CN104808107A - XLPE cable partial discharge defect type identification method - Google Patents

Abstract

The invention discloses an XLPE cable partial discharge defect type identification method. The method comprises the following steps: obtaining the de-noised XLPE cable ultrahigh frequency PRPS graph; extracting and computing four dimension statistical characteristic parameters and six non-dimension characteristic parameters of the PRPS graph; inputting the characteristic parameters to SOM neural network model; determining the win nerve cell, continuously learning through updating the weight vector of the win nerve cell by the SOM neural network, until the input sample corresponding to the output layer win nerve cell is stable; and outputting the type identification result of the SOM neural network. The XLPE cable partial discharge defect type identification method is capable of overcoming the defects of the traditional method that the convergence speed is slow and the judgment accuracy is bad, satisfying the visual requirement of the fault analysis, and improving the intelligent level of the XLPE cable partial discharge detection system, has the characters of rapid detection speed, simple process and high diagnosis accuracy, and is good for accurately evaluating the running state of the XLPE cable by the maintenance worker.

Description

A kind of XLPE cable shelf depreciation default kind identification method

Technical field

The present invention relates to XLPE cable insulation defect detection technique field, particularly relate to a kind of XLPE cable shelf depreciation default kind identification method.

Background technology

In the net power transmission system of city, crosslinked polyethylene (XLPE) power cable has become the main flow equipment of electric power conveying.The part XLPE cable of current operation is about to reach tenure of use, and its Insulation Problems is day by day remarkable.The severe local defect of laying environment and cable itself also substantially reduces cable life simultaneously, and cause insulation ag(e)ing serious, line fault takes place frequently.As the Important Parameters of reflection cable machinery state of insulation, shelf depreciation (hereinafter referred to as PD) and its insulation status have close ties.Effective pattern-recognition is carried out to PD signal, defect type character and feature in XLPE cable can be understood and grasped exactly, to judge its insulating reliability further.Therefore, the research of XLPE cable PD Pattern Recognition, for the safe and reliable operation ensureing XLPE cable, grasps the insulation status of XLPE cable and instructs the service work of XLPE cable to have very important meaning.

Summary of the invention

The object of the invention is for overcoming the deficiencies in the prior art, provide a kind of XLPE cable shelf depreciation default kind identification method, the method is based on wavelet analysis and SOM neural network, realize the Intelligent Recognition of XLPE cable insulation defect, improve the intelligent level of XLPE cable partial discharge detecting system.

For achieving the above object, the present invention adopts following technical proposals:

A kind of XLPE cable shelf depreciation default kind identification method, comprises the steps:

(1) the XLPE cable shelf depreciation PRPS collection of illustrative plates gathered, and the process of 2-d wavelet threshold deniosing is carried out to described PRPS collection of illustrative plates, obtain the XLPE cable superfrequency PRPS collection of illustrative plates after denoising;

(2) extract and calculate the characteristic parameter of PRPS collection of illustrative plates, comprising peak-to-peak value X _pk, rectification average X _av, standard deviation X _sdwith root mean square X _rmfour have dimension statistical nature parameter and kurtosis X _ku, measure of skewness X _sk, peak factor X _c, pulse factor X _i, shape factor X _swith nargin factor X _lsix dimensionless characteristic parameters;

(3) above-mentioned characteristic parameter is input to SOM neural network model; By calculating the distance between input vector and each neuronic weight vectors, determine triumph neuron, SOM neural network is by upgrading neuronic weight vector unceasing study of winning, until input amendment corresponding to output layer triumph neuron is stablized;

(4) export the type identification result of SOM neural network, namely export corresponding XLPE cable shelf depreciation defect type.

The concrete grammar of described step (1) is:

1) data each in PRPS collection of illustrative plates are normalized:

$x_{i,j}^{\′}=\frac{x_{i,j}-\min \left\{x_{i,j}\right\}}{\max \left\{x_{i,j}\right\}-\min \left\{x_{i,j}\right\}}$

Wherein, x _i,jbe the i-th row, peak value of pulse, x ' are put in the original office of j row _i,jfor peak value of pulse, min{x are put in the office after normalization _i,jput peak value, the max{x of minimum pulse in pulse for all original offices _i,jput the peak value of maximum impulse in pulse for all original offices;

2) PRPS collection of illustrative plates is represented with the form of gray level image, by wavelet transformation, 2-d wavelet decomposition is carried out to gray level image;

3) hard-threshold process is carried out to the wavelet low frequency signal after decomposition:

$W_{\mathrm{\δ}}=\left\{\begin{array}{c}W,\left|W\right|>\mathrm{\δ}\\ 0,\left|W\right|\≤\mathrm{\δ}\end{array}\right.$

Wherein, W _δfor wavelet coefficient values, δ is threshold value;

4) low frequency component of high fdrequency component and hard-threshold process is carried out small echo inversion, obtain the denoising gray-scale map after reconstructing, and renormalization process is carried out to described denoising gray-scale map, obtain XLPE cable superfrequency PRPS collection of illustrative plates after final denoising.

In described step (2), four have the computing method of dimension statistical nature parameter to be specially:

$X_{\mathrm{pk}}=\max \left\{x_{i,j}\right\}-\min \left\{x_{i,j}\right\};X_{\mathrm{av}}=\frac{1}{\mathrm{MN}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}\left|x_{\mathrm{ij}}\right|;$

$X_{\mathrm{sd}}=\sqrt{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{(x_{i,j}-\stackrel{\‾}{x})}^2};X_{\mathrm{rm}}=\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{x}_{i,j}^2$

Wherein, X _pkfor the peak-to-peak value of pulse signal is put in office, X _avfor rectification average, X _sdfor standard deviation, X _rmfor mean square value, x _i,jbe the i-th row, peak value of pulse is put in the original office of j row, and M is number of phases, and N is power frequency period number, min{x _i,jput the peak value of minimum pulse in pulse, max{x for all original offices _i,jput the peak value of maximum impulse in pulse for all original offices, x is the average that pulse signal peak value is put in all original offices.

In described step (2), the computing method of six dimensionless statistical nature parameters are specially:

$X_{\mathrm{ku}}=\frac{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N}{\mathrm{\Σ}}}\underset{j=0}{\overset{M}{\mathrm{\Σ}}}{\left(\right|x_{i,j}|-\stackrel{\‾}{x})}^4}{{X_{\mathrm{rm}}}^2};X_{\mathrm{sk}}=\frac{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{(x_{i,j}-\stackrel{\‾}{x})}^3}{{\left(\sqrt{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{(x_{i,j}-\stackrel{\‾}{x})}^2}\right)}^3};$

$X_c=\frac{\max \left\{x_{i,j}\right\}}{\sqrt{X_{\mathrm{rm}}}};X_i=\frac{\max \left\{x_{i,j}\right\}}{X_{\mathrm{av}}};X_s=\frac{\sqrt{X_{\mathrm{rm}}}}{X_{\mathrm{av}}};X_l=\frac{\max \left\{x_{i,j}\right\}}{{\left[\frac{1}{N}\underset{i=0}{\overset{N}{\mathrm{\Σ}}}\underset{j=0}{\overset{M}{\mathrm{\Σ}}}\sqrt{\left|x_{i,j}\right|}\right]}^2};$

Wherein, X _kufor the kurtosis of pulse signal is put in office, X _skfor measure of skewness, X _cfor peak factor, X _ifor the pulse factor, X _sfor shape factor, X _lfor the nargin factor, X _pkfor peak-to-peak value, X _avfor rectification average, X _sdfor standard deviation, X _rmfor mean square value, x _i,jbe the i-th row, peak value of pulse is put in the original office of j row, and M is number of phases, and N is power frequency period number, min{x _i,jput the peak value of minimum pulse in pulse, max{x for all original offices _i,jput the peak value of minimum pulse in pulse for all original offices, for the average of pulse signal peak value is put in the original office of entirety.

The concrete grammar of described step (3) is:

A) initialization SOM neural network model:

The neuronal quantity of input layer and competition layer in SOM neural network model is set to n, m respectively, determines topology of networks; The iterations and the sample size that arrange SOM network are T and K respectively; By neuronic weights W each in SOM network _j,ibe initialized as the random number that [0,1] is interval;

B) by input vector, i.e. 10 statistical nature parameter vectors, bring SOM neural network model into:

Wherein, the input vector that kth is secondary is X _k=[x _k1, x _k2..., x _k10], X _kvalue random selecting or from training centralized cycle choose;

C) all input vector X are calculated _k=[x _k1, x _k2..., x _k10] and each neuronic weight vectors W _j=[W _j, W _j..., W _j10] between distance d _jk;

D) triumph neuron is determined:

By weight vectors and input vector X _knearest neuron is as the triumph neuron C of SOM neural network model;

E) weight vector is upgraded:

Once triumph neuron C is located, SOM neural network model is constantly learnt by upgrading neuronic weight vector of winning;

Renewal learning rate formula is

η(t)＝η ₀(1-t/T)

Wherein, η (t) is renewal learning rate, η ₀for renewal learning rate initial value, t is time constant, and T is phase constant;

F) select new input vector, step (c) to (e) is carried out in circulation, until input amendment corresponding to output layer triumph neuron is stablized;

H) t=t+1 is set, repeats step 2 and carry out T iteration.

Described step c) in calculate the formula of the spacing of all input vectors and each neuronic weight vectors as follows:

$d_{\mathrm{jk}}=\left|\right|X_k-W_j\left|\right|=\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(X_{\mathrm{ki}}-W_{\mathrm{ji}})}^2,}j=1,2,3\·\·\·m$

Wherein, X _kfor kth time input vector, W _jfor a jth neuronic weight vectors, X _kifor i-th input vector that kth is secondary, W _jifor the weight vectors of corresponding i-th input vector of a jth neuron.

Described step e) in the update method of weight vectors be:

W _j(t+1)＝W _j(t)+η(t)h _jc(t)[X _k(t)-W _j(t)]

Wherein, W _j(t+1) be weight vector after renewal rewards theory, W _jt () is the weight vector before renewal rewards theory, η (t) is learning rate, h _ict () is neighborhood function.

Described neighborhood function h _ict the computing formula of () is:

$h_{\mathrm{jc}}=\exp (-\frac{\left|\right|r_c-r_j\left|\right|}{{2\mathrm{\σ}}^2\left(t\right)})$

Wherein, r _cand r _jbe the point of neuron in two-dimensional planar array of triumph neuron and other competition respectively, σ (t) is the radius of neighbourhood, and along with continuous renewal, the radius of neighbourhood is finally reduced to 0.

Described step e) in, learning rate needs to be reduced to guarantee convergence.

The concrete grammar exporting the type identification result of SOM neural network in described step (4) is:

By arranging corresponding neuron as triumph neuron sample, neuronic activation value of winning is 1, and other values are 0; Using the classification results of triumph neuron sample as input vector, realize the classification feature of SOM neural network.

The invention has the beneficial effects as follows:

The present invention realizes the Intelligent Recognition of XLPE cable insulation defect in conjunction with wavelet analysis and SOM neural network model, make use of the feature that two-dimensional wavelet transformation is given prominence to the noise reduction of PRPS image, bring extraction statistical nature parameter vector into SOM neural network again and carry out fault diagnosis, overcome classic method speed of convergence slow, the shortcoming of accuracy of judgement degree difference, there is detection speed fast, step is simple, the feature that diagnosis accuracy is high, and meet the visualization requirement of fault analysis, be conducive to the running status of service work personnel accurate evaluation XLPE cable, improve the intelligent level of XLPE cable partial discharge detecting system.

Accompanying drawing explanation

Fig. 1 is two layers of wavelet decomposition schematic diagram of image;

Fig. 2 is XLPE cable shelf depreciation default kind identification method process flow diagram;

Fig. 3 (a) is the original PRPS collection of illustrative plates of XLPE cable superfrequency

Fig. 3 (b) is the PRPS collection of illustrative plates after XLPE cable superfrequency noise reduction;

Fig. 4 is the SOM neural network recognization figure of XLPE cable shelf depreciation type.

Embodiment

Below in conjunction with drawings and Examples, the present invention will be further described.

A kind of XLPE cable shelf depreciation default kind identification method, as shown in Figure 2, comprises the steps:

(1) carry out the process of 2-d wavelet threshold deniosing to the XLPE cable shelf depreciation PRPS collection of illustrative plates gathered, its basic thought is, is first normalized data each in PRPS collection of illustrative plates:

$x_{i,j}^{\′}=\frac{x_{i,j}-\min \left\{x_{i,j}\right\}}{\max \left\{x_{i,j}\right\}-\min \left\{x_{i,j}\right\}}$

PRPS collection of illustrative plates, as shown in Fig. 3 (a), represents with the form of gray level image, carries out wavelet decomposition by wavelet transformation to gray level image by the original PRPS collection of illustrative plates of XLPE cable superfrequency, if gray image signals f (x, y) is in resolution 2 ^junder, decomposing through 2-d wavelet, can be A by picture breakdown _jf (x, y), and 4 subgraphs, namely

A _jf(x,y)＝{f(x,y),Φ _j,n(x)Φ _j,m(y)}；

$D_j^{1}f(x,y)=\{f(x,y),{\mathrm{\Φ}}_{j,n}\left(x\right){\mathrm{\Ψ}}_{j,m}\left(y\right)\};$

$D_j^{2}f(x,y)=\{f(x,y),{\mathrm{\Ψ}}_{j,n}\left(x\right){\mathrm{\Φ}}_{j,m}\left(y\right)\};$

$D_j^{3}f(x,y)=\{f(x,y),{\mathrm{\Ψ}}_{j,n}\left(x\right){\mathrm{\Ψ}}_{j,m}\left(y\right)\};$

In formula: Φ and Ψ is corresponding scaling function and wavelet function; A _jf (x, y) is being similar to, also referred to as low frequency part on original image; represent this approximate error, i.e. the HFS of image; for horizontal edge information; for vertical edge information; for the high-frequency information of diagonal.

Through wavelet transformation, original image will be broken down into 4 subimages, and what each subimage represented last tomographic image respectively smoothly approaches information component and horizontal vertical and diagonal line information component.Fig. 1 gives two layers of wavelet decomposition process of two dimensional image, and its wavelet reconstruction is undertaken by inverse process.Wherein, LL1, HL1, LH1 and HH1 are respectively the subimage of smoothly the approach information of original image after wavelet transformation and level, vertical and object line information structure; LL2, HL2, LH2 and HH2 be LL1 subimage again through wavelet transformation formed 4 subimages, as shown in Figure 1.

Hard-threshold process is carried out to wavelet low frequency signal

$W_{\mathrm{\δ}}=\left\{\begin{array}{c}W,\left|W\right|>\mathrm{\δ}\\ 0,\left|W\right|\≤\mathrm{\δ}\end{array}\right.$

Wherein, W _δfor wavelet coefficient values, δ is threshold value.

Low frequency component for high fdrequency component and hard-threshold process carries out small echo inversion, obtains the denoising gray-scale map after reconstructing, and carries out renormalization process, obtains XLPE cable superfrequency PRPS collection of illustrative plates after final denoising, as shown in Fig. 3 (b).

(2) extract the characteristic parameter of PRPS collection of illustrative plates, wherein comprising four has dimension statistical nature parameter (peak-to-peak value X _pk, rectification average X _av, standard deviation X _sdwith root mean square X _rm) and six dimensionless characteristic parameter (kurtosis X _ku, measure of skewness X _sk, peak factor X _c, pulse factor X _i, shape factor X _swith nargin factor X _l).

Formula for 10 statistical nature parameters is respectively:

$X_{\mathrm{pk}}=\max \left\{x_{i,j}\right\}-\min \left\{x_{i,j}\right\};X_{\mathrm{av}}=\frac{1}{\mathrm{MN}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}\left|x_{\mathrm{ij}}\right|;X_{\mathrm{sd}}=\sqrt{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{(x_{i,j}-\stackrel{\‾}{x})}^2};$

$X_{\mathrm{rm}}=\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{x}_{i,j}^2;X_{\mathrm{ku}}=\frac{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N}{\mathrm{\Σ}}}\underset{j=0}{\overset{M}{\mathrm{\Σ}}}{\left(\right|x_i|-\stackrel{\‾}{x})}^4}{{X_{\mathrm{rm}}}^2};X_{\mathrm{sk}}=\frac{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{(x_{i,j}-\stackrel{\‾}{x})}^3}{{\left(\sqrt{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{(x_{i,j}-\stackrel{\‾}{x})}^2}\right)}^3};$

$X_c=\frac{\max \left\{x_{i,j}\right\}}{\sqrt{X_{\mathrm{rm}}}};X_i=\frac{\max \left\{x_{i,j}\right\}}{X_{\mathrm{av}}};X_s=\frac{\sqrt{X_{\mathrm{rm}}}}{X_{\mathrm{av}}};X_l=\frac{\max \left\{x_{i,j}\right\}}{{\left[\frac{1}{N}\underset{i=0}{\overset{N}{\mathrm{\Σ}}}\underset{j=0}{\overset{M}{\mathrm{\Σ}}}\sqrt{\left|x_{i,j}\right|}\right]}^2}.$

(3) SOM neural metwork training is carried out to the 315 groups of PRPS collection of illustrative plates gathered, and obtain the SOM neural network recognization model that trains;

(4) extract 10 characteristic ginseng values to the PRPS collection of illustrative plates of collection in worksite, as the input quantity of the SOM neural network trained, the output node of SOM neural network is 5*4.The topological structure of output layer is the hexagonal mesh of stratiform, carries out insulation defect type identification;

Characteristic parameter is input to SOM neural network model, concrete steps are as follows:

1) initialization SOM

The neuronal quantity of input layer and competition layer is set to n, and m determines topology of networks.All weights W _ijbe initialized as the random number that [0,1] is interval.The iterations and the sample size that arrange SOM network are T and K respectively.

2) input pattern is proposed

Input vector is brought into SOM neural network.Wherein, the input vector that kth is secondary is X _k=[x _k1, x _k2.。。，x _kn]。X _kvalue will be randomly picked or from training centralized cycle choose.

3) all neuronic distances are calculated

In this research, input vector X _k=[x _k1,-x _k2..., x _kn] and each neuronic weight vectors W _j=[W _j1, W _j2..., W _jn] between distance, use standard Euclidean distance calculate.This formula is as follows:

$d_{\mathrm{jk}}=\left|\right|X_k-W_j\left|\right|=\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(X_{\mathrm{ki}}-W_{\mathrm{ji}})}^2,}j=1,2,3\·\·\·m$

4) triumph neuron is determined

Output normally weight vectors and the input vector X of SOM _knearest neuron.Suppose that triumph neuron is C, the weight vectors between triumph neuron and input neuron is W _c.That is: || X _k-W _c||=min{d _jk.

5) weight and adjacent node is upgraded

Once triumph neuron is located, SOM obtains unceasing study by upgrading neuronic weight vector of winning.According to input vector, weight vectors should pass through formula, carries out upgrading and strengthening.

W _j(t+1)＝W _j(t)+η(t)h _jc(t)[X _k(t)-W _j(t)]

Wherein, W _j(t+1) be weight vector after renewal rewards theory, W _jt () is the weight vector before renewal rewards theory, η (t) is learning rate, h _it () is neighborhood function, its formula is:

$h_{\mathrm{jc}}=\exp (-\frac{\left|\right|r_c-r_j\left|\right|}{{2\mathrm{\σ}}^2\left(t\right)})$

Wherein, r _cand r _jbe the point of neuron in two-dimensional planar array of triumph neuron and other competition respectively, σ (t) is the radius of neighbourhood.

6) select new input vector, the step (3) that algorithm is carried out in circulation is to (5), until input amendment corresponding to output layer triumph neuron is stablized.

7) renewal learning rate and neighborhood function

η(t)＝η ₀(1-t/T)

Wherein, η (t) is renewal learning rate, η ₀for renewal learning rate initial value, t is time constant, and T is phase constant.Generally, learning rate needs to be reduced to guarantee convergence.

$\left\{\begin{array}{c}\mathrm{\σ}\left(t\right)={\mathrm{\σ}}_0\exp (-t/\mathrm{\τ})\\ h_{\mathrm{iC}}\left(t\right)=\exp (-d_{\mathrm{ij}}^2/2\mathrm{\σ}{\left(t\right)}^2)\end{array}\right.$

Wherein, σ (t) is the radius of neighbourhood, and along with continuous renewal, the radius of neighbourhood is finally reduced to 0.Triumph neuron will constantly reduce neighbouring neuronic impact, to strengthen the response to determining classification.T is time constant, can by τ=1000/log (σ ₀) obtain.

8) t=t+1 is set, repeats step 2 and carry out T iteration.

Can find from learning process, weight vector is close to input model.Weight vectors collection is the description to all samples, and single weight vector can as the cluster centre of whole sample, by arranging corresponding neuron as triumph neuron sample.Then, SOM neural fusion cluster and classification feature.

Once cluster is formed, new data category will be imported into SOM, judge the cluster (using identical similarity criterion to train SOM) belonging to it.Suppose A=[a ₁, a ₂..., a _n], be a new input quantity, then find nearest competition neurons, as triumph neuron.Neuronic activation value of winning is 1, and other values are 0.The neuron of winning represents the classification results of A.

(5) classification results of XLPE cable shelf depreciation is exported: use recognition methods of the present invention, classification results accuracy rate can reach 92.1%, and as shown in Figure 4, wherein the hexagon form of different grey-scale represents each defect, totally 4 kinds.

By reference to the accompanying drawings the specific embodiment of the present invention is described although above-mentioned; but not limiting the scope of the invention; one of ordinary skill in the art should be understood that; on the basis of technical scheme of the present invention, those skilled in the art do not need to pay various amendment or distortion that creative work can make still within protection scope of the present invention.

Claims (10)

1. an XLPE cable shelf depreciation default kind identification method, is characterized in that, comprises the steps:

(1) the XLPE cable shelf depreciation PRPS collection of illustrative plates gathered, and the process of 2-d wavelet threshold deniosing is carried out to described PRPS collection of illustrative plates, obtain the XLPE cable superfrequency PRPS collection of illustrative plates after denoising;

(2) extract and calculate the characteristic parameter of PRPS collection of illustrative plates, comprising peak-to-peak value X _pk, rectification average X _av, standard deviation X _sdwith root mean square X _rmfour have dimension statistical nature parameter and kurtosis X _ku, measure of skewness X _sk, peak factor X _c, pulse factor X _i, shape factor X _swith nargin factor X _lsix dimensionless characteristic parameters;

(3) above-mentioned characteristic parameter is input to SOM neural network model; By calculating the distance between input vector and each neuronic weight vectors, determine triumph neuron, SOM neural network is by upgrading neuronic weight vector unceasing study of winning, until input amendment corresponding to output layer triumph neuron is stablized;

(4) export the type identification result of SOM neural network, namely export corresponding XLPE cable shelf depreciation defect type.

2. a kind of XLPE cable shelf depreciation default kind identification method as claimed in claim 1, it is characterized in that, the concrete grammar of described step (1) is:

1) data each in PRPS collection of illustrative plates are normalized:

$x_{i,j}^{\′}=\frac{x_{i,j}-\min \left\{x_{i,j}\right\}}{\max \left\{x_{i,j}\right\}-\min \left\{x_{i,j}\right\}}$

Wherein, x _i,jbe the i-th row, peak value of pulse, x ＇ are put in the original office of j row _i,jfor peak value of pulse, min{x are put in the office after normalization _i,jput peak value, the max{x of minimum pulse in pulse for all original offices _i,jput the peak value of maximum impulse in pulse for all original offices;

2) PRPS collection of illustrative plates is represented with the form of gray level image, by wavelet transformation, 2-d wavelet decomposition is carried out to gray level image;

3) hard-threshold process is carried out to the wavelet low frequency signal after decomposition:

$W_{\mathrm{\δ}}=\left\{\begin{array}{c}W,\left|W\right|>\mathrm{\δ}\\ 0,\left|W\right|\≤\mathrm{\δ}\end{array}\right.$

Wherein, W _δfor wavelet coefficient values, δ is threshold value;

4) low frequency component of high fdrequency component and hard-threshold process is carried out small echo inversion, obtain the denoising gray-scale map after reconstructing, and renormalization process is carried out to described denoising gray-scale map, obtain XLPE cable superfrequency PRPS collection of illustrative plates after final denoising.

3. a kind of XLPE cable shelf depreciation default kind identification method as claimed in claim 1, is characterized in that, in described step (2), four have the computing method of dimension statistical nature parameter to be specially:

$X_{\mathrm{pk}}=\max \left\{x_{i,j}\right\}-\min \left\{x_{i,j}\right\};X_{\mathrm{av}}=\frac{1}{\mathrm{MN}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}\left|x_{\mathrm{ij}}\right|;$

$X_{\mathrm{sd}}=\sqrt{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{(x_{i,j}-\stackrel{\‾}{x})}^2};X_{\mathrm{rm}}=\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{x}_{i,j}^2$

Wherein, X _pkfor the peak-to-peak value of pulse signal is put in office, X _avfor rectification average, X _sdfor standard deviation, X _rmfor mean square value, x _i,jbe the i-th row, peak value of pulse is put in the original office of j row, and M is number of phases, and N is power frequency period number, min{x _i,jput the peak value of minimum pulse in pulse, max{x for all original offices _i,jput the peak value of maximum impulse in pulse for all original offices, for the average of pulse signal peak value is put in the original office of entirety.

4. a kind of XLPE cable shelf depreciation default kind identification method as claimed in claim 1, it is characterized in that, in described step (2), the computing method of six dimensionless statistical nature parameters are specially:

$X_{\mathrm{ku}}=\frac{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N}{\mathrm{\Σ}}}\underset{j=0}{\overset{M}{\mathrm{\Σ}}}{\left(\right|x_{i,j}|-\stackrel{\‾}{x})}^4}{{X_{\mathrm{rm}}}^2};X_{\mathrm{sk}}=\frac{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{(x_{i,j}-\stackrel{\‾}{x})}^3}{{\left(\sqrt{\frac{1}{\mathrm{NM}}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{(x_{i,j}-\stackrel{\‾}{x})}^2}\right)}^3};$

$X_c=\frac{\max \left\{x_{i,j}\right\}}{\sqrt{X_{\mathrm{rm}}}};X_i=\frac{\max \left\{x_{i,j}\right\}}{X_{\mathrm{av}}};X_s=\frac{\sqrt{X_{\mathrm{rm}}}}{X_{\mathrm{av}}};X_l=\frac{\max \left\{x_{i,j}\right\}}{[\frac{1}{N}\underset{i=0}{\overset{N}{\mathrm{\Σ}}}\underset{j=0}{\overset{M}{\mathrm{\Σ}}}\sqrt{\left|x_{i,j}\right|}{]}^2};$

Wherein, X _kufor the kurtosis of pulse signal is put in office, X _skfor measure of skewness, X _cfor peak factor, X _ifor the pulse factor, X _sfor shape factor, X _lfor the nargin factor, X _pkfor peak-to-peak value, X _avfor rectification average, X _sdfor standard deviation, X _rmfor mean square value, x _i,jbe the i-th row, peak value of pulse is put in the original office of j row, and M is number of phases, and N is power frequency period number, min{x _i,jput the peak value of minimum pulse in pulse, max{x for all original offices _i,jput the peak value of minimum pulse in pulse for all original offices, for the average of pulse signal peak value is put in the original office of entirety.

5. a kind of XLPE cable shelf depreciation default kind identification method as claimed in claim 1, it is characterized in that, the concrete grammar of described step (3) is:

A) initialization SOM neural network model:

The neuronal quantity of input layer and competition layer in SOM neural network model is set to n, m respectively, determines topology of networks; The iterations and the sample size that arrange SOM network are T and K respectively; By neuronic weights W each in SOM network _j,ibe initialized as the random number that [0,1] is interval;

B) by input vector, i.e. 10 statistical nature parameter vectors, bring SOM neural network model into:

Wherein, the input vector that kth is secondary is X _k=[x _k1, x _k2..., x _k10], X _kvalue random selecting or from training centralized cycle choose;

C) all input vector X are calculated _k=[x _k1, x _k2..., x _k10] and each neuronic weight vectors W _j=[W _j, W _j..., W _j10] between distance d _jk;

D) triumph neuron is determined:

By weight vectors and input vector X _knearest neuron is as the triumph neuron C of SOM neural network model;

E) weight vector is upgraded:

Once triumph neuron C is located, SOM neural network model is constantly learnt by upgrading neuronic weight vector of winning;

Renewal learning rate formula is

η(t)＝η ₀(1-t/T)

Wherein, η (t) is renewal learning rate, η ₀for renewal learning rate initial value, t is time constant, and T is phase constant;

F) select new input vector, step (c) to (e) is carried out in circulation, until input amendment corresponding to output layer triumph neuron is stablized;

H) t=t+1 is set, repeats step 2 and carry out T iteration.

6. a kind of XLPE cable shelf depreciation default kind identification method as claimed in claim 5, is characterized in that, described step c) in calculate the formula of the spacing of all input vectors and each neuronic weight vectors as follows:

$d_{\mathrm{jk}}=\left|\right|X_k-W_j\left|\right|=\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(X_{\mathrm{ki}}-W_{\mathrm{ji}})}^2},j=\mathrm{1,2,3}...m$

Wherein, X _kfor kth time input vector, W _jfor each neuronic weight vectors, X _kifor i-th input vector that kth is secondary, W _jifor the weight vectors of corresponding i-th input vector of a jth neuron.

7. a kind of XLPE cable shelf depreciation default kind identification method as claimed in claim 5, is characterized in that, described step e) in the update method of weight vectors be:

W _j(t+1)＝W _j(t)+η(t)h _jc(t)[X _k(t)-W _j(t)]

Wherein, W _j(t+1) be weight vector after renewal rewards theory, W _jt () is the weight vector before renewal rewards theory, η (t) is learning rate, h _ict () is neighborhood function.

8. a kind of XLPE cable shelf depreciation default kind identification method as claimed in claim 7, is characterized in that, described neighborhood function h _ict the computing formula of () is:

$h_{\mathrm{jc}}=\exp (-\frac{\left|\right|r_c-r_j\left|\right|}{2{\mathrm{\σ}}^2\left(t\right)})$

Wherein, r _cand r _jbe the point of neuron in two-dimensional planar array of triumph neuron and other competition respectively, σ (t) is the radius of neighbourhood, and along with continuous renewal, the radius of neighbourhood is finally reduced to 0.

9. a kind of XLPE cable shelf depreciation default kind identification method as claimed in claim 5, is characterized in that, described step e) in, learning rate needs to be reduced to guarantee convergence.

10. a kind of XLPE cable shelf depreciation default kind identification method as claimed in claim 1, is characterized in that, the concrete grammar exporting the type identification result of SOM neural network in described step (4) is:

By arranging corresponding neuron as triumph neuron sample, neuronic activation value of winning is 1, and other values are 0; Using the classification results of triumph neuron sample as input vector, realize the classification feature of SOM neural network.