Advanced
Application of SA-SVM Incremental Algorithm in GIS PD Pattern Recognition
Application of SA-SVM Incremental Algorithm in GIS PD Pattern Recognition
Journal of Electrical Engineering and Technology. 2016. Jan, 11(1): 192-199
Copyright © 2016, The Korean Institute of Electrical Engineers
  • Received : November 06, 2013
  • Accepted : May 26, 2015
  • Published : January 01, 2016
Download
PDF
e-PUB
PubReader
PPT
Export by style
Article
Author
Metrics
Cited by
TagCloud
About the Authors
Ju Tang
Corresponding Author: State Key Laboratory of Power Transmission Equipment & System Security and New Technology, Chongqing University, China. (cqtangju@cqu.edu.cn)
Ran Zhuo
State Key Laboratory of Power Transmission Equipment & System Security and New Technology, Chongqing University, China. (zrhungry@yahoo.com.cn)
DiBo Wang
State Key Laboratory of Power Transmission Equipment & System Security and New Technology, Chongqing University, China. (zrhungry@yahoo.com.cn)
JianRong Wu
State Key Laboratory of Power Transmission Equipment & System Security and New Technology, Chongqing University, China. (zrhungry@yahoo.com.cn)
XiaoXing Zhang
State Key Laboratory of Power Transmission Equipment & System Security and New Technology, Chongqing University, China. (zrhungry@yahoo.com.cn)

Abstract
With changes in insulated defects, the environment, and so on, new partial discharge (PD) data are highly different from the original samples. It leads to a decrease in on-line recognition rate. The UHF signal and pulse current signal of four kinds of typical artificial defect models in gas insulated switchgear (GIS) are obtained simultaneously by experiment. The relationship map of ultra-high frequency (UHF) cumulative energy and its corresponding apparent discharge of four kinds of typical artificial defect models are plotted. UHF cumulative energy and its corresponding apparent discharge are used as inputs. The support vector machine (SVM) incremental method is constructed. Examples show that the PD SVM incremental method based on simulated annealing (SA) effectively speeds up the data update rate and improves the adaptability of the classifier compared with the original method, in that the total sample is constituted by the old and new data. The PD SVM incremental method is a better pattern recognition technology for PD on-line monitoring.
Keywords
1. Introduction
Recognition of partial discharge (PD) pattern in gas insulated switchgear (GIS) is vital for on-line monitoring and faults diagnosis of electric equipment [1] . PD pattern recognition is divided into three parts, namely, data acquisition, characteristics extraction, and pattern classifycation. First, data are acquired on-line or off-line by the sensors and data acquisition system. Quality of data is associated with the character of sensors, accuracy of the acquisition system, interference from the environment, and so on. Subsequently, data are processed by artificial intelligence (AI) to obtain the features extracted and to finish the pattern recognition. With different AI, different process methods possess a distinct recognition ability. Thus, improving the recognition ability cannot be achieved by merely ameliorating the hardware. Determining the new characteristic quantities and better classifiers are extremely important as well.
There are two main kinds of characteristic quantities for PD recognition. The first is based on the statistical regularities of the discharge on power cycles, such as figures in two or three dimensions [2 - 5] , while the second is based on the real-time characteristics of PD signals [6] . The former reflects better on PD’s physical features. Its characteristic quantities, discharge phase, discharge quantity, and discharge repetition rate have a tight relationship with PD’s severity and an extremely high data dimension, which becomes its main shortage [3] . Considering its high data dimension, reducing dimensions is necessary. However, balancing between redundancy and abandonment of information is difficult.
The second kind emphasizes on the detailed features of PD signals from different defects. Moreover, it can provide a high recognition rate. However, a PD pulse is in nanosecond scale, which provides a strict requirement to the data acquisition system and does no good for the extension of this method. Similarly, both kinds require quantities of data that lead to difficulty in data storage forming, and training the classifier. For the reasons stated above, searching for new characteristic quantities will remain to be the hot point in this research field.
In [7] , certain regularities between the apparent charge and the amplitude of ultra-high frequency (UHF) signal from a PD are found to exist. The existence of the regularity is proved in theory [8] , which provides us with a good foundation. At present, the main algorithm for pattern recognition includes neural network [4 , 9] , cluster analysis [10] , and Fisher kernel [3] , to name a few. Though each offers its own advantages and has been applied to a different extent, the classifier can hardly adapt when the data updated are still the common question. A classifier using support vector machine (SVM) [15] can save memory space and simultaneously retain the information of the original classifier when data are updated. Furthermore, given that the constructed classifier depends only on the support vectors in the samples, abandonment of non-support vectors does not affect the classifier. Considering adaptability to small sample, nonlinear, high-dimensional samples, there have been some instances in which SVM has been applied in PD pattern recognition [11 , 12] . Nevertheless, SVM’s advantage of fast adaptation has not been reflected in PD on-line monitoring. The study on parametric optimization for SVM requires further work as well.
In this article, we aimed at decreasing recognition rate caused by changes in insulated defects, the environment, and so on.(here we mean that as the change of types of insulated detect, the environment and other factors, the recognition rate would always decrease, thus we put our eyes on this phenomenon to reduce the decreasing level.) making the new PD data extremely different from the original sample. Four typical PD models were established to generate PD signals. The PD was detected using the UHF antenna and pulse current method simultaneously. Numerous experiments were conducted on these four models. Acquired data were divided into the following three groups: sample set T1 for training, T2 for testing, and incremental data set T3 . A classifier for the PD pattern recognition was constructed using the SVM incremental method. In this process, training sample set T1 was used to construct the original classifier. Testing sample set T2 was used to prove its availability. Moreover, incremental data set T3 was used to prove the fast adaptability of the classifier.
2. PD Experiment
- 2.1 Experiment set
To perform the experiments, we placed four types of defect models [13] in a GIS simulator: N type for needle defect, P type for particle defect, M type for metal defect, and G type for gap defect. The experiment circuit is shown in Fig. 1 . The power source is a testing transformer (YDYW-25/100) free of corona, the sensor is a homemade microstrip antenna [14] , while the oscilloscope is a Tectronix7104 (bandwidth 1GHz, maximum sample rate 20G/s). Environmental temperature was 10 ℃. The pulse current method is adopted according to the IEC60270 standard.
PPT Slide
Lager Image
Experiment circuit
- 2.2 Calibration for pulse current
The pulse voltage amplitude obtained using IEC60270 is proportionate to the apparent discharge. However, the proportionality factor is affected by the characteristics of elements in the circuit and the performance of the instrument. The proportionality factor Kc can only be obtained through calibration. Subsequently, the apparent discharge Q can be calculated. To achieve this without the power source, an adjustable PD calibrator was connected to the ends of the model. The PD calibrator produced a pulse with known discharge quantity; it caused a certain voltage pulse between the ends of the model, through which we obtained a point. Furthermore, by adjusting the PD calibrator to produce different pulses with different discharge quantities, a series of points was obtained. Accordingly, the relationship between the voltage on the detection impendence U (mV) and the apparent discharge Q (pC) was described. Fig. 2. presents their tight linear relation.
PPT Slide
Lager Image
Calibration curve of pulse current method
- 2.3 Experiment procedure
The main experiment steps are described as follows:
1) Source voltage was slowly increased until breakdown occurred in the test sample. The inceptive discharge voltage and breakdown voltage were recorded. To generate a stable PD without breakdown, three test points were properly selected and marked as U 1 , U 2 , and U 3 between these two voltages.
2) The oscilloscope was set at two-channel mode to collect the PD UHF signal and pulse current signal simultaneously.
3) Five hundred sets of data were collected, respectively, on all three voltages and four defect models.
- 2.4 PD signal collected
PD signals on all four defect models were collected in the experiments. Considering the limit of length, however, we only provided the signals collected on needle defect ( Fig. 3. ). In the figure, the upper part represents the UHF signal in 50 mV/scale while the lower part pertains to the pulse current signal in 200 mV/scale. Time was plotted in 1 us/scale.
PPT Slide
Lager Image
Actually measured signals on needle defect using UHF and IEC60270
3. Characteristic Quantities
- 3.1 UHF cummulative energyW
The PD UHF signals collected in the experiments are voltage signals. Voltage points are read by oscilloscope and UHF cumulative energy is calculated using the following equation [8] :
PPT Slide
Lager Image
In the experiments, sample rate was 2.5 GS/s; data were recorded for 10 us. Thus, the sample interval Δ t is 400 ps; RL is a matching impendence of 50 Ω ; N is the total number of the sample points, which is 2.5GS/s 10 × μs = 25000 ; and Un stands for the value of the nth sample point.
- 3.2 Apparent dischargeQ
Apparent discharge can be acquired by the following steps:
1) Voltage pulse signals obtained from IEC60270 are used. Amplitude of the signal H is found using a computer program or an oscilloscope.
2) Proportionality factor Kc is determined using the method mentioned in Section 2.2.
3) Apparent discharge Q is calculated as shown below:
PPT Slide
Lager Image
- 3.3 Selection of characteristic quantities
After data were collected, the UHF cumulative energy W and the apparent discharge Q in those four typical defect models was plotted (shown in Fig. 4. ) using Equations (1) and (2).
PPT Slide
Lager Image
Relationship between cumulative energy of UHF signal and apparent discharge
In the figure, changing of the UHF cumulative energies with apparent discharges is observed to obey quadratic functions. Moreover, given that obvious differences exist among the four defects, we decide to use W and Q as the input characteristic quantities to achieve pattern recognition.
4. Support Vector Machine
- 4.1 Basic principle
To describe it simply, a support vector machine performs a nonlinear transformation, ϕ , on the input vector to map it into a considerably higher-dimensional space F , and conducts a linear separation in this space [15] . For a given sample set
PPT Slide
Lager Image
, xi Rn is the input, N is the total number of samples, and yi ∈{±1} is the class number.
In this article, UHF cumulative energy W and apparent discharge Q are used as inputs. Thus, xi = {( Wi,Qi )] and yi are the class numbers of the four typical defects. If data are separable in the space F , then SVM will construct a hyperplane in that space to separate the data. The hyperplane can be described as f ( x ) = ( w ϕ ( x )) + b , where w is the normal vector of the hyperplane and b is the offset. w and b are optimized by the following expressions:
PPT Slide
Lager Image
PPT Slide
Lager Image
Here, ξ pertains to the relaxation factor, which is regarded as the distance between the training samples and the separate hyperplane. C > 0 refers to the penalty factor, whose discrimination function is as follows:
PPT Slide
Lager Image
In the equation, SV stands for the support vector while k ( xj,x ) is the kernel function. In our research, Gaussian radial basis function (RBF) kernel, expressed in the following equation, was selected.
PPT Slide
Lager Image
- 4.2 Dual-tree classification for insulation defects
Given that GIS PD pattern recognition is a multiclass question while SVM is a binary classifier, algorithms such as one-to-one, one-to-many, or dual-tree should be used to solve the multi-classification question [16] . Among these algorithms, a dual-tree requires the least SVMs. Thus, we chose dual-tree to extend SVM to a multi-classifier and perform the recognition. Its basic thinking is shown in Fig. 5.
PPT Slide
Lager Image
Dual-tree SVM of insulation defects
In a training sample set with C classes, C −1 SVM should be trained. The first SVM, signed as SVM 1 , yields positive results to Class 1 and negative results to Classes 2,3⋯ C . In the same vein, SVM i provides positive results to Class i and negative results to Classes i +1, i + 2⋯ C . Until SVM C−1 is achieved, all classes are separated
In our research, PD defects include the N, P, M, and G types. Thus, the constructed dual-tree will be similar to Fig. 5(b) , where the order of N, P, M, and G is variable. SVM N uses UHF cumulative energy W and apparent discharge Q as inputs. Furthermore, it is trained to separate type N from the others. In the same way, we can determine the usage of SVM P and SVM M . By these three SVMs, all four kinds of defects can be separated.
- 4.3 Parameter optimization in SVM by SA
Apart from kernel function and input {( W ,Q )} , two parameters must be determined before the construction of SVM: penalty factor C and parameter of the kernel function σ . Determination of these two parameters has a serious effect on SVM’s recognition performance. In this article, C and σ are optimized by simulated annealing (SA) [17] algorithm.
- 4.3.1 SA settings
The object of optimization is to achieve the maximum recognition rate, according to which the following shows several settings on SA optimization.
1) Objective function
PPT Slide
Lager Image
In the function, Ttrain is the recognition rate of the training sample set that is equal to ( Ncorrect / Ntotal ) ×100% , where Ncorrect is the number of correct determination among the training samples and Ntotal is the total number of the training samples. To increase sensitivity in parameter searching, parameters C and σ are described as C = 2 p′ , σ = 2 p′′ .
2) Renewal function for parameter p
PPT Slide
Lager Image
In the function,
PPT Slide
Lager Image
is a set searching step length, where α is set manually according to the amount of data, t is the temperature parameter [18] in the original SA physical model, and k refers to the iterations. In our problem, we call t as the step length controller. Given that the original solutions [ p′, p″ ] 0 are random, a larger step length is desired at the beginning. Thus, the initial step length controller t 0 was set to 100 according to the experience.
3) Renewal function for step length controller t
According to the needs of accuracy and searching speed, and considering the procedure of the PD SVM incremental method, two kinds of renewal functions for step length controller were designed: slow and fast. The slow kind is aimed at extending the searching range, and it is appropriate for the generation of initial SVMs from the original samples. Meanwhile, the fast kind aims at detail searching, and it is appropriate for the construction of incremental method based on the original SVM method and associates with the incremental data.
Slow renewal function for step length controller:
PPT Slide
Lager Image
Fast renewal function for step length controller:
PPT Slide
Lager Image
Here, k ≥ 3 .
4) Acceptance criteria
PPT Slide
Lager Image
Here, delE = F (
PPT Slide
Lager Image
)− F ( p* ), where p* is the present solution and
PPT Slide
Lager Image
is a new feasible solution, which resulted from equation 8. If delE < 0 , which indicates the objective function value of
PPT Slide
Lager Image
, is smaller than the objective function value of p* , then the new feasible solution
PPT Slide
Lager Image
is accepted. h = 1/(1 + exp( delE / max( t ))) , h > rand indicate that when the step controller is a certain t , and within a set time range, delE < 0 has never occurred. Subsequently, a random number rand ( rand has been defined as a random number) will be generated. Moreover, if h > rand , the new feasible solution
PPT Slide
Lager Image
is accepted.
5) Loop termination criteria
PPT Slide
Lager Image
When any of these situations occurs, the loop will be terminated.
- 4.3.2 Main steps of SA algorithm
1) The initial step length controller t 0 is determined; let p 0 be the initial solution and p * be the present solution; the objective function F ( p * ) is calculated. Subsequently, let iterations k =0.
2) A new feasible solution,
PPT Slide
Lager Image
by p * , is generated using the renewal function for parameter p , and F (
PPT Slide
Lager Image
) is calculated.
3) Whether
PPT Slide
Lager Image
is accepted by the acceptance criteria is decided. If accepted, p* is replaced by
PPT Slide
Lager Image
, described as p * =
PPT Slide
Lager Image
. Let k = k +1 , and move forward. If it does not keep p * , then we revert back to Step 2.
4) The step length controller t is updated using the renewal function. The loop termination criteria are checked if they are satisfied. If they are satisfied, the loop and output are terminated in the optimal solution pbest = p* . If they fail to be satisfied, we revert back to Step 2.
5. PD SVM Incremental Recognition Method
With changes in insulated defects, the environment, and so on, the new PD data are highly different from the original samples. Subsequently, they lead to a decrease in on-line recognition rate. In this article, a PD SVM incremental recognition method [19] is presented. It can create accurate prediction toward new data samples while keeping the pre-existing information intact. It can obtain superior adaptability while saving a sizeable storage space.
Taking those four typical defects as examples the dual-tree SVM classifier, denoted by Γ , constructed with the training sample set T1 , includes SVM N , SVM P , and SVM M . The incremental data sets are T3i = { xi | i = 1,2,..., I } and T3'i = { yi | i = 1, 2,..., I } = {+1,−1} , where xi Rn are inputs, yi are class numbers, and l is the total number of the samples. The concrete steps are described as follows:
1) The support vector set T1sv , which is used to construct the classifier Γ is extracted from T1 : T 1 SV = {[ SV N , SV P , SV M ]} . SV N , SV P ,,and SV M are support vectors of SVM N , SVM P and SVM M .
2) The combination of T1 SV and incremental data set T3 is taken as the new incremental sample set, expressed as follows: T 1 SV + T 3 . Let B be the non-support vectors in T 1. B is used to store the non-support vectors.
3) Let T 1 = T 1 SV + T 3 and train the new classifier Γ * . Extract the support vector set T 1 SV and non-support vector set B * in the new T 1.
4) Let B = B + B * , which is used to back up samples that have no contribution to the classifier. Steps 1, 2, and 3 are continually repeated to learn the new samples.
5) Iterate for many times until the classifier converges to Γ * .
Equations should be placed at the center of the line and provided consecutively with equation numbers in parentheses flushed to the right margin, as in (1).
PPT Slide
Lager Image
Be sure that the symbols used in your equation have been defined before the equation appears or immediately following.
It is strongly encouraged that the authors may use SI units only.
6. Instance Simulation
The 1989 PD samples obtained in our experiments were divided into three groups. Training sample set T 1 was used to construct the original classifier; testing sample set T 2 was used to prove its availability; and incremental data set T 3, collected after a slight change in defect models, was used to prove the fast adaptability of the classifier. Samples in each group are shown in Table 1.
Properties of the sample
PPT Slide
Lager Image
Properties of the sample
In the process of constructing the original classifier, T 1 was used as a training sample set. The initial values of C and σ in SA algorithm were generated randomly. The parameter in searching step length α was set to 10 to improve the ability to jump out from a local optimum and to lengthen the step. The parameters in SVMs were updated using Equation 9, the slow renewal function. After 55 iterations, the acquired SVM parameters are C = 79.26 and σ = 0.863 , and the number of support vectors is NSV = 465. The constructed classifier was denoted as Γ , whose support vector set is T 1 SV . The changing of the objective function value is shown in Fig. 6. Using classifier Γ to recognize T 2, we obtained 88.67% correct solutions, as demonstrated in Table 2.
PPT Slide
Lager Image
The value of optimizing objective function
Recognition results of incremental PD SVM
PPT Slide
Lager Image
PS: C and are the result of slow optimization
In actual situation, the collected PD samples may change considerably because of the changing defects or the environment. To test the fast adaptability of the classifier, several slight changes were made on the defect models before the incremental data set T 3 was collected.
After using the classifier Γ – trained by T 1 only – to recognize T 3, we obtained the result showing that the accuracy rate is extremely low. To improve this, the training sample set was replaced by combining T 1 and T 3. The initial values of C and σ in SA algorithm are still generated randomly. Moreover, the searching step length parameter α was set to 10 to lengthen the step and to improve the ability to jump out from a local optimum. The parameters in SVMs were updated using Equation 9, the slow renewal function. After being fully trained, that is, being subjected to 64 iterations lasting 15 hours, the objective function value reached stability. The obtained parameters are C = 97.68 and σ = 0.832 , and the number of support vectors is NSV = 672 . Testing it on T 2, the obtained recognition rate is 89.63%, as seen in Table 2. Noticeably, once there is a new sample set, the entire processes of construction and optimization must be repeated; thus, a large amount of time is wasted.
Adopting the PD SVM incremental method mentioned in the previous section, we replaced the training sample set as the combination of support vector set T 1 SV and T 3. C = 79.26 and σ = 0.863 were used as the initial values of C and σ . Given that these values had been optimized by T 1, performing a global optimization became unnecessary. However, a local particular searching was conducted using the fast update function, Equation 10. With α = 200 , after 46 iterations, the objective function value became stable. The changing of the objective function value is shown in Fig. 6.
In the results, C = 96.85 , σ = 0.837 , and NSV = 673 . The process lasted six hours. The acquired classifier is denoted by Γ * , whose support vector set is
PPT Slide
Lager Image
Testing on T 2, we obtained the recognition rate of 89.63%, as seen in Table 2. Assuming that there is a new sample set T 4, the combination of T 4 and
PPT Slide
Lager Image
will be the new incremental sample set. Using C = 96.85 and σ = 0.837 as the initial values, and repeating the former process, the classifier Γ * can be updated within a short time while information on the original classifier can be kept intact.
By comparing the PD SVM incremental method based on SA optimization and the method using the combination of training sample set and incremental data set as the new training samples set, it can be observed that though both have the same recognition rate and similar results of NSV , C , and σ , the incremental method decreases the samples to a large extent. Thus, training time is significantly reduced. As a result, the memory space could be saved a lot, and the higher order of magnitude of data could be proceed. Moreover, the incremental method keeps the information on the original classifier intact, and it can be updated quickly with the entry of new data.
Finally, a comparison was made among the PD SVM incremental method based on SA optimization, the neutral network based on fuzzy C-means clustering algorithm (FCM) and K-means clustering algorithm. The results we used the combination of support vector set T 1 SV and T 3 as the training sample set are shown in Table 3. The SA-SVM method has obvious advantages in all three items, which are training accuracy, recognition accuracy, and training time. More important, both FCM algorithm and K-means clustering algorithm are difficult to use in dealing with the incremental data.
Comparison of recognition results
PPT Slide
Lager Image
Comparison of recognition results
5. Conclusion
1) PD UHF signals and pulse current signals of all four defect models were collected simultaneously. The relationship between the UHF cumulative energy W and the apparent discharge Q was mapped out. The method of using W and Q as the characteristic quantity to recognize the four types of defects was proposed.
2) PD SVM incremental method based on SA optimization was constructed. It solves the problem of the drop in the PD recognition rate, which results from the relative difference between the newly acquired PD data and the original data. The said difference between the data sets results from the changing insulation defects and the environment.
3) The slow and fast update functions of the step length controller, used in the optimization of SVM parameters based on SA, were provided. When parameters were unknown, the slow one should be adopted. Meanwhile, when parameters were known, the fast one was more appropriate. Thus, the classifier can be updated quickly and adaptability of the classifier is improved.
4) By comparison, the PD SVM incremental method was determined to have the capacity to save storage space and shorten training time to a large extent. Furthermore, the PD SVM incremental method was shown to have the capacity to deal with incremental data while keeping the original information intact, which both FCM method and K-means algorithm cannot achieve at the same time.
Even though a conclusion may review the main results or contributions of the paper, do not duplicate the abstract or the introduction. For a conclusion, you might elaborate on the importance of the work or suggest the potential applications and extensions.
BIO
Ju Tang was born in Pengxi, Si Chuan Province, China, in 1960. He received a Bachelor’s degree from Xi’an Jiaotong University, and Master’s and Doctoral degrees from Chongqing University. Dr. Tang is a professor in the Key Laboratory of High Voltage Engineering, Chongqing University. Currently, as a chief scientist he is involved in high voltage electric equipment insulation on-line monitoring and fault diagnosis.
Ran Zhuo was born in Guiyang, Gui Zhou Province, China, in 1986. He received his Bachelor’s degree in Electrical Engineering at Chongqing University, China. He is now working toward his Doctor’s degree in the Key Laboratory of High Voltage Engineering, Chongqing University. He is involved in high voltage electric equipment insulation online monitoring and fault diagnosis.
Di-bo Wang was born in Liuzhou, Guangxi Province, China, in 1988. He received his Bachelor’s degree in Electrical Engineering at Chongqing University, China. He is now working toward his Master’s degree in the Key Laboratory of High Voltage Engineering, Chongqing University. He is involved in high voltage electric equipment insulation online monitoring and fault diagnosis.
Jian-rong Wu was born in Pixian, Sichuan Province, China, in 1986. She received her Bachelor’s degree in Electrical Engineering at Chongqing University, China. She is now working toward her Master’s degree in the Key Laboratory of High Voltage Engineering, Chongqing University. She is involved in high voltage electric equipment insulation online monitoring and fault diagnosis.
Xiao-xing Zhang was born in Qianjiang City, Hubei Province, China, in 1972. He received Bachelor’s and Master’s degrees at the Hubei Institute of Technology and Doctoral degree at Chongqing University. Dr. Zhang is a professor in the College of Electric Engineering, Chongqing University. He is involved in high voltage electric equipment insulation on-line monitoring and fault diagnosis.
References
Kranz H. -G. 2002 “Fundamentals in computer aided PD processing, PD pattern recognition and automated diagnosis in GIS” IEEE Transactions on Dielectrics and Electrical Insulation 7 12 - 20
Li Jian , Sun Caixin 2006 “Partial Discharge Image Recognition Using a New Group of Features” IEEE Transactions on Dielectrics and Electrical Insulation 13 (6) 1245 - 1253    DOI : 10.1109/TDEI.2006.258196
Zhang Xiaoxing , Ren Jiangbo , Tan Ju , Sun Caixin 2009 “Kernel Statistical Uncorrelated Optimum Discriminant Vectors Algorithm for GIS PD Recognition” IEEE Transactions on Dielectrics and Electrical Insulation 16 (1) 206 - 213    DOI : 10.1109/TDEI.2009.4784569
Salama M. M. A. , Bartnikas R. 2002 “Determination of Neural-Network Topology for Partial Discharge Pulse Pattern Recognition” IEEE Transactions on Networks 13 (2) 446 - 456    DOI : 10.1109/72.991430
Sahoo N. C. , Salama M. M. A. 2005 “Trends in Partial Discharge Pattern Classification: A Survey” IEEE Transactions on Dielectrics and Electrical Insulation 12 (2) 248 - 264    DOI : 10.1109/TDEI.2005.1430395
Xie Yanbin , Tang Ju , Zhou Qian 2009 “Feature extraction and recognition of UHF partial discharge signals in GIS based on dual-tree complex wavelet transform” European Transactions on Electrical Power 20 639 - 649
Lijun Jin , Mingrui Zhang , Weidong Liu 2005 “Examination and Research on the Partial Discharge Fault Diagnostics in Gas Insulated Switchgear” Transactions of China Electrotechnical Society 20 (11) 88 - 92
TANG Ju , WU Jianrong , ZHUO Ran , XIE Yanbin 2010 “Relationship Between VHF Signals and Discharge Magnitude of Partial Discharge from Needle Plate Electrode” High Volt age Engineering 36 (5) 1083 - 1089
Kranz H. G. 1993 “Diagnosis of partial discharge signals using neural networks and minimum distance classification” IEEE Transactions on Electrical Insulation 28 1016 - 1024    DOI : 10.1109/14.249375
Contin A. , Cavallini A. , Montanari G. C. , Pasini G. , Puletti F. 2002 “Digital detection and fuzzy classification of partial discharge signals” IEEE Trans. On Dielectrics and Electrical Insulation 9 (2) 335 - 348    DOI : 10.1109/TDEI.2002.1007695
Sharkawy R. M. , Mangoubi R.S. , Abdel Galil T. K. , Salama M. M. A. , Bartnikas R 2007 “SVM Classification of Contaminating Particles in Liquid Dielectrics Using Higher Order Statistics of Electrical and Acoustic PD Measurements” IEEE Transactions on Dielectrics and Electrical Insulation 14 (3) 669 - 678    DOI : 10.1109/TDEI.2007.369530
Hao L. , Lewin P. L. 2010 “Partial Discharge Source Discrimination using a Support Vector Machine” IEEE Transactions on Dielectrics and Electrical Insulation 17 (1) 189 - 197    DOI : 10.1109/TDEI.2010.5412017
Tang Ju , Zhou Qian , Tang Ming , Xie Yanbin 2007 “Study on Mathematical Model for VHF Partial Discharge of Typical Insulated Defects in GIS” IEEE Transactions on Dielectrics and Electrical Insulation 14 (1) 30 - 38    DOI : 10.1109/TDEI.2007.302869
Zhang Xiaoxing , Tang Ju , Peng Wenxiong , Meng Yanhui , Sun Caixin 2006 “Study on the Outer UHF Microstrip Patch Antenna for Partial Discharge Detection in GIS” Chinese Journal of Scientific Instrument 27 (12) 1595 - 1599
Vapnik V. N. 1995 The Nature of Statistical Learning Theory Springer-Verlag Inc.
Cristianini Nello , Shawe-Taylor John , Holloway Royal 2000 An Introduction to Support Vector Machines and Other Kernel-based Learning Methods Cambridge University Press, Inc. Cambridge
Kirkpatrick Scott 1984 “Optimization by Simulated Annealing: Quantitative Studies” Journal of Statistical Physics 34 (975-986)
Cao Longhan , Zhou Shanquan , Li Rui , Wu Fan , Liu Tao 2008 “Application of optimizing the parameters of SVM using genetic simulated annealing algorithm” WCICA 2008. 7th World Congress on Intelligent Control and Automation 5381 - 5385
BAI Dong ying , WANG Xiao dan , MA Fei 2007 “ SVMbased Incremental Learning Method and its Application” Aeronautical Computing Technique 37 (4) 23 - 26
Vardi Joseph , Avi Itzhak Benjamin 1981 MIT 75 - 94