CN103558529A - Method for pattern recognition of three-phase drum-sharing type ultrahigh voltage GIS partial discharge - Google Patents

Abstract

The invention discloses a method for pattern recognition of three-phase drum-sharing type ultrahigh voltage GIS partial discharge. The method comprises the steps that three-phase drum-sharing type GIS partial discharge is detected through ultrahigh frequency and partial discharging signals are sampled through a UHF sensor; denoising is conducted on the collected partial discharging signals according to the improved wavelet threshold value filtering method so that real partial discharging signals can be obtained; characteristic parameters of the sampling signals are extracted through an algorithm based on a phase position analytical pattern; dimension reduction processing is conducted on characteristic space composed of the characteristic parameters according to the improved core principal component analysis method so that a characteristic parameter matrix after dimension reduction can be obtained; pattern recognition is conducted on the GIS insulation detect type according to the K adjacent classification method based on cluster. According to the method for pattern recognition of three-phase drum-sharing type ultrahigh voltage GIS partial discharge, defects of less functions, a small application range, poor accuracy and the like in the prior art are overcome and the advantages of more functions, a wide application range and good accuracy are achieved.

Description

A kind of three-phase is the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation altogether

Technical field

The present invention relates to high-tension electricity electric discharge recognition technology field, particularly, relate to a kind of three-phase mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation altogether.

Background technology

Gas-insulated switchgear (Gas lnsulated Switchgear, be called for short GIS) be one of important composition equipment in extra-high voltage grid, last concentrating is assembled in one and fills and using in the monolithic case of SF6 as insulating medium in being contained in respectively between each self sealss after the optimal design such as the isolating switch in Ta Jiangyizuo transformer station, current transformer, voltage transformer (VT), lightning arrester, disconnector, grounding switch, bus, cable termination, inlet-outlet sleeve.

The defect of GIS internal influence insulating medium performance mainly contains: loose contact between serious setup error, conductor, high-pressure conductor protrusion, fixedly particulate, defects of insulator, steam etc.

The development of GIS is tending towards three-phase altogether cylinderization, Composite and intellectuality, owing to having realized in miniaturization ，Ke factory, carries out with the form at interval, transporting to scene after final assembly and stand the test, therefore can shorten on-the-spot installation period, the raising of having got back of while reliability.

Altogether cartridge type GIS is in inner structure for three-phase, and it is significantly different that the aspects such as Electric Field Distribution and close coupled type GIS have, and prior art research mainly collects at close coupled type GIS, but less for the research of the common cartridge type UHV (ultra-high voltage) GIS Partial Discharge Detection pattern-recognition of three-phase.

In realizing process of the present invention, inventor finds at least to exist in prior art that function is few, the scope of application is little and the defect such as poor accuracy.

Summary of the invention

The object of the invention is to, for the problems referred to above, propose the altogether mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation of a kind of three-phase, and advantage that accuracy good many, applied widely with practical function.

For achieving the above object, the technical solution used in the present invention is: a kind of three-phase is the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation altogether, comprises the following steps:

Step 1: adopt ultrahigh frequency to detect three-phase cartridge type GIS shelf depreciation altogether, utilize UHF sensor to sample to local discharge signal;

Step 2: utilize improved small echo threshold values filtering method to carry out denoising Processing to the local discharge signal collecting, obtain real local discharge signal;

Step 3: by extract the characteristic parameter of sampled signal based on phase analysis pattern algorithm;

Step 4: the feature space that utilizes improved core principle component analysis method to form characteristic parameter carries out dimension-reduction treatment, obtains the characteristic parameter matrix after dimensionality reduction;

Step 5: utilize the k nearest neighbor classification based on bunch thought to carry out pattern-recognition to GIS insulation defect type.

Further, in step 2, describedly utilize improved small echo threshold values filtering method the local discharge signal collecting to be carried out to the operation of denoising Processing, specifically adopt adaptive thresholding value calculating method; This adaptive thresholding value calculating method is as follows:

$T_j=\frac{\mathrm{Median}\left(\right|C_{j,k}\left|\right)\sqrt{2\ln \left(N_j\right)}}{0.6745}\mathrm{\αexp}(-\frac{1}{{\mathrm{\β}}_j});$

Wherein, j is yardstick, N _jfor the number of wavelet coefficient on this yardstick, Median (| C _j,k|) be the median of all wavelet coefficients on this yardstick, α is called the signal to noise ratio (S/N ratio) factor, is the embodiment in threshold values calculates of the signal to noise ratio (S/N ratio) of signal, β _jbeing called scale factor, is that the maximal value of wavelet coefficient on yardstick is corrected the evaluated error that sample sequence length difference causes, T _jfor the threshold values calculating.

Further, in step 3, described characteristic parameter comprises measure of skewness Sk, steepness Ku, local peaks count Pe, cross-correlation coefficient Cc and discharge factor Q.

Further, described measure of skewness Sk is specially:

$\mathrm{Sk}=\underset{i=1}{\overset{W}{\mathrm{\Σ}}}{(x_i-\mathrm{\μ})}^3\·p_i\mathrm{\Δx}/{\mathrm{\σ}}^3;$

In above formula, w is the phase window number in the semiperiod; Xi is the phase place of i phase window;

$p_i=y_i/\underset{i=1}{\overset{W}{\mathrm{\Σ}}}{y}_i,$

Wherein, y _ibe the ordinate of spectrogram, represent Apparent discharge magnitude q or discharge time n; The position of the shelf depreciation collection of illustrative plates center that parameter μ representative collects, σ represents the precipitous situation that the center axis of symmetry of collection of illustrative plates embodies, Δ x be about one of shelf depreciation collection of illustrative plates to be evenly distributed relevant parameter,

certain in collection of illustrative plates is put corresponding phase place;

Measure of skewness Sk reflection spectrogram shape is with respect to the left and right deflection situation of normal distribution: Sk=0 illustrates that this spectrogram shape is symmetrical; Sk>0 illustrate this spectrogram with respect to normal distribution shape to left avertence; Sk<0 illustrate this spectrogram with respect to normal distribution shape to right avertence.

Further, described steepness Ku is specially:

$\mathrm{Ku}=[\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{(x_i-\mathrm{\&mu;})}^4{p}_i\mathrm{\&Delta;x}/{\mathrm{\&sigma;}}^4]-3;$

In above formula, w is the phase window number in the semiperiod; Xi is the phase place of i phase window;

$p_i=y_i/\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{y}_i,$

Wherein, y _ibe the ordinate of spectrogram, represent Apparent discharge magnitude q or discharge time n; The position of the shelf depreciation collection of illustrative plates center that parameter μ representative collects, σ represents the precipitous situation that the center axis of symmetry of collection of illustrative plates embodies, Δ x be about one of shelf depreciation collection of illustrative plates to be evenly distributed relevant parameter,

certain in collection of illustrative plates is put corresponding phase place;

Steepness Ku in contrast to the projection degree of normal distribution shape for describing the distribution of certain shape: the steepness Ku of normal distribution is 0; If Ku>0, illustrates that this spectrogram profile is sharply more precipitous than normal distribution profile; If Ku<0, illustrates that this spectrogram profile is more smooth than normal distribution profile.

Further, the described local peaks Pe that counts, local peaks is counted for describing the number of local peaks on spectrogram profile; In point

whether place has local peaks, needs to judge with difference equation below:

（y _i-y _i-1）>0，（y _i+1-y _i）<0；

The halved phase window of phase shaft is more, and local peaks is counted larger.

Further, described cross-correlation coefficient Cc is specially:

$\mathrm{Cc}=\frac{\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{+}{q}_i^{-}-\left(\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{+}\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{-}\right)/W}{\sqrt{[\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{\left(q_i^{+}\right)}^2-{\left(\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{-}\right)}^2/W][\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{\left(q_i^{-}\right)}^2-{\left(\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{-}\right)}^2]/W}};$

In formula,

be the discharge capacity in phase window i, subscript "+", "-" are corresponding to the positive and negative semiaxis of spectrogram; C has reacted the correlativity of the strong and weak and PHASE DISTRIBUTION of electric discharge in positive and negative half period, and cross-correlation coefficient Cc means close to 1

the profile of spectrogram positive-negative half-cycle is quite similar; Cc is close to 0,

spectrogram profile difference is huge.

Further, described discharge factor Q is specially:

$Q=\frac{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{-}{q}_i^{-}}{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{-}}/\frac{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{+}{q}_i^{+}}{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{+}};$

In formula, the electric discharge repetition rate in phase window i, subscript "+", "-" corresponding to

the positive-negative half-cycle of spectrogram.

Further, in step 4, described improved core principle component analysis method is improved core principle component analysis method, and the kernel function of sampling is:

$k(x_i,x_j)={(<x_i,x_j>+a)}^b\exp \left(\frac{-{\left|\right|x_i-x_j\left|\right|}^2}{2{\mathrm{\&sigma;}}^2}\right);$

Wherein, (a ∈ R, b ∈ N, σ >0), the selection of parameter a, b and σ is to determine according to the numerical values recited of element in eigenmatrix, parameter σ is for controlling the radial effect scope of kernel function; x _iand x _jrepresent different sample vector ， ﹙ x _i, x _jthe vector product of ﹚ representative sample vector, the span of R representation vector is at set of real numbers, and N represents set of integers, k (x _i, x _j) the new kernel function that obtains in conjunction with the advantage of polynomial kernel function and gaussian kernel function of representative.

Further, in step 5, the algorithm of described k nearest neighbor classification specifically comprises:

Step1: in training set, first all Partial Discharge Datas are carried out to pre-service and become space vector;

Step2: from first class, to belonging to all signal datas of this classification, carry out between two similarity and calculate, set a minimum threshold, according to statistics, obtain that similarity approaches one by one bunch;

Step3: for each bunch, all signal datas are wherein merged, then calculate its center vector; In addition, compute cluster number/classification sum, this value represents the contribution coefficient of this bunch to this class;

Step4: after new text arrives, carry out the vector space that pre-service obtains it;

Step5: the center vector of every cluster that the space vector of new text and Step3 are generated calculates distance, these distances are multiplied each other with the contribution coefficient of corresponding bunch, the results added that belongs to other bunch of calculating of same class, relatively obtaining that maximum classification is exactly classification under typical defect shelf depreciation to be sorted.

The three-phase of various embodiments of the present invention is the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation altogether, owing to comprising step: adopt ultrahigh frequency to detect three-phase cartridge type GIS shelf depreciation altogether, utilize UHF sensor to sample to local discharge signal; Utilize improved small echo threshold values filtering method to carry out denoising Processing to the local discharge signal collecting, obtain real local discharge signal; By extract the characteristic parameter of sampled signal based on phase analysis pattern algorithm; The feature space that utilizes improved core principle component analysis method to form characteristic parameter carries out dimension-reduction treatment, obtains the characteristic parameter matrix after dimensionality reduction; The k nearest neighbor classification of utilization based on bunch thought carried out pattern-recognition to GIS insulation defect type; Can overcome the defect of prior art, improve the three-phase accuracy of cartridge type UHV (ultra-high voltage) GIS Partial Discharge Detection pattern-recognition altogether; Thereby can overcome, in prior art, function is few, the scope of application is little and the defect of poor accuracy, and advantage that accuracy good many, applied widely with practical function.

Other features and advantages of the present invention will be set forth in the following description, and, partly from instructions, become apparent, or understand by implementing the present invention.

Below by drawings and Examples, technical scheme of the present invention is described in further detail.

Accompanying drawing explanation

Accompanying drawing is used to provide a further understanding of the present invention, and forms a part for instructions, for explaining the present invention, is not construed as limiting the invention together with embodiments of the present invention.In the accompanying drawings:

Fig. 1 is the altogether structural representation of local discharge of gas-insulator switchgear test unit in the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation of three-phase of the present invention;

Fig. 2 is the three-phase of the present invention schematic flow sheet of the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation altogether;

Fig. 3 be three-phase of the present invention altogether in the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation statistical distribution and Sk, Ku be related to schematic diagram;

Fig. 4 is the three-phase of the present invention action effect schematic diagram of the mode identification method Kernel Function of cartridge type UHV (ultra-high voltage) GIS shelf depreciation altogether;

Fig. 5 is altogether comparison of wave shape figure before and after filtering in the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation of three-phase of the present invention.

By reference to the accompanying drawings, in the embodiment of the present invention, Reference numeral is as follows:

1-water resistance; 2-HT testing transformer; 3-transformer; 4-bushing; 5-disk insulator.

Embodiment

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described, should be appreciated that preferred embodiment described herein, only for description and interpretation the present invention, is not intended to limit the present invention.

For the defect existing in prior art, according to the embodiment of the present invention, as Figure 1-Figure 5, provide a kind of three-phase mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation altogether.

As shown in Figure 1, the three-phase of the present embodiment local discharge of gas-insulator switchgear test unit that the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation is used altogether, comprise disk insulator 5, be arranged on the bushing 4 on disk insulator 5, the water resistance 1 being connected with bushing 4 successively, HT testing transformer 2, transformer 3 and local discharge signal detector (PDSG); Local discharge signal detector (PDSG) is also connected with disk insulator 5, the common port of water resistance 1 and bushing 4 by the voltage divider being formed by electric capacity after ground connection.

This local discharge of gas-insulator switchgear test unit, mainly comprises transformer 3, the voltage divider being comprised of electric capacity, oscillograph, three-phase cartridge type gas insulated combined electrical equipment (GIS), sensor and local discharge signal detector (PDSG) altogether; By the defects such as high-pressure conductor metal protrusion, free metal particulate, insulator surface fixing metal, insulator air gap are set respectively in GIS, detect corresponding local discharge signal, carry out pattern-recognition.

The three-phase of the present embodiment is the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation altogether, comprises the following steps:

1) adopt ultrahigh frequency (UHF) to detect three-phase cartridge type GIS shelf depreciation altogether, utilize UHF sensor to shelf depreciation (PD) signal sampling;

In step 1), adopt ultrahigh frequency method (UHF) to be total to cartridge type GIS shelf depreciation to three-phase and detect.While detecting GIS shelf depreciation by UHF method, can identify different faults type with the position that electric discharge occurs on power frequency voltage waveform according to the spectral characteristic that records signal.

In step 1), adopt actual three-phase cartridge type GIS altogether, shelf depreciation collection of illustrative plates under typical defect condition is measured.Wherein three-phase altogether cartridge type GIS should have professional high-tension switch gear enterprise to provide.The pressuring method of three-phase conductor is two phase ground, and the high pressure that joins, is only provided with a bushing in model.According to coaxial type GIS test model test result, the typical insulation defect with contrast property is set.There is free metal particulate in cartridge type GIS cavity in three-phase, insulator surface exists fixing metal particle, has void defects on insulator, and shelf depreciation physical model is placed in to GIS analogue means, carries out the measurement of shelf depreciation altogether.

2) utilize improved small echo threshold values filtering method to carry out denoising Processing to the local discharge signal collecting, obtain real local discharge signal;

In step 2) in, improved small echo threshold values filtering algorithm, from median Median(Cj, k) computing method, the selection of threshold value and the length relation of analyzed signal are very close.In the middle of practical application, can not guarantee that useful signal is constant in whole section of sampling number shared ratio and position in whole section of sample sequence.Because the length of signal has determined after wavelet transformation the number Nj of wavelet coefficient on each yardstick, it affects median, Median(Cj, k) value, thereby affected the size of threshold value.This impact can cause such result: same useful signal sequence u is comprised in the middle of one section of sample sequence s, if s length different (width that is time view window is different), the filtered result of wavelet threshold has larger difference.For the s of the different length that comprises same UHF PD signal, the result of application soft-threshold filtering algorithm.Original signal sample frequency is 20GHz, and the wider long sample sequence of time window is 50000 points, and the narrower short sample sequence of time window is 16000 points, and two sequences are all complete has comprised useful UHF PD signal.Obviously,, there is obvious difference in the part that the filtered UHF PD of application soft-threshold algorithm signal, especially oscillatory extinction will finish, and both related coefficients are 0.6677.

The basic reason that produces this consequence is to have ignored the amplitude of useful signal and the factor of signal to noise ratio (S/N ratio) completely in the computing formula of threshold value.Given this, the present embodiment is through repeatedly analyzing, considered the factor of useful signal amplitude and signal to noise ratio (S/N ratio), a kind of new adaptive thresholding value calculating method has been proposed, this adaptive thresholding value calculating method greatly reduces the sensitivity of wavelet threshold filtering result to sampled point, as shown in Figure 5, in Fig. 5, (a) is figure before filtering to graphic correlation after worry wavefront, is (b) figure after filtering:

$T_j=\frac{\mathrm{Median}\left(\right|C_{j,k}\left|\right)\sqrt{2\ln \left(N_j\right)}}{0.6745}\mathrm{\&alpha;exp}(-\frac{1}{{\mathrm{\&beta;}}_j});$

Wherein, j is yardstick, N _jfor the number of wavelet coefficient on this yardstick, Median (| C _j,k|) be the median of all wavelet coefficients on this yardstick, α is called the signal to noise ratio (S/N ratio) factor, is the embodiment in threshold values calculates of the signal to noise ratio (S/N ratio) of signal, β _jbeing called scale factor, is that the maximal value of wavelet coefficient on yardstick is corrected the evaluated error that sample sequence length difference causes, T _jfor the threshold values calculating.

The validity of Methods for Wavelet Denoising Used depends primarily on wavelet basis function, wavelet decomposition yardstick, threshold values function, threshold values and several aspects such as chooses.In the present embodiment with the analysis verification of a large amount of emulation experiments, laboratory simulation and field measurement data the validity of method therefor.Result shows, this Methods for Wavelet Denoising Used is compared with the denoising algorithm of other threshold values rule, has obviously improved the de-noising ability in signal processing of partial discharge, but also have process after signal waveform distortion little, extract the advantages such as more accurate, suffered influence factor is few.

3) by extract the characteristic parameter of sampled signal based on phase analysis pattern algorithm, preferred, characteristic parameter comprises: measure of skewness Sk, steepness Ku, local peaks count Pe, cross-correlation coefficient Cc and discharge factor Q;

In step 3), for the signal after sampling, adopt the extraction characteristic parameter based on phase analysis pattern (PRPD), wherein: measure of skewness Sk, the definition of Sk:

$\mathrm{Sk}=\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{(x_i-\mathrm{\&mu;})}^3\&CenterDot;p_i\mathrm{\&Delta;x}/{\mathrm{\&sigma;}}^3;$

In above formula, w is the phase window number in the semiperiod; Xi is the phase place of i phase window;

$p_i=y_i/\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{y}_i,$

Wherein, y _ibe the ordinate of spectrogram, represent Apparent discharge magnitude q or discharge time n; The position of the shelf depreciation collection of illustrative plates center that parameter μ representative collects, σ represents the precipitous situation that the center axis of symmetry of collection of illustrative plates embodies, Δ x be about one of shelf depreciation collection of illustrative plates to be evenly distributed relevant parameter,

certain in collection of illustrative plates is put corresponding phase place.

Measure of skewness Sk reflection spectrogram shape is with respect to the left and right deflection situation of normal distribution: Sk=0 illustrates that this spectrogram shape is symmetrical; Sk>0 illustrate this spectrogram with respect to normal distribution shape to left avertence; Sk<0 illustrate this spectrogram with respect to normal distribution shape to right avertence.

The definition of phase window: the building method of Φ-q-n space curved surface: by operating frequency phase according to.0-360 ° is divided into 256 minizones, and discharge pulse amplitude q is divided into 128 minizones by maximum amplitude, thereby Φ-q plane is divided into 128 * 256 minizones; Discharge time in statistics Φ-q plane in each minizone, obtains middle Φ-q-n space curved surface.

Steepness Ku, the definition of Ku:

$\mathrm{Ku}=[\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{(x_i-\mathrm{\&mu;})}^4{p}_i\mathrm{\&Delta;x}/{\mathrm{\&sigma;}}^4]-3;$

The definition of wherein each variable is identical with variable-definition in measure of skewness.Steepness Ku in contrast to the projection degree of normal distribution shape for describing the distribution of certain shape: the steepness Ku of normal distribution is 0; If Ku>0, illustrates that this spectrogram profile is sharply more precipitous than normal distribution profile; If Ku<0, illustrates that this spectrogram profile is more smooth than normal distribution profile.

The local peaks Pe that counts, local peaks is counted for describing the number of local peaks on spectrogram profile.In point

whether place has local peaks, and available following formula is judged:

and

Above formula becomes difference equation:

Difference equation can be reduced to:

（y _i-y _i-1）>0，（y _i+1-y _i）<0；

In actual computation, local peaks is counted closely related with the phase window number of spectrogram.Generally speaking, the halved phase window of phase shaft is more, and local peaks is counted larger.

Cross-correlation coefficient Cc, the definition of cross-correlation coefficient Cc:

$\mathrm{Cc}=\frac{\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{+}{q}_i^{-}-\left(\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{+}\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{-}\right)/W}{\sqrt{[\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{\left(q_i^{+}\right)}^2-{\left(\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{-}\right)}^2/W][\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{\left(q_i^{-}\right)}^2-{\left(\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{-}\right)}^2]/W}};$

In formula,

be the discharge capacity in phase window i, subscript "+", "-" are corresponding to the positive and negative semiaxis of spectrogram.Cc has reacted the correlativity of the strong and weak and PHASE DISTRIBUTION of electric discharge in positive and negative half period, and cross-correlation coefficient Cc means close to 1

the profile of spectrogram positive-negative half-cycle is quite similar; Cc is close to 0,

spectrogram profile difference is huge.

Discharge factor Q,

$Q=\frac{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{-}{q}_i^{-}}{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{-}}/\frac{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{+}{q}_i^{+}}{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{+}};$

In formula,

the electric discharge repetition rate in phase window i, subscript "+", "-" corresponding to the positive-negative half-cycle of spectrogram.Discharge capacity factor Q has reacted

the difference of mean discharge magnitude in spectrogram positive-negative half-cycle.

According to the formula of each Statistical Operator above, by analysis of spectra and algorithm, calculate Statistical Operator, extract that characteristic parameter drift rate (SK), steepness (Ku), local peaks are counted (Pe), discharge factor (Q), cross-correlation coefficient (Cc) carry out pattern-recognition.

4) feature space that utilizes improved core principle component analysis method to form characteristic parameter carries out dimension-reduction treatment, obtains the characteristic parameter matrix after dimensionality reduction;

In step 4), owing to cannot knowing that in advance which characteristic parameter can construct the simplest feature space of UHF PD signal, i.e. the break-even characteristic parameter matrix of being completely lost, the feature space dimension of structure is higher, and may there is dimension redundancy, unfavorable to operation and recognition result.Therefore, have and need to carry out the dimension-reduction treatment of feature space.

In the middle of the application of KPCA, the selection of nonlinear transformation (being kernel function) is extremely important.Conventional kernel function has polynomial kernel function (Polynomial), gaussian kernel function (Gauss) and Sigmoid kernel function, as follows respectively:

k(x _i,x _j)=(<x _i,x _j>+a) ^b，(a∈R,b∈N)；

$k(x_i,x_j)=\exp \left(\frac{-{\left|\right|x_i-x_j\left|\right|}^2}{2{\mathrm{\&sigma;}}^2}\right);$

k(x _i,x _j)=tanh(<x _i,x _j>+a)，(a∈R)；

Wherein, <x _i, x _j> is sample vector, x _iand x _jvector product, || x _i-x _j|| be both Euclidean Norms.Polynomial kernel function is distance (<x _i, x _j>+a) b power is <x _i, x _jthe monotonic increasing function of >.Through conversion, if (<x _i, x _j>+a) >1, raw range <x _i, x _j> can be exaggerated; On the contrary, if (<x _i, x _j>+a) <1, raw range <x _i, x _j> can be compressed.The effect of visible polynomial kernel function is that small distance compression is further expanded large distance.Gaussian kernel function, also referred to as radial basis kernel function, is also defined as the index monotonically decreasing function of two vectorial Euclidean distances conventionally, and it is a kind of scalar function of radial symmetry.Wherein σ is called width parameter, for the radial effect scope of control function, the i.e. width of Gauss pulse.But the reach of gaussian kernel function is less conventionally, and action effect is just in time contrary with polynomial kernel function, small distance is expanded and a large Range compress.

The effect of kernel function should be that raw range is further expanded in fact, or the Range compress between similar sample is expanded the distance between non-similar sample, so that carry out Classification and Identification.Given this, the present embodiment, in conjunction with the relative merits of polynomial kernel function and gaussian kernel function, has proposed a kind of new kernel function, and its action effect as shown in Figure 4.

In step 4), the kernel function that improved core principle component analysis method is sampled is:

$k(x_i,x_j)={(<x_i,x_j>+a)}^b\exp \left(\frac{-{\left|\right|x_i-x_j\left|\right|}^2}{2{\mathrm{\&sigma;}}^2}\right);$

Wherein, (a ∈ R, b ∈ N, σ >0), the selection of parameter a, b and σ is to determine according to the numerical values recited of element in eigenmatrix, parameter σ is for controlling the radial effect scope of kernel function; x _iand x _jrepresent different sample vector ， ﹙ x _i, x _jthe vector product of ﹚ representative sample vector, the span of R representation vector is at set of real numbers, and N represents set of integers, k (x _i, x _j) the new kernel function that obtains in conjunction with the advantage of polynomial kernel function and gaussian kernel function of representative.Get a=5 herein, b=1 is in order to make the distance after conversion do a certain proportion of variation with former distance.Parameter σ is for controlling the radial effect scope of kernel function, because two vectorial distances in primitive character matrix are generally no more than 7, so get σ=7.As seen from Figure 4, nucleus function proposed by the invention becomes large by original small distance, and large distance is suitably dwindled.

5) utilize the k nearest neighbor classification based on bunch thought to carry out pattern-recognition to GIS insulation defect type;

In step 5), its basic thought of K arest neighbors method is: provide test document, system is searched the K nearest with it neighbours in the training set of having classified, and obtains the classification of test document according to these neighbours' category distribution situation.Wherein can be weighted by the similarity of these neighbours and test document, thereby obtain good classifying quality.So-called bunch, the meaning is exactly the set that a class has the text of similar quality, the present invention closes and thinks one bunch belonging between other text of same class those local discharge signal data subsets of distance maximum in training set, and therefore, the algorithm of k nearest neighbor classification can be described below:

Step1: in training set, first all Partial Discharge Datas are carried out to pre-service and become space vector;

Step2: from first class, to belonging to all signal datas of this classification, carry out between two similarity and calculate, set a minimum threshold, according to statistics, obtain that similarity approaches one by one bunch;

Step3: for each bunch, all signal datas are wherein merged, then calculate its center vector, in addition, and compute cluster number/classification sum, this value represents the contribution coefficient of this bunch to this class, is denoted as C;

Step4: after new text arrives, carry out the vector space that pre-service obtains it;

Step5: the center vector of every cluster that the space vector of new text and Step3 are generated calculates distance, these distances are multiplied each other with the contribution coefficient of corresponding bunch, the results added that belongs to other bunch of calculating of same class, relatively obtaining that maximum classification is exactly classification under typical defect shelf depreciation to be sorted.

The basis of this algorithm is which text of how finding out in same classification belongs to same cluster, below provides the generation bunch algorithm idea of finding out same classification bunch: suppose classification:

c={d1，d2，...................，dm}

Step1: the threshold value a that sets a similarity;

Step2: first create one bunch, be denoted as T0, with the number of documents comprising in Ki record bunch, the number of clusters amount that total record creates, the processed document i=2 of initialization;

Step3: from di;

Step4: carry out the similarity value of calculating s with first text in Tn;

Step5: if s>=a and also has the sample that sample does not compare therewith in Tn, proceed so similarity and calculate and upgrade s; If there is no not comparative sample, so these data joined in bunch Tn and gone; If s<a, if having that other does not compare bunch, n++, returns to step4; If do not compare bunch, create so new bunch, be designated as T++total; This document is classified as in T++total bunch;

Step6: if i unequal to is m, i++ so; Return to Step3; Otherwise, finish.

In order to overcome the defect that nearest neighbor method false determination ratio is higher, arest neighbors is generalized to k nearest neighbor, k nearest neighbor method is not to choose an arest neighbors to classify, but chooses from K nearest representative point of text to be sorted, then according to the classification information of this K representative point, determines the classification of text to be sorted.

For the characteristic parameter matrix after dimensionality reduction, a half-sample is used for training k nearest neighbor sorter, and second half is for the performance of testing classification device.For the characteristic parameter matrix of applying respectively after KPCA, RST and CCMDR algorithm dimensionality reduction, application k nearest neighbor sorter is identified GIS insulation defect type.The present embodiment has been write program file under C lingware environment, realizes design, training and the Classification and Identification test of sorter.Because the output of the sorter of the present embodiment design does not distribute as being dispersed BP neural network centered by certain point, but corresponding to 4 class GIS defect types, output value only comprises 4 kinds of results [1,2,3,4], so pattern-recognition result only represents with recognition correct rate, as shown in table 1.

table 1:K nearest neighbor algorithm pattern-recognition accuracy

Defect type	K nearest neighbor method recognition correct rate
		High-pressure conductor metal protrusion	92%
Free metal particulate	91.5%
		Insulator surface fixing metal	88%
Insulator void defects	90%

In sum, the three-phase of the various embodiments described above of the present invention is the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation altogether, comprise step: adopt ultrahigh frequency (UHF) to detect three-phase cartridge type GIS shelf depreciation altogether, utilize UHF sensor to shelf depreciation (PD) signal sampling; Utilize improved small echo threshold values filtering method to carry out denoising Processing to local discharge signal; The characteristic parameter that extracts sampled signal based on phase analysis pattern, characteristic parameter comprises: measure of skewness Sk, steepness Ku, local peaks count Pe, cross-correlation coefficient Cc and discharge factor Q; The feature space that utilizes improved core principle component analysis method to form characteristic parameter carries out dimension-reduction treatment, obtains the characteristic parameter matrix after dimensionality reduction; Utilize k nearest neighbor classification to carry out pattern-recognition to GIS insulation defect type.This three-phase beneficial effect that the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation at least can reach altogether comprises: overcome the defect of prior art, improved the three-phase accuracy of cartridge type UHV (ultra-high voltage) GIS Partial Discharge Detection pattern-recognition altogether.

Finally it should be noted that: the foregoing is only the preferred embodiments of the present invention, be not limited to the present invention, although the present invention is had been described in detail with reference to previous embodiment, for a person skilled in the art, its technical scheme that still can record aforementioned each embodiment is modified, or part technical characterictic is wherein equal to replacement.Within the spirit and principles in the present invention all, any modification of doing, be equal to replacement, improvement etc., within all should being included in protection scope of the present invention.

Claims (10)

1. three-phase is total to a mode identification method for cartridge type UHV (ultra-high voltage) GIS shelf depreciation, it is characterized in that, comprises the following steps:

Step 1: adopt ultrahigh frequency to detect three-phase cartridge type GIS shelf depreciation altogether, utilize UHF sensor to sample to local discharge signal;

Step 2: utilize improved small echo threshold values filtering method to carry out denoising Processing to the local discharge signal collecting, obtain real local discharge signal;

Step 3: by extract the characteristic parameter of sampled signal based on phase analysis pattern algorithm;

Step 4: the feature space that utilizes improved core principle component analysis method to form characteristic parameter carries out dimension-reduction treatment, obtains the characteristic parameter matrix after dimensionality reduction;

Step 5: utilize the k nearest neighbor classification based on bunch thought to carry out pattern-recognition to GIS insulation defect type.

2. three-phase according to claim 1 is total to the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation, it is characterized in that, in step 2, describedly utilize improved small echo threshold values filtering method the local discharge signal collecting to be carried out to the operation of denoising Processing, specifically adopt adaptive thresholding value calculating method; This adaptive thresholding value calculating method is as follows:

$T_j=\frac{\mathrm{Median}\left(\right|C_{j,k}\left|\right)\sqrt{2\ln \left(N_j\right)}}{0.6745}\mathrm{\&alpha;exp}(-\frac{1}{{\mathrm{\&beta;}}_j});$

Wherein, j is yardstick, N _jfor the number of wavelet coefficient on this yardstick, Median (| C _j,k|) be the median of all wavelet coefficients on this yardstick, α is called the signal to noise ratio (S/N ratio) factor, is the embodiment in threshold values calculates of the signal to noise ratio (S/N ratio) of signal, β _jbeing called scale factor, is that the maximal value of wavelet coefficient on yardstick is corrected the evaluated error that sample sequence length difference causes, T _jfor the threshold values calculating.

3. three-phase according to claim 1 and 2 is total to the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation, it is characterized in that, in step 3, described characteristic parameter comprises measure of skewness Sk, steepness Ku, local peaks count Pe, cross-correlation coefficient Cc and discharge factor Q.

4. three-phase according to claim 3 is total to the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation, it is characterized in that, described measure of skewness Sk is specially:

$\mathrm{Sk}=\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{(x_i-\mathrm{\&mu;})}^3\&CenterDot;p_i\mathrm{\&Delta;x}/{\mathrm{\&sigma;}}^3;$

In above formula, w is the phase window number in the semiperiod; Xi is the phase place of i phase window;

$p_i=y_i/\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{y}_i,$

Wherein, y _ibe the ordinate of spectrogram, represent Apparent discharge magnitude q or discharge time n; The position of the shelf depreciation collection of illustrative plates center that parameter μ representative collects, σ represents the precipitous situation that the center axis of symmetry of collection of illustrative plates embodies, Δ x be about one of shelf depreciation collection of illustrative plates to be evenly distributed relevant parameter,

certain in collection of illustrative plates is put corresponding phase place;

Measure of skewness Sk reflection spectrogram shape is with respect to the left and right deflection situation of normal distribution: Sk=0 illustrates that this spectrogram shape is symmetrical; Sk>0 illustrate this spectrogram with respect to normal distribution shape to left avertence; Sk<0 illustrate this spectrogram with respect to normal distribution shape to right avertence.

5. three-phase according to claim 3 is total to the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation, it is characterized in that, described steepness Ku is specially:

$\mathrm{Ku}=[\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{(x_i-\mathrm{\&mu;})}^4{p}_i\mathrm{\&Delta;x}/{\mathrm{\&sigma;}}^4]-3;$

In above formula, w is the phase window number in the semiperiod; Xi is the phase place of i phase window;

$p_i=y_i/\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{y}_i,$

Wherein, y _ibe the ordinate of spectrogram, represent Apparent discharge magnitude q or discharge time n; The position of the shelf depreciation collection of illustrative plates center that parameter μ representative collects, σ represents the precipitous situation that the center axis of symmetry of collection of illustrative plates embodies, Δ x be about one of shelf depreciation collection of illustrative plates to be evenly distributed relevant parameter,

certain in collection of illustrative plates is put corresponding phase place;

Steepness Ku in contrast to the projection degree of normal distribution shape for describing the distribution of certain shape: the steepness Ku of normal distribution is 0; If Ku>0, illustrates that this spectrogram profile is sharply more precipitous than normal distribution profile; If Ku<0, illustrates that this spectrogram profile is more smooth than normal distribution profile.

6. the three-phase according to claim 3 mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation altogether, is characterized in that, the described local peaks Pe that counts, and local peaks is counted for describing the number of local peaks on spectrogram profile; In point

whether place has local peaks, needs to judge with difference equation below:

（y _i-y _i-1）>0，（y _i+1-y _i）<0；

The halved phase window of phase shaft is more, and local peaks is counted larger.

7. three-phase according to claim 3 is total to the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation, it is characterized in that, described cross-correlation coefficient Cc is specially:

$\mathrm{Cc}=\frac{\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{+}{q}_i^{-}-\left(\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{+}\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{-}\right)/W}{\sqrt{[\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{\left(q_i^{+}\right)}^2-{\left(\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{-}\right)}^2/W][\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{\left(q_i^{-}\right)}^2-{\left(\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}{q}_i^{-}\right)}^2]/W}};$

In formula,

be the discharge capacity in phase window i, subscript "+", "-" are corresponding to the positive and negative semiaxis of spectrogram; C has reacted the correlativity of the strong and weak and PHASE DISTRIBUTION of electric discharge in positive and negative half period, and cross-correlation coefficient Cc means close to 1

the profile of spectrogram positive-negative half-cycle is quite similar; Cc is close to 0,

spectrogram profile difference is huge.

8. three-phase according to claim 3 is total to the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation, it is characterized in that, described discharge factor Q is specially:

$Q=\frac{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{-}{q}_i^{-}}{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{-}}/\frac{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{+}{q}_i^{+}}{\underset{i=1}{\overset{w}{\mathrm{\&Sigma;}}}{n}_i^{+}};$

In formula, the electric discharge repetition rate in phase window i, subscript "+", "-" corresponding to

the positive-negative half-cycle of spectrogram.

9. three-phase according to claim 1 and 2 is total to the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation, it is characterized in that, in step 4, described improved core principle component analysis method is improved core principle component analysis method, and the kernel function of sampling is:

$k(x_i,x_j)={(<x_i,x_j>+a)}^b\exp \left(\frac{-{\left|\right|x_i-x_j\left|\right|}^2}{2{\mathrm{\&sigma;}}^2}\right);$

Wherein, (a ∈ R, b ∈ N, σ >0), the selection of parameter a, b and σ is to determine according to the numerical values recited of element in eigenmatrix, parameter σ is for controlling the radial effect scope of kernel function; x _iand x _jrepresent different sample vector ， ﹙ x _i, x _jthe vector product of ﹚ representative sample vector, the span of R representation vector is at set of real numbers, and N represents set of integers, k (x _i, x _j) the new kernel function that obtains in conjunction with the advantage of polynomial kernel function and gaussian kernel function of representative.

10. three-phase according to claim 1 and 2 is total to the mode identification method of cartridge type UHV (ultra-high voltage) GIS shelf depreciation, it is characterized in that, in step 5, the algorithm of described k nearest neighbor classification specifically comprises:

Step1: in training set, first all Partial Discharge Datas are carried out to pre-service and become space vector;

Step2: from first class, to belonging to all signal datas of this classification, carry out between two similarity and calculate, set a minimum threshold, according to statistics, obtain that similarity approaches one by one bunch;

Step3: for each bunch, all signal datas are wherein merged, then calculate its center vector; In addition, compute cluster number/classification sum, this value represents the contribution coefficient of this bunch to this class;

Step4: after new text arrives, carry out the vector space that pre-service obtains it;

Step5: the center vector of every cluster that the space vector of new text and Step3 are generated calculates distance, these distances are multiplied each other with the contribution coefficient of corresponding bunch, the results added that belongs to other bunch of calculating of same class, relatively obtaining that maximum classification is exactly classification under typical defect shelf depreciation to be sorted.

Images