nsulator on the low-voltage sde, as shown n Fg. 2, (d) nternal fssure dscharge (FH) caused by an ar fssure nsde the epoxy resn nsulator on the hghvoltage sde, as shown n Fg. 3, (e) nternal dscharge (MH) caused by a metal-lne mpurty nsde the epoxy resn nsulator on the hgh-voltage sde, as shown n Fg. 4. The PD events were detected by a PD detectng system set up n our laboratory. The structure of the PD detectng system s shown n Fg. 5. It ncludes a step-up transformer, capactor couplng crcut, PD Fg. 2 VL on the low-voltage sde of CRCT Fg. 3 FH on the hgh-voltage sde of CRCT Fg. 4 MH on the hgh-voltage sde of CRCT ISSN: Issue 3, Volume 3, March 2008

3 Servce Voltage Table 1 Specfcatons of model CRCTs Prmary Current Secondary Current Burden V 20 A 5 A 40VA detector, and the CRCT under test. Through the testng processes, all the data measured were dgtally converted n order to save them n the computer memory. Then, the phase-related dstrbutons of PD derved from the orgnal PD data are obtaned n relaton to the waveform of the feld-test hgh voltage. The hgh voltage n the feld tests s assumed to be held constant and the voltage phase angle s dvded nto a sutable number of wndows (blocks). The PD detector, shown n Fg. 5, s used for acquston of all the ndvdual quas-ntegrated pulses and quantfyng each of these PD pulses by ther dscharge magntude (q), the correspondng phase angle (), at whch PD pulses occur and the number of dscharge (n) over the chosen block. The analyss software (DDX DA3) plots these data as functons of the phase postons [17]. The three phase-related dstrbutons refer to the peak pulse magntude dstrbuton H qmax (), the average pulse magntude dstrbuton H qn (), and the number of pulse dstrbuton H n (). The typcal phase-related dstrbutons of PD patterns for the four knds of defects (VH, VL, FH, and MH) are shown n Fgs. 6 to 9, respectvely. As shown n Fgs. 6 to 9, the PD patterns of deferent defects dsplay dscrmnatve features. descrbed n ths secton; fve statstcal operators are extracted from phase-related dstrbutons. Defntons of the operators are descrbed below. The profle of all these dscrete dstrbuton functons can be put n a general framework,.e., y = f(x ) [17]. The statstcal operators of mean () and varance ( 2 ) can be computed as follows: x f ( x ) (1) f ( x ) Fg. 6 Typcal phase-related dstrbutons of PD for the VH defect Fg. 7 Typcal phase-related dstrbutons of PD for the VL defect 3 Statstcal Feature Extracton Feature extracton s a technque essental n PD pattern recognton to reduce the dmenson of the orgnal data. The features are ntended to denote the characterstcs of dfferent PD statuses [18]. Several statstcal methods of feature extracton are Hgh Voltage Control Plate Step-up Transformer Capactor Couplng Crcut Object under Detecton (CRCT) Fg. 8 Typcal phase-related dstrbutons of PD for the FH defect PD Pattern Analyss Unt Data Acqurement & Analyss Unt PD Detector Personal Computer Fg. 5 System confguraton of the PD detectng system Fg. 9 Typcal phase-related dstrbutons of PD for the MH defect ISSN: Issue 3, Volume 3, March 2008

4 2 2 ( x ) f ( x ) (2) f ( x ) Skewness (S k ) s extracted from each phaserelated dstrbuton of PD to denote the asymmetry of the dstrbuton. It can be represented as: 3 ( x ) p S k (3) 3 Kurtoss (K u ) s extracted to descrbe the sharpness of the dstrbuton as: 4 ( x ) p K u 3 (4) 4 In (1) and (2), x s the statstcal value n the phase wndow, p s the related probablty of appearance. Skewness s a measure of asymmetry degree wth respect to normal dstrbuton. If the dstrbuton s totally symmetrc, then S k =0; f the dstrbuton s asymmetrc to the left of mean, S k >0; and f t s asymmetrc to the rght of mean, S k <0. Kurtoss s an ndcator of sharpness of dstrbuton. If the dstrbuton has the same sharpness as a normal dstrbuton, K u =0; and f t s sharper than normal, K u >0; and f t s flatter than normal, K u <0 [17]. Peaks (P e ) count the number of peaks n the postve or negatve half of a cycle of the dstrbuton. Asymmetry (D a ) represents the asymmetrcal characterstc of partal pulses n both postve and negatve cycles. It s gven by: N q D a (5) N q where N - s the number of PD pulses n the negatve cycle, N + s the number of PD pulses n the postve cycle. q - s the ampltude of the PD pulse at a phase wndow n the negatve cycle, and q + s the ampltude of the PD pulse at a phase wndow n the postve cycle. The cross correlaton factor (C c ) can be expressed as: x y x y / n Cc (6) ( x ( x ) / n) ( y ( y ) / n) where x s the statstcal value n the phase wndow of the postve half cycle, y s the statstcal value n the correspondng wndow of the negatve half cycle, and n s the number of phase wndow per half cycle. Cross correlaton factor ndcates the dfference n the dstrbuton sharps of both postve and negatve half cycles. C c =1 means the sharps are totally symmetrc, C c =0 means sharps are totally asymmetrc. As S k, K u and P e are appled to both postve and negatve cycles of H qmax (), H qn (), and H n (), a total of 18 features can be extracted from a PD pattern. D a and C c are appled to ndcate the dfference or asymmetry n postve and negatve cycles of H qmax (), H qn (), and H n (), and a total of 6 features can be extracted from a PD pattern. Therefore, after the feature extracton procedure, a feature vector of 24 statstcal features s bult for each PD pattern. The typcal statstcal features extracted by the analyss software (DDX DA3) from PD patterns for the four knds of defects (VH, VL, FH, and MH) are shown n Fgs. 10 to 13, respectvely. Fg. 10 Typcal statstcal features of PD for VH Fg. 11 Typcal statstcal features of PD for VL Fg. 12 Typcal statstcal features of PD for FH Fg. 13 Typcal statstcal features of PD for MH ISSN: Issue 3, Volume 3, March 2008

5 The use of statstcal featurng operators for the patterns nstead of the dstrbuton profles can sgnfcantly reduce the dmenson of the database. To a certan degree, they can characterze the PD patterns wth reasonable dscrmnaton [19]. 4 SOM-Based PD Pattern Recognton Method In ths secton, the algorthms of SOM and SOMbased PD pattern recognton scheme are descrbed. The PD recognton through SOM n multdmensonal feature space s also valdated on the bass of the laboratory PD dataset as mentoned above. 4.1 SOM Algorthm The SOM s a typcal unsupervsed neural network, whch maps the multdmensonal space onto a two dmensonal space by preservng the orgnal order. It smulates the self-organzng feature map s functon of the human cerebrum. The SOM s a twolayer neural network that conssts of an nput layer n a lne and an output layer constructed of neurons n a two-dmensonal grd as shown n Fg. 14. The arthmetc of SOM maps random dmenson nput vectors to one or two-dmenson dspersed graphcs and mantan ts orgnal topologes. Wth contnuous compettve learnng, weght vectors would separate from each other n the nput space and form one knd of pattern representaton. So, SOM learns to recognze groups of smlar nput vectors n such a way that neurons whch are physcally close to each other n the neuron layer respond to smlar nput vectors. Dfferent from other clusterng mappng methods for unsupervsed data, mappng relatonshp of SOM can be hghly nonlnear, drectly showng the smlar nput vectors n the source space by ponts close n the two-dmensonal target space [18]. Along wth the smlarty of the nput data, SOM potentally leads to a classfcaton result. It has been appled for PD pattern recognton of turbo-generators [18] and gas nsulated swtchgear [20], and for power system voltage stablty assessment [21]. 4.2 SOM-based PD Pattern Recognzng Procedure The proposed SOM-based PD pattern recognton scheme for CRCTs has been successfully mplemented n the PC-based software (MATLAB). The overall operaton flowchart s shown n Fg. 15. The procedure of the proposed recognton scheme s descrbed brefly as follows. Step1 A grd of SOM output layer neurons s set up wth ntal gven weght vectors. Step2 An nput vector s chosen randomly from the nput space. Step3 A wnnng neuron on the output layer s determned by calculatng the Eucldean dstance between the nput vector and the weght vectors of all neurons n the grd. Step4 The weght vector of the wnner and the weght vectors of ts neghbourng neurons are adjusted accordng to the learnng rate. Step5 Iterate the procedures from Steps 2 to 4 above, tll the tranng process s fnshed. Step6 Save the weght vectors of the traned SOM. Step7 Use the traned SOM to dentfy the defect types of CRCTs. Weght Input Layer Output Layer Fg. 14 System structure of SOM No Start Set up SOM Structure Intal Setup of SOM Neurons Weght Tranng the SOM Tranng Procedure Fnshed? PD Pattern Recognton for CRCT Stop Yes Save the Weght of SOM Fg. 15 Flowchart of the SOM-based recognton scheme ISSN: Issue 3, Volume 3, March 2008

6 5 Expermental Results To verfy the proposed approach, a practcal experment s conducted to demonstrate the effectveness of the PD pattern recognton scheme. The proposed method has been mplemented accordng to the feld-test PD patterns collected from the laboratory. Fve types of expermental models wth artfcal defects are purposely embedded to produce fve common PD events n CRCTs. The proposed method has been mplemented accordng to the feld-test PD patterns collected from our laboratory. The nput data to a PD recognton system are the peak pulse magntude dstrbuton H qmax (), the average pulse magntude dstrbuton H qn (), and the number of pulse dstrbuton H n ().. Assocated wth ther real defect types, there are a total of 250 sample data for dfferent PD events. Each PD event contans 50 patterns of sample data, of whch 30 patterns are tranng data and 20 patterns are testng data. The statstcal feature extracton methods are used to extract 24 statstcal features for each pattern. But, some of the statstcal features are futle for pattern recognton. So, the combnaton of feature vector wll nfluence the accuracy of pattern recognton. In ths paper, the selectng ndex of statstcal features s the standard devaton of each feature calculated from the tranng data. To evaluate the best combnaton of feature vector, we set up three systems of tranng sets. In, the feature vector ncludes 10 features, whch have the lower standard devaton. In, the feature vector ncludes 12 features; and n the feature vector ncludes 14 features. Table 2 shows the combnaton of feature vector for Systems 1 to 3. In, the number of neurons n the nput layer of SOM s desgned to comprse the 10 statstcal featurng operators mentoned above. The numbers of neurons n nput layer of SOM are set to be 12 and 14 for and, respectvely. To evaluate the performance of dfferent structure of SOM, the expermental tests are carred out on 3 types of SOM. The output layer of Type 1 SOM n the three systems s a twodmensonal space comprsng 15 by 15 neurons. The output layer of Type 2 SOM n the three systems s a two-dmensonal space comprsng 17 by 17 neurons. The output layer of Type 3 SOM n the three systems s a two-dmensonal space comprsng 20 by 20 neurons. The structures of 3 types of SOM are shown n Table 3. Table 2 Combnaton of feature vector for 3 systems Dstrbuton Cycle S k K u P e D a C c Postve H qmax () Negatve H qn () H n () Postve Negatve Postve Negatve Dstrbuton Cycle S k K u P e D a C c Postve H qmax () Negatve H qn () H n () Postve Negatve Postve Negatve Dstrbuton Cycle S k K u P e D a C c Postve H qmax () Negatve H qn () H n () Postve Negatve Postve Negatve :selected feature SOM Type 1 Type 2 Type 3 Table 3 The structures of 3 types of SOM System Neurons n Input layer Neurons n Output layer ISSN: Issue 3, Volume 3, March 2008

7 The tranng data consst of 150 patterns, whch were randomly chosen from the 250 sets of sample data. The other 100 patterns were used as the testng data. After the tranng process, the weght vectors of the traned SOM were saved. To verfy the tranng effectveness of the SOM, tranng data are appled to the SOM agan. Tables 4 to 6 show the test results of the tranng data for Types 1 to 3 SOM, respectvely. From Tables 4 to 6, they are shown that the proposed method has 100% accuracy for the 150 tranng feature vectors n three SOMs. Table 7 demonstrates the promsng performance of Type 1 SOM when 300 testng patterns of three systems were tested. As shown n Table 7 among the 100 testng patterns of, there are only two errors of recognton, one for VL, and the other for MH defects. The Table shows that among the 100 testng patterns of, there s only one error of recognton for FH defect. It s shown n the Table that the proposed method has 100% accuracy for the 100 testng patterns of. The test results gve that Type 1 SOM s able to accurately recognze the testng defects for three systems. The number of features n the feature vector wll nfluence the accuracy of pattern recognton. The best combnaton of feature vector for Type 1 SOM s, the feature vector ncludes 14 features. Table 8 demonstrates the promsng performance of Type 2 SOM when 300 testng patterns of three systems were tested. As shown n Table 8 among the 100 testng patterns of, there s only one error of recognton MH defects. The Table shows that among the 100 testng patterns of System 2, there s only one error of recognton for VH defect. The Table also dsplays that proposed method has 100% accuracy for the 100 testng patterns of and the Type 2 SOM s able to accurately recognze the testng defects for three systems. The best combnaton of feature vector for Type 2 SOM s, the feature vector ncludes 14 features. Table 9 demonstrates the promsng performance of Type 3 SOM when 300 testng patterns of three systems were tested. As shown n Table 9 among the 100 testng patterns of, there are only two errors of recognton, one for VH, and the other for MH defects. The Table shows that among the 100 testng patterns of, there s only one error of recognton for FH defect. As shown n the Table, among the 100 testng patterns of, only two errors of recognton exst, one for VH, and the other for MH defects. The best combnaton of feature vector for Type 3 SOM s, the feature vector ncludes 12 features. Table 4 Recognton performance of Type 1 SOM n tranng data (150 patterns) Table 5 Recognton performance of Type 2 SOM n tranng data (150 patterns) ISSN: Issue 3, Volume 3, March 2008

8 Table 6 Recognton performance of Type 3 SOM n tranng data (150 patterns) 6 Conclusons Ths paper has proposed an SOM based pattern recognton technque for PD of CRCTs. The effectveness of the proposed technque has been verfed usng expermental results. It has been shown that through the feature extracton procedure, the extracted statstcal featurng operators can sgnfcantly reduce the sze of the PD pattern database. Also, the SOM based PD pattern recognton scheme s very effectve for clusterng the defects of CRCTs. The expermental results show that the number of features n the feature vector nfluences the accuracy of pattern recognton. To further mprove the recognton accuracy of the proposed approach, the optmal search methods, such as genetc programmng and evolutonary programmng, etc., for the best combnaton selecton of feature vectors can be nvestgated and ntegrated n the proposed SOM based PD pattern recognton for the CRCTs and other hgh-voltage equpment. Besdes, the structures of SOM have also been found to nfluence the accuracy of pattern recognton. To amelorate further the recognton accuracy of the proposed approach, the optmzed structure of the SOM can be studed n the future researches. Table 7 Recognton performance of Type 1 SOM n testng data (100 patterns) VL 95% MH 95% FH 95% References: [1] L. Nemeyer, A Generalzed Approach to Partal Dscharge Modelng, IEEE Transactons on Delectrcs and Electrcal Insulaton, Vol. 2, No. 4, August 1995, pp [2] C. Cachn and H.J. Wesmann, PD Recognton wth Knowledge-Based Preprocessng and Neural Networks, IEEE Transactons on Delectrcs and Electrcal Insulaton, Vol. 2, No. 4, 1995, pp [3] M.M.A. Salama and R. Bartnkas, Determnaton of Neural Network Topology for Partal Dscharge Pulse Pattern Recognton, IEEE Transactons on Neural Networks, Vol. 13, No. 2, 2002, pp [4] M.A. Bussab, J.I. Bernardo, and A. R. Hrakawa, Neural Networks Modelng n Greenhouse wth Spatal Varablty Identfcaton, WSEAS Transactons on Computer Research, Volume 2, Issue 2, February 2007, pp [5] M.H. Wang, Partal Dscharge Pattern Recognton of Current Transformers Usng an ENN, IEEE Transactons on Power Delvery, Vol. 20, No. 3, 2005, pp ISSN: Issue 3, Volume 3, March 2008

9 [6] I. Ok, T. Hada, S. Wakabayash, R. Tsuge, T. Sakakbarb, and H. Muraseg, Development of Partal Dscharge Montorng Technque Usng a Neural Network n a Gas Insulated Substaton, IEEE Transactons on Power Systems, Vol. 12, No. 2, May 1997, pp [7] K. Zals, Applcatons of Expert Systems n Evaluaton of Data from Partal Dscharge Dagnostc Measurement, Proceedngs of the 7th Internatonal Conference on Propertes and Applcatons of Delectrc Materals, 2003, pp [8] M.S. Yang, W.L. Hung, and C.H. Chang, A Penalzed Fuzzy Clusterng Algorthm, WSEAS Transactons on Computer Research, Volume 1, Issue 2, December 2006, pp [9] S.C. Wang and P.H. Huang, Fuzzy C-Means Clusterng for Power System Coherency, Proceedngs of the 2005 IEEE Internatonal Conference on Systems, Man and Cybernetcs, 2005, pp [10] W.Y. Chang and H.T. Yang, Applcaton of Fuzzy C-Means Clusterng Approach to Partal Dscharge Pattern Recognton of Cast-Resn Current Transformers, Proceedngs of the 8th Internatonal Conference on Propertes and Applcaton of Delectrc Materals, ICPADM 2006, Bal, Indonesa, 2006, pp [11] J. P and M.F. Lao, Rollng Bearng Fault Dagnoss wth Wavelet-Based Method, WSEAS Transactons on Computer Research, Volume 2, Issue 1, January 2007, pp [12] Y. Zhang and L. Wu, Research on Tme Seres Modelng by Genetc Programmng and Wavelet De-nosng Performance of the Model, WSEAS Transactons on Computer Research, Volume 2, Issue 1, January 2007, pp [13] Y. Tan, P.L. Lewn, S.J. Sutton, and S.G. Swngler, PD Characterzaton Usng Wavelet Decomposton of Acoustc Emsson Sgnals, Proceedngs of the 2004 Internatonal Conference on Sold Delectrcs, Toulouse, France, July 5-9, [14] Y. Tan, P.L. Lewn, A.E. Daves, S.G. Swngler, S.J. Sutton, and G.H. Hathaway, Comparson of On-Lne PD Detecton Methods for Hgh Voltage Cable Jonts, IEEE Transactons on Delectrcs and Electrcal Insulaton, Vol. 9, No. 3, 2002, pp [15] L. Satsh and B. Nazneen, Wavelet-Based Denosng of Partal Dscharge Sgnals Bured n Excessve Nose and Interference, IEEE Transactons on Delectrcs and Electrcal Insulaton, Vol. 10, No. 2, 2003, pp Table 8 Recognton performance of Type 2 SOM n testng data (100 patterns) MH 95% VH 95% Table 9 Recognton performance of Type 3 SOM n testng data (100 patterns) VH 95% MH 95% FH 95% VH 95% MH 95% ISSN: Issue 3, Volume 3, March 2008

10 [16] P. Wang, P.L. Lewn, Y. Tan, S.J. Sutton, and S.G. Swngler, Applcaton of Wavelet-Based Denosng to Onlne Measurement of Partal Dscharge, Proceedngs of the 2004 Internatonal Conference on Sold Delectrcs, Toulouse, France, July 5-9, [17] N.C. Sahoo and M.M.A. Salama, Trends n Partal Dscharge Pattern Classfcaton: A Survey, IEEE Transactons on Delectrcs and Electrcal Insulaton, Vol. 12, No. 2, 2005, pp [18] Y. Han and Y.H. Song, Usng Improved Selforganzng Map for Partal Dscharge Dagnoss of Large Turbogenerators, IEEE Transactons on Energy Converson, Vol. 18, No. 3, 2003, pp [19] R.E. James and B.T. Phung, Development of Computer-based Measurements and Ther Applcaton to PD Pattern Analyss, IEEE Transactons on Delectrcs and Electrcal Insulaton, Vol. 2, No. 5, 1995, pp [20] T. Ln, R.K. Aggarwal, and C.H. Km, Identfcaton of the Defectve Equpments n GIS Usng the Self Organsng Map, IEE Proc.- Generaton Transmsson Dstrbuton, Vol. 151, No. 5, 2004, pp [21] Y. H. Song, H. B. Wan, and A. T. Johns, Power System Voltage Stablty Assessment Usng a Self-Organzng Neural Network Classfer, Proceedngs of the 4th Internatonal Conference on Power System Control and Management, 1996, pp ISSN: Issue 3, Volume 3, March 2008