CN104155585A - GIS partial discharge type identification method based on GK fuzzy clustering - Google Patents

Abstract

The present invention discloses a GIS partial discharge type identification method based on GK fuzzy clustering. The method is characterized by comprising the steps of S01, constructing a GIS partial discharge gray scale image according to the acquired data; S02, extracting the fractal features of the GIS partial discharge gray scale image, namely a box dimension and an information dimension; S03, processing the fractal feature data further by a GK fuzzy clustering algorithm, isolating a GIS site interference signal; S04, designing a GIS partial discharge mode recognizer based on a least squares support vector machine classification algorithm; S05, identifying the GIS partial discharge type. By a GK fuzzy clustering method, the GIS site interference signal is isolated, thereby improving the accuracy of extracting the fractal features of the partial discharge signal. Meanwhile, by identifying the GIS partial discharge type by the least squares support vector machine classification algorithm, the correctness and rapidity of the discharge type identification are improved.

Description

GIS shelf depreciation kind identification method based on GK fuzzy clustering

Technical field

The present invention relates to a kind of GIS shelf depreciation kind identification method based on GK fuzzy clustering.

Background technology

Gas insulated combined electric appliance equipment (gas insulated switchgear, GIS, also claims SF6 fully closed combined electric unit) has the features such as floor area is little, operational reliability is high, easy to maintenance, in electric system, is widely applied.Yet, GIS because of manufacture or installation process in the fault that causes of defect under residual happen occasionally, its internal fault type is main mainly with insulativity fault, mainly comprises Four types: corona discharge, suspension electric discharge, the electric discharge of insulator surface defect and the electric discharge of free particle.When fault occurs, by defect distortion electric field, produce shelf depreciation.Therefore, by Partial Discharge Detection, judge that the potentiality fault of GIS is to grasping the insulation status of GIS and instructing its maintenance all to have great importance, and GIS local discharge signal is carried out to type identification, become the effective method of judgement GIS inherent vice type characteristic.

The feature extraction of GIS local discharge signal and pattern classification are two key issues in PD Pattern Recognition research.At present, the main method of extracting discharge signal characteristic parameter both at home and abroad comprises: statistical nature, wavelet character, waveform character, moment characteristics, assemblage characteristic and fractal characteristic.Wherein, fractal characteristic is due to its separating capacity and pattern description ability is superior, characteristic parameter is few and can reflect very complicated local discharge signal, and has obtained increasingly extensive application.But in current document, in the extraction of GIS local discharge signal fractal characteristic, do not consider the on-the-spot impact of disturbing fractal characteristic of GIS, if do not consider the on-the-spot impact of disturbing discharge signal, likely can, using undesired signal as discharge signal, cause thus fractal characteristic to extract the accurate problem of owing.

Aspect PD Pattern Recognition, document will have the sorter of the artificial neural network of good adaptive ability, robust performance and non-linear mapping capability as GIS shelf depreciation pattern at present, yet, artificial neural network, owing to having the determining of network structure, local minimum and owing study and cross the problems such as study, may bring certain impact to the correctness of recognition result.

Summary of the invention

For the problems referred to above, the invention provides a kind of GIS shelf depreciation kind identification method based on GK fuzzy clustering, adopt GK fuzzy clustering method, by the on-the-spot undesired signal isolation of GIS, greatly improve the accuracy that local discharge signal fractal characteristic extracts, meanwhile, adopt least square method supporting vector machine sorting algorithm (Least Squares Support Vector Machine, LS-SVM) identification GIS shelf depreciation type, improves correctness and the rapidity of electric discharge type identification.

For realizing above-mentioned technical purpose, reach above-mentioned technique effect, the present invention is achieved through the following technical solutions:

GIS shelf depreciation kind identification method based on GK fuzzy clustering, is characterized in that, comprises the steps:

S01: according to the data construct GIS shelf depreciation gray level image gathering;

S02: extract the fractal characteristic of GIS shelf depreciation gray level image, i.e. box counting dimension and information dimension;

S03: adopt GK fuzzy clustering algorithm further to process fractal characteristic data, the on-the-spot undesired signal of isolation GIS;

S04: based on least square method supporting vector machine design of algorithm GIS PD Pattern Recognition device;

S05: identification GIS shelf depreciation type.

The shelf depreciation sample data of the method based under corona discharge, suspension electric discharge, four kinds of different electric pressures of defect of the electric discharge of insulator surface defect and free particle electric discharge, extract the fractal property of shelf depreciation gray level image as recognition feature amount, consider the on-the-spot impact of disturbing local discharge signal of GIS simultaneously, utilize GK fuzzy clustering algorithm further to process fractal characteristic amount, to extract the fractal characteristic amount after isolation is disturbed.Finally, based on least square method supporting vector machine design of algorithm GIS PD Pattern Recognition device, can effectively identify GIS shelf depreciation type, than Artificial Neural Network, have advantages of that discrimination is high, good stability simultaneously.

The invention has the beneficial effects as follows:

1) the method takes into full account the impact of the on-the-spot undesired signal of GIS on local discharge signal fractal characteristic, adopts GK fuzzy clustering algorithm that undesired signal is isolated, and can greatly improve the accuracy that GIS local discharge signal fractal characteristic extracts.

2) the method can be avoided the determining of network structure in Artificial Neural Network, local minimum and owe study and cross the problems such as study, can effectively improve the correctness that electric discharge type is identified.

3) the GIS shelf depreciation type that the method identifies can be used as the direct basis of judgement GIS inherent vice type, and to grasping the insulation status of GIS and instructing its maintenance all to have great importance.

Accompanying drawing explanation

Fig. 1 is the process flow diagram that the present invention is based on the GIS shelf depreciation kind identification method of GK fuzzy clustering;

Fig. 2 is the fractal characteristic schematic diagram of the untreated four kinds of front GIS shelf depreciation defects of the present invention;

Fig. 3 is the fractal characteristic schematic diagram of four kinds of GIS shelf depreciation defects after GK fuzzy clustering of the present invention is processed;

Fig. 4 is the schematic diagram that GIS of the present invention inside manually arranges four kinds of defects;

Fig. 5 is the schematic diagram of GIS testing ground undesired signal in the embodiment of the present invention;

Fig. 6 is recognition result (σ=9.8463 before LS-SVM parameter optimization of the present invention; γ=94730.521);

Fig. 7 is recognition result (σ=13.5741528 after LS-SVM parameter optimization of the present invention; γ=23478.3341).

Embodiment

Below in conjunction with accompanying drawing and specific embodiment, technical solution of the present invention is described in further detail, so that those skilled in the art can better understand the present invention also, can be implemented, but illustrated embodiment is not as a limitation of the invention.

The invention discloses a kind of gas insulated combined electrical equipment (Gas Insulated Switchgear based on Gustafson-Kessel (GK) fuzzy clustering, GIS) shelf depreciation kind identification method, belongs to fault detect and the state estimation field of high voltage electric equipment.

Fig. 1 is the process flow diagram of the GIS shelf depreciation kind identification method based on GK fuzzy clustering, comprises the steps:

S01: according to the data construct GIS shelf depreciation gray level image gathering;

S02: extract the fractal characteristic of GIS shelf depreciation gray level image, i.e. box counting dimension and information dimension;

S03: adopt GK fuzzy clustering algorithm further to process fractal characteristic data, the on-the-spot undesired signal of isolation GIS;

S04: based on least square method supporting vector machine design of algorithm GIS PD Pattern Recognition device;

S05: identification GIS shelf depreciation type.

Be specially: in GIS local discharge signal feature extraction, first utilize the data construct GIS shelf depreciation gray level image gathering, then extract the fractal characteristic of GIS shelf depreciation gray level image, that is: the information dimension of gray-scale map negative half period the information dimension of positive half cycle the box counting dimension of negative half period and the box counting dimension of positive half cycle and using this characteristic parameter as GIS shelf depreciation type identification.Wherein, concrete methods of realizing can be referring to [Sun Caixin, permitted peak, Tang Ju, Deng. take the GIS Partial Discharge Pattern Recognition Method [J] that box counting dimension and information dimension be recognition feature amount. Proceedings of the CSEE, 2005, 25 (3): 100-104.] institute's extracting method, the result of the gray level image fractal property that the GIS Partial Discharge Data based on gathering builds as shown in Figure 2, wherein " ☆ " represents pollution severity of insulators electric discharge (electric discharge of insulator surface defect), "●" represents metal tip electric discharge (corona discharge), "+" represents suspension electrode electric discharge (suspension electric discharge), " * " represents metal particle electric discharge (free particle electric discharge).

As can be seen from Figure 2, the fractal property of 4 kinds of discharge defects (box counting dimension, information dimension) is all gathered in a certain region, and fractal characteristic has stronger separating capacity.Yet testing ground exists larger interference, inevitably can have undesired signal in the local discharge signal gathering, even stronger undesired signal can covering internal discharge signal.But in local discharge signal fractal characteristic extracts, prior art is not all considered the impact of undesired signal on normal discharge signal, also be in extracted fractal characteristic data as shown in Figure 2, to comprise most probably the fractal characteristic data of undesired signal, if do not processed disturbing a minute graphic data, and the directly input using the fractal characteristic extracting as electric discharge type mode discriminator, this will bring certain impact to the precision of recognition result.Therefore,, for eliminating the impact of disturbing, this method, on the basis of extracted fractal characteristic data, adopts GK fuzzy clustering algorithm further to process fractal characteristic data, to obtain pure electric discharge type recognition feature amount.

GK fuzzy clustering algorithm is to utilize the self-adaptation distance of cluster covariance matrix to measure, the center vector V=(v of its fuzzy clustering ₁, v ₂..., v _c) ^tdegree of membership matrix (or fuzzy partition matrix) U=[μ with data set _ij] _{c * N}by minimizing objective function, try to achieve the number that wherein N is sample, the number that c is cluster, μ _ijfor the degree of membership of data point with respect to cluster centre, and meet:

$\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{\mathrm{\μ}}_{\mathrm{ij}}=1,{\mathrm{\μ}}_{\mathrm{ij}}\∈\left[\mathrm{0,1}\right],1\≤j\≤N---\left(1\right)$

Suppose that data sequence to be identified is that the data sequence that step S02 obtains is X=(x ₁, x ₂..., x _n), it minimize objective function and can be expressed as:

$J(X,V,U)=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\left({\mathrm{\μ}}_{\mathrm{ij}}\right)}^m{D}_{\mathrm{ij}}^2---\left(2\right)$

Wherein, and m ∈ [1, ∞), be a weighted index that characterizes classification fog-level; for data point x in data sequence _jto cluster centre v _idistance, and be a square of inner product apart from norm, be expressed as:

$D_{\mathrm{ij}}^2={\left|\right|x_j-v_i\left|\right|}_{M_i}^2={(x_j-v_i)}^T{M}_i(x_j-v_i)---\left(3\right)$

Wherein, M _ifor following matrix:

$M_i=\det {\left(F_i\right)}^{\frac{1}{n}}{F}_i^{-1}---\left(4\right)$

Wherein, F _iit is the covariance matrix of i cluster centre.

Utilize Lagrangian multiplication, to minimizing target function type (2), be optimized, make objective function obtain minimizing two necessary conditions and be respectively:

${\mathrm{\μ}}_{\mathrm{ij}}=\frac{1}{\underset{k=1}{\overset{c}{\mathrm{\Σ}}}{(D_{\mathrm{ij}}/D_{\mathrm{kj}})}^{2/(m-1)}}---\left(5\right)$

$v_i=\frac{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\left({\mathrm{\μ}}_{\mathrm{ij}}\right)}^m{x}_j}{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\left({\mathrm{\μ}}_{\mathrm{ij}}\right)}^m}---\left(6\right)$

The concrete execution step of GK fuzzy clustering algorithm is as follows:

(I) data-oriented sequence X=(x ₁, x ₂..., x _n), and select clusters number 1<c<N, Weighting exponent m>=1 and termination allowable error ε >0.

(II) arranges fuzzy partition matrix U=[μ _ij] _{c * N}initial value, and make it meet constraint equation (1), set iteration number of times l=0,1 ... (iterations l is a variable, and algorithm repeats how many times), upgrades cluster centre v according to formula (6) _i.

(III) calculates the covariance matrix F of i cluster centre _i:

$F_i=\frac{\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left({\mathrm{\&mu;}}_{\mathrm{ij}}\right)}^m(x_j-v_j){(x_j-v_i)}^T}{\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left({\mathrm{\&mu;}}_{\mathrm{ij}}\right)}^m}---\left(7\right)$

(IV) utilizes the covariance matrix F of step (III) _i, according to formula (4) compute matrix M _i, then according to formula (3), ask for data point x in data sequence _jto cluster centre v _idistance norm

(V) is according to the distance norm of step (IV) gained utilize formula (5) to upgrade fuzzy partition matrix U.

For any one, stop allowable error positive number ε > 0, if meet fuzzy partition matrix || U ^(l+1)-U ^(l)|| < ε, computing stops, otherwise with the fuzzy partition matrix U iteration number of times l=0 after upgrading, 1 ..., according to formula (6), upgrade cluster centre v _i, and perform step successively (III) to (V), until satisfy condition.Namely when not satisfying condition, whole algorithm continues to carry out, i.e. iteration, until satisfy condition.

The fractal characteristic data that extract are further processed, in the situation that taking into full account on-the-spot undesired signal, extract the fractal characteristic data of undesired signal, and utilize above-mentioned GK fuzzy clustering algorithm to carry out cluster to normal shelf depreciation fractal characteristic data, reject the fractal characteristic data of mixing undesired signal therein.

The parameter of GK fuzzy clustering algorithm is selected as follows: c=23, m=1, ε=0.001.Fractal characteristic in Fig. 2 is carried out to fractal characteristic after GK fuzzy clustering processing as shown in Figure 3, and wherein circle represents the fractal characteristic of undesired signal after GK fuzzy clustering is processed.

In step S04, based on least square method supporting vector machine design of algorithm GIS PD Pattern Recognition device, specific implementation process is:

Suppose to have λ sample training set, z wherein _i∈ R ⁿbe i input data; y _i{ 1 ,+1} is i output data to ∈.Constructing a sorter with following form is category support vector machines simulated target:

So that sample z can correctly be classified by f (z).

The Function Estimation problem of least square method supporting vector machine sorting algorithm can be described as the following problem that solves:

$\underset{w,e}{\min }J(w,e)=\frac{1}{2}{w}^{T}w+\mathrm{\&gamma;}\underset{i=1}{\overset{\mathrm{\&lambda;}}{\mathrm{\&Sigma;}}}{e}_i^2---\left(9\right)$

Meet equality constraint

Introduce Lagrangian function:

α wherein _i, (i=1,2 ..., λ) be Lagrange multiplier, represent nuclear space mapping function, e _i∈ R represents error variance, w ∈ R ^nhfor weight vector, b is departure, and γ is error penalty factor, and J (w, e) is the optimal function shown in formula (9).

The necessary condition existing according to extreme value, has following equation to form vertical:

Formula (12) is arranged and can be obtained:

Wherein e=[e ₁, e ₂..., e _λ] ^t, y=[y ₁, y ₂..., y _λ] ^t, α=[α ₁, α ₂..., α _λ] ^t.

According to Mercer condition:

Thus, the system of equations obtaining is relevant with α and b.

Therefore, system of equations (13) can be converted into:

By formula (15), can be obtained:

The sorter that can obtain supporting vector machine model according to the first formula of formula (8), formula (12), formula (14) and formula (16) is:

$f\left(x\right)=\mathrm{sign}[\underset{i=1}{\overset{\mathrm{\&lambda;}}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{y}_i\mathrm{\&psi;}(z,z_i)+b]---\left(17\right)$

ψ (x, x wherein _i) be any symmetric function and meet Mercer condition.Based on RBF type function, have advantages of, the kernel function of selecting is herein following radial basis function:

ψ(z _i,z _j)＝exp(-||z _i-z _j|| ²/(2σ ²)) (18)

In GIS partial discharge test, the artificial defect arranging has 4 kinds, therefore need the shelf depreciation mode type of identification also to have accordingly 4 classes, therefore when solving PD Pattern Recognition, LS-SVM recognizer need to be expanded to multicategory classification problem, yet multicategory classification problem is compared with two class classification problems, need to consider a plurality of samples, decision boundary is complicated, be difficult to training and precision not high.For the identification problem of 4 kinds of electric discharge types, this method adopts the mode of a plurality of 2 classification LS-SVM recognizer combinations to realize when structure LS-SVM pattern recognition classifier device, and this mode has simply, is easy to the advantages such as realization.

Based on above-mentioned analysis, for better identifying 4 kinds of GIS shelf depreciation types, need to design 62 classification LS-SVM recognizers, and it is defined as respectively to SVM ₁₁, SVM ₁₂, SVM ₁₃, SVM ₂₁, SVM ₂₂, SVM ₂₃so, by 6 LS-SVM recognizers, take the output of ballot method can judge the electric discharge type of GIS shelf depreciation sample, its voting rule is as shown in table 1.

Table 1 ballot method discharge mode recognition rule

In addition, LS-SVM recognizer exists kernel function ψ (z _i, z _j) the selection problem of width cs and error penalty factor γ.It is whether suitable that these parameters are selected, and will produce material impact to accuracy of identification.For obtaining optimum parameter, this method (specifically can be referring to document SUYKENS J.A.K by the method for cross validation (CV), VANDEWALLE J.Least squares support vector machine classifiers[J] .Neural Processing letters, 1999,9 (3): 293-30.) to LS-SVM kernel function ψ (z _i, z _j) width cs and error penalty factor γ choose.

For the GIS shelf depreciation kind identification method based on GK fuzzy clustering described in clear explanation this method, the GIS intensive care monitoring system of Jiangsu Province's DianKeYuan of take is experimental study platform, and using the local discharge signal gathering on its platform as fundamental research data.

In the collection of each test figure sample, corona discharge (bus spine and barrel spine), suspension electric discharge, the electric discharge of insulator surface defect and 4 kinds of artificial defects of free particle electric discharge are set in GIS inside respectively, the setting up procedure of each defect is as follows:

1) corona discharge: a long draw point that is 1mm for 53mm, its most advanced and sophisticated place equivalent redius is installed on GIS cavity bus surface and is carried out the most advanced and sophisticated defect of simulated high-pressure conductor metal, equally also same draw point can be fixed on GIS cavity inner wall, as shown in Fig. 4 (a).

2) floating potential defect: by polyethylene insulation screw rod, life is belonged to aluminium flake and be fixed on high-pressure conductor in GIS cavity, both are at a distance of 1.5mm, insulation blanket stud can be for regulating the clearance distance of suspension aluminium flake and guide rod, its length is enough to avoid edge flashing, with this, simulate floating potential defect, as shown in Fig. 4 (b).

3) pollution severity of insulators: coat at insulator inside surface-layer has the silver sulfide powder simulation surface filth of semiconduction character, produces along face electric field and concentrates, and then simulation creeping discharge, as shown in Fig. 4 (c).

4) free particulate electric discharge: due in GIS, electrically conductive microparticle is maximum to the harm of insulation, therefore the harm that it has been generally acknowledged that non-conductive particulate is much smaller, and is that the circular spheroid of stainless steel of 1mm is simulated free particle defects, by sprinkle diameter on cavity as shown in Fig. 4 (d).

Above-mentioned 4 kinds of insulation defects carry out respectively the partial discharge test under 47kV, 50kV, 55kV and 60kV voltage, during test, ultra-high frequency signal adopts the external mode of sensor, by the high-speed oscilloscope of 4GHz/s, sample respectively and store the Partial Discharge in continuous 100 power frequency periods under every kind of defect situation, under each trial voltage, gathering 30 groups of data.

In GIS partial discharge test, on-the-spot electromagnetic interference (EMI) and background interference have larger impact to local discharge signal.Obtaining Accurate local discharge signal from strong interference how is to realize that GIS real-time online detects and the key point of electric discharge type pattern-recognition.For the impact of Analysis interference on shelf depreciation, when carrying out above-mentioned 4 kinds of tests, also on-the-spot undesired signal is gathered, it disturbs as shown in Figure 5, from on-the-spot undesired signal, can find out, there is stronger interference source in testing ground, moreover power frequency correlativity is strong.Therefore, be necessary to take appropriate measures, to suppress the impact of undesired signal on normal discharge signal and electric discharge type pattern-recognition.

The present embodiment selects 30 groups of data that 4 kinds of defect local discharge signals processed through GK fuzzy clustering algorithms as training sample, and every group of sample data comprises: the information dimension of negative half period , positive half cycle information dimension the box counting dimension of negative half period and the box counting dimension of positive half cycle

For checking LS-SVM kernel function ψ (x _i, x _j) width cs and error penalty factor γ choose the impact on accuracy of identification, take pollution severity of insulators defect as example, by the recognition result before parameter optimization with utilize cross validation method to compare to the identification of the electric discharge type after parameter optimization, its comparing result as shown in Figure 6,7.

By comparing result Fig. 6 and Fig. 7, can find out that the accuracy of identification after parameter optimization obviously improves, and illustrates kernel function ψ (x _i, x _j) selection of width cs and error penalty factor γ has a direct impact accuracy of identification.

For further checking the performance of institute's extracting method herein, build respectively the partial discharges fault recognizer based on BP neural network and RBF neural network, and with contrasting based on GK fuzzy clustering recognition device of carrying herein, by square error MSE and accuracy, weigh the quality of recognition performance.

Square error MSE computing formula is as follows:

$E_{\mathrm{MSE}}=\frac{1}{M}\underset{k=1}{\overset{M}{\mathrm{\&Sigma;}}}{(y_k-\hat{y}\left(k\right))}^2---\left(19\right)$

Accuracy is described as:

$\mathrm{Accuracy}=(1-\sqrt{\frac{1}{M}\underset{k=1}{\overset{M}{\mathrm{\&Sigma;}}}{[(y_k-\hat{y}\left(k\right))/y_k]}^2})---\left(20\right)$

Wherein, M is the right number of training data, y _kfor actual value, identification Output rusults for LS-SVM.

The PD Pattern Recognition result of 4 kinds of defects based on BP neural network, RBF neural network and GK fuzzy clustering recognition method is as shown in table 2-5.

Corona discharge recognition result under three kinds of algorithms of table 2

Corona discharge	The inventive method	BP	RBF
				MSE	0.001	2.283	1.896
Accuracy	95.6％	81.6％	87.13％
				Time	0.911	63.36	36.84

Floating potential discharge recognition result under three kinds of algorithms of table 3

Corona discharge	The inventive method	BP	RBF
				MSE	0.00135	2.239	1.811
Accuracy	91.1％	85.6％	88.4％
				Time	0.911	63.36	36.86

Pollution severity of insulators electric discharge recognition result under three kinds of algorithms of table 4

Corona discharge	The inventive method	BP	RBF
				MSE	0.00087	1.233	0.935
Accuracy	97.9％	94.6％	95.4％
				Time	0.911	63.36	36.86

Free particle electric discharge recognition result contrast under three kinds of algorithms of table 5

Corona discharge	The inventive method	BP	RBF
				MSE	0.0013	2.201	1.813
Accuracy	92.6％	84.5％	87.4％
				Time	0.911	63.36	36.86

From the recognition result contrast of table 2-table 5, can find out, adopt the GIS partial discharges fault recognizer of institute of the present invention extracting method all to there is higher discrimination to 4 kinds of discharge faults, especially corona discharge and the electric discharge of free metal particle are had than BP network and the better recognition effect of RBF network, simultaneously the training time of algorithm the shortest (in table, the unit of Time is second), real-time performance is good.

The invention has the beneficial effects as follows:

1) the method takes into full account the impact of the on-the-spot undesired signal of GIS on local discharge signal fractal characteristic, adopts GK fuzzy clustering algorithm that undesired signal is isolated, and can greatly improve the accuracy that GIS local discharge signal fractal characteristic extracts.

2) the method can be avoided the determining of network structure in Artificial Neural Network, local minimum and owe study and cross the problems such as study, can effectively improve the correctness that electric discharge type is identified.

3) the GIS shelf depreciation type that the method identifies can be used as the direct basis of judgement GIS inherent vice type, and to grasping the insulation status of GIS and instructing its maintenance all to have great importance.

These are only the preferred embodiments of the present invention; not thereby limit the scope of the claims of the present invention; every equivalent structure or conversion of equivalent flow process that utilizes instructions of the present invention and accompanying drawing content to do; or be directly or indirectly used in other relevant technical fields, be all in like manner included in scope of patent protection of the present invention.

Claims (6)

1. the GIS shelf depreciation kind identification method based on GK fuzzy clustering, is characterized in that, comprises the steps:

S01: according to the data construct GIS shelf depreciation gray level image gathering;

S02: extract the fractal characteristic of GIS shelf depreciation gray level image, i.e. box counting dimension and information dimension;

S03: adopt GK fuzzy clustering algorithm further to process fractal characteristic data, the on-the-spot undesired signal of isolation GIS;

S04: based on least square method supporting vector machine design of algorithm GIS PD Pattern Recognition device;

S05: identification GIS shelf depreciation type.

2. the GIS shelf depreciation kind identification method based on GK fuzzy clustering according to claim 1, is characterized in that, step S03 specifically comprises the steps:

A1: suppose the data sequence X=(x that step S02 obtains ₁, x ₂..., x _n), select clusters number 1<c<N, Weighting exponent m>=1 and termination allowable error ε >0;

A2: it is degree of membership matrix U=[μ that fuzzy partition matrix is set _ij] _{c * N}initial value, and make it meet constraint equation (1):

$\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{\mathrm{\&mu;}}_{\mathrm{ij}}=1,{\mathrm{\&mu;}}_{\mathrm{ij}}\&Element;\left[\mathrm{0,1}\right],1\&le;j\&le;N---\left(1\right)$

A3: set iteration number of times l=0,1 ..., according to formula (6):

$v_i=\frac{\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left({\mathrm{\&mu;}}_{\mathrm{ij}}\right)}^m{x}_j}{\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left({\mathrm{\&mu;}}_{\mathrm{ij}}\right)}^m}---\left(6\right)$

Upgrade cluster centre v _i;

A4: the covariance matrix F that calculates i cluster centre _i:

$F_i=\frac{\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left({\mathrm{\&mu;}}_{\mathrm{ij}}\right)}^m(x_j-v_j){(x_j-v_i)}^T}{\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left({\mathrm{\&mu;}}_{\mathrm{ij}}\right)}^m}---\left(7\right)$

A5: the covariance matrix F that utilizes step a4 _i, according to formula (4)

$M_i=\det {\left(F_i\right)}^{\frac{1}{n}}{F}_i^{-1}---\left(4\right)$

Compute matrix M _i, then according to formula (3)

$D_{\mathrm{ij}}^2={\left|\right|x_j-v_i\left|\right|}_{M_i}^2={(x_j-v_i)}^T{M}_i(x_j-v_i)---\left(3\right)$

Ask for data point x in data sequence _jto cluster centre v _idistance norm

A6: according to the distance norm of step a5 gained utilize formula (5)

${\mathrm{\&mu;}}_{\mathrm{ij}}=\frac{1}{\underset{k=1}{\overset{c}{\mathrm{\&Sigma;}}}{(D_{\mathrm{ij}}/D_{\mathrm{kj}})}^{2/(m-1)}}---\left(5\right)$

Upgrade fuzzy partition matrix U;

A7: if meet fuzzy partition matrix || U ^(l+1)-U ^(l)|| < ε, computing stops, otherwise increases the iterations of step a3, calculates successively, until satisfy condition according to step a3-a7;

In above formula: the number that N is sample, the number that c is cluster, μ _ijfor the degree of membership of data point with respect to cluster centre, for data point x in data sequence _jto cluster centre v _idistance, F _ibe the covariance matrix of i cluster centre, and m ∈ [1, ∞), be a weighted index that characterizes classification fog-level.

3. the GIS shelf depreciation kind identification method based on GK fuzzy clustering according to claim 1, is characterized in that, in step S04, adopts several 2 classification LS-SVM recognizer, and adopts the type of ballot method judgement GIS local discharge signal.

4. the GIS shelf depreciation kind identification method based on GK fuzzy clustering according to claim 3, is characterized in that, adopts 62 classification LS-SVM recognizers.

5. the GIS shelf depreciation kind identification method based on GK fuzzy clustering according to claim 4, is characterized in that, the kernel function width cs of LS-SVM and error penalty factor parameter γ adopt cross validation method to determine.

6. according to the GIS shelf depreciation kind identification method based on GK fuzzy clustering described in claim 1-5 any one, it is characterized in that, GIS shelf depreciation type comprises corona discharge, suspension electric discharge, the electric discharge of insulator surface defect and the electric discharge of free particle Four types altogether.