CN106291275A - A kind of local discharge superhigh frequency single waveform frequency domain character extracts and recognition methods - Google Patents

Abstract

The invention discloses a kind of local discharge superhigh frequency single waveform frequency domain character to extract and recognition methods, belong to power equipment Partial Discharge Detecting Technology field, in the method, use the method that fast Fourier and singular value decomposition combine, it is extracted the frequency domain character parameter of local discharge superhigh frequency single waveform, this characteristic vector represents different electric discharge type discharge characteristic on different frequency point, the characteristic vector extracted according to the present invention, utilizing has the recognition methods of supervision can effectively identify shelf depreciation type, as a example by support vector machine, accuracy of identification reaches more than 96%.

Description

A kind of local discharge superhigh frequency single waveform frequency domain character extracts and recognition methods

Technical field:

The invention belongs to power equipment Partial Discharge Detecting Technology field, be specifically related to a kind of local discharge superhigh frequency single Waveform frequency domain character extracts and recognition methods, for the identification of power equipment shelf depreciation type.

Background technology:

The insulant of power equipment is to ensure that the significant components that power equipment is properly functioning, but owing to insulant exists Aging or insulant manufacturing deficiency under highfield effect, there will be local in power equipment runs and puts inside insulant Electricity, the development of shelf depreciation can be accelerated the aging of insulant, thus cause the power equipment lost of life, so must send out as early as possible Now with identify shelf depreciation type, employing measure slows down the aging of power equipment.

Local discharge superhigh frequency detection is by highly sensitive, and the advantages such as on-the-spot anti-jamming effectiveness is good, in power equipment local Discharge examination has obtained relatively broad application.The bandwidth of local discharge superhigh frequency signal, from 0MHz～3000MHz, contains Abundant discharge information, analysis and the feature extraction to hyperfrequency single signal, contribute to the identification of shelf depreciation electric discharge type.

Summary of the invention:

The problem that the invention aims to solve shelf depreciation type identification, it is provided that a kind of local discharge superhigh frequency Single waveform frequency domain character extracts and recognition methods, and demonstrates characteristic parameter effectiveness in shelf depreciation type identification.

For reaching above-mentioned purpose, the present invention adopts the following technical scheme that and realizes:

A kind of local discharge superhigh frequency single waveform frequency domain character extracts and recognition methods, comprises the following steps:

1) many groups local discharge superhigh frequency impulse discharge signal of different shelf depreciation type is gathered；

2) the hyperfrequency impulse discharge signals organizing same electric discharge type are carried out fast Fourier transform more, obtain organizing one more Dimension power spectrum array FreArray, many group One-dimensional power spectrum arrays FreArray constitute the spectral power matrix of same electric discharge type FreMatrix；

3) the spectral power matrix FreMatrix of same electric discharge type is carried out singular value decomposition, take singular value maximum pair The one-dimensional vector answered, obtains the characteristic vector of the spectral power matrix FreMatrix of same electric discharge type, i.e. this feature vector corresponding The characteristic vector of same electric discharge type hyperfrequency single waveform；

4) different electric discharge types are repeated steps 2) and 3) operation, obtain difference electric discharge type hyperfrequency single waveform Characteristic vector；

5) step 4 is utilized) characteristic vector extracted, use the identification of the learning method training local electric discharge type having supervision Model, and utilize identification model that electric discharge type is identified.

The present invention is further improved by, step 2) concrete methods of realizing as follows:

201) hyperfrequency single waveform is used fast Fourier transform, obtain the One-dimensional power spectrum array of single waveform FreArray；

202) multiple hyperfrequency single waveforms of same electric discharge type are repeated 201) step, multiple by FFT FreArray obtains the spectral power matrix FreMatrix of same electric discharge type by row combination, and wherein row represents a single waveform Power spectrum, sample number is shown in list.

The present invention is further improved by, step 5) concrete methods of realizing as follows:

501) obtain the FreMatrix of different electric discharge type, FreMatrix is normalized in the range of [-1,1], it After be classified as test set and training set；

502) method using supervised learning, utilizes training set to be identified model；

503) being taken to by test set in identification model, inspection identifies the effect of model, and according to accuracy of identification, adjusts model Parameter；

504) 502 are repeated) and 503) step, until it reaches meet the accuracy of identification required, i.e. obtain the shelf depreciation of correspondence Hyperfrequency single waveform recognition model.

The present invention is further improved by, step 501) in, FreMatrix is normalized in the range of [-1,1], Formula is as follows:

$y=(ymax-y\min )*\frac{(x-xmin)}{(xmax-x\min )}+y\min ---\left(1\right)$

Wherein, ymax is 1, and ymin is-1, and x is characterized parameter, and xmin is characterized the minima of parameter, and xmax is characterized ginseng The maximum of amount.

The present invention is further improved by, step 502) in, utilize support vector machine to be identified mould according to training set The training of type, obtains the identification model of SVM, and the kernel function of support vector machine is as follows:

K (x, y)=exp (-gamma × | x-y |²) (2)

Wherein, x is characterized parameter, and y is the value that electric discharge type is corresponding, and gmmma is default value 1/ (eigenvalue number).

The present invention is further improved by, and the value of gmmma is 1/2000；Punishment parameter is default value 1.

The present invention contrasts prior art and has a following innovative point:

The mode using fast Fourier transform and singular value decomposition to combine first, is extracted hyperfrequency single waveform Characteristic vector, it is possible to the frequency domain character of the different electric discharge type of reflection directly perceived；

2, use the frequecy characteristic that the present invention extracts, in the case of not to identifying that model carries out parameter optimization, just can reach To the highest accuracy of identification.

The present invention contrasts prior art and has a following remarkable advantage:

1, the frequecy characteristic parameter that the present invention extracts, it is possible to comprise the frequency domain character of ultra-high frequency signal, and energy as far as possible Enough characteristic quantity sizes shown on each Frequency point directly perceived；

2, utilize the characteristic parameter that the present invention extracts, in the case of not Statistical error model parameter, just can reach to compare High accuracy of identification, as a example by support vector machine, discrimination reaches more than 96%.Model Kernel Function is RBF, ginseng Number c and g all uses default value 1 and (1/ eigenvalue number).

In sum, the present invention proposes a kind of new local discharge superhigh frequency single waveform frequency domain character extracting method, Utilize the characteristic parameter extracted in the present invention can represent the frequency domain character of different electric discharge type intuitively, utilize in the present invention and carry The characteristic parameter gone out can effectively carry out shelf depreciation type identification.

Accompanying drawing illustrates:

Fig. 1 is four kinds of shelf depreciation type ultra-high frequency signal single waveform frequency domain character vectors that the inventive method is extracted； Wherein, Fig. 1 (a) is the ultra-high frequency signal single waveform frequency domain character vector of electric discharge type P1, and Fig. 1 (b) is electric discharge type P2's Ultra-high frequency signal single waveform frequency domain character vector, Fig. 1 (c) is the ultra-high frequency signal single waveform frequency domain character of electric discharge type P3 Vector, Fig. 1 (d) is the ultra-high frequency signal single waveform frequency domain character vector of electric discharge type P4.

Detailed description of the invention:

Below in conjunction with accompanying drawing, the present invention is made further instructions.

The basic thought of the present invention is the mode combined based on fast Fourier transform and singular value decomposition, extracts superelevation Frequently the frequency domain character parameter of single waveform, utilizes the characteristic parameter extracted to carry out shelf depreciation type identification, and idiographic flow is as follows:

1) four kinds of electric discharge types (be electric discharge type 1,2,3,4 respectively, hereinafter referred P1, P2, P3 and P4) multiple offices are gathered Portion's electric discharge hyperfrequency single waveform, the oscillograph of use a width of 100MHz～3GHz of its band, its sample rate is 5GS/s；

2) the hyperfrequency single waveform of local electric discharge type P1 is carried out fast Fourier transform (FFT), obtain power spectrum FreArray；

3) many groups hyperfrequency single waveform of local electric discharge type P1 is repeated step 2) operation, obtain organizing hyperfrequency more The FreArray of single waveform, as shown in Fig. 1 (a)；

4) many groups hyperfrequency single waveform of four kinds of electric discharge types P2, P3 and P4 is repeated step 2) and 3), obtain each Spectral power matrix FreMatrix, wherein row represent a single waveform power spectrum, list shows sample number, such as Fig. 1 (b) extremely Shown in (d)；

5) FreMatrix to four kinds of electric discharge types carries out singular value decomposition, a group that extraction singular value maximum is corresponding Vector, obtains the frequency domain character parameter of four kinds of electric discharge type hyperfrequency single waveforms, as shown in Figure 1；

6) utilize the characteristic vector obtained to carry out shelf depreciation identification, specifically comprise the following steps that

A) normalization characteristic parameter, normalizes to characteristic quantity in the range of [-1,1], and formula is as follows:

$y=(ymax-y\min )*\frac{(x-xmin)}{(xmax-x\min )}+y\min ---\left(1\right)$

Wherein, ymax is 1, and ymin is-1, and x is characterized parameter, and xmin is characterized the minima of parameter, and xmax is characterized ginseng The maximum of amount；

B) sample of four kinds of electric discharge types is randomly divided into training set and test set；

C) utilize support vector machine (SVM) to be identified the training of model according to training set, obtain the identification model of SVM, The kernel function of support vector machine is as follows:

K (x, y)=exp (-gamma × | x-y |²) (2)

Wherein, x is characterized parameter, y be value that electric discharge type is corresponding (defined in the present invention: P1 correspondence 1, P2 correspondence 2, P3 Corresponding 3, P4 correspondence 4), gmmma is default value 1/ (eigenvalue number), is 1/2000 in the present invention；Punishment parameter is default value 1；

D) test set being identified validity check, the accuracy of identification obtained is more than 96%.

Claims (6)

1. a local discharge superhigh frequency single waveform frequency domain character extracts and recognition methods, it is characterised in that include following step Rapid:

1) many groups local discharge superhigh frequency impulse discharge signal of different shelf depreciation type is gathered；

2) the hyperfrequency impulse discharge signals organizing same electric discharge type are carried out fast Fourier transform more, obtain organizing one-dimensional merit more Rate spectrum array FreArray, many group One-dimensional power spectrum arrays FreArray constitute the spectral power matrix of same electric discharge type FreMatrix；

3) the spectral power matrix FreMatrix of same electric discharge type is carried out singular value decomposition, take singular value maximum corresponding One-dimensional vector, obtains the characteristic vector of the spectral power matrix FreMatrix of same electric discharge type, i.e. this feature vector correspondence same The characteristic vector of electric discharge type hyperfrequency single waveform；

4) different electric discharge types are repeated steps 2) and 3) operation, obtain the feature of difference electric discharge type hyperfrequency single waveform Vector；

5) step 4 is utilized) characteristic vector extracted, use the identification mould of the learning method training local electric discharge type having supervision Type, and utilize identification model that electric discharge type is identified.

Local discharge superhigh frequency single waveform frequency domain character the most according to claim 1 extracts and recognition methods, its feature Be, step 2) concrete methods of realizing as follows:

201) hyperfrequency single waveform is used fast Fourier transform, obtain the One-dimensional power spectrum array of single waveform FreArray；

202) multiple hyperfrequency single waveforms of same electric discharge type are repeated 201) step, multiple by FFT FreArray obtains the spectral power matrix FreMatrix of same electric discharge type by row combination, and wherein row represents a single waveform Power spectrum, sample number is shown in list.

Local discharge superhigh frequency single waveform frequency domain character the most according to claim 1 extracts and recognition methods, its feature Be, step 5) concrete methods of realizing as follows:

501) obtain the FreMatrix of different electric discharge type, FreMatrix is normalized in the range of [-1,1], afterwards will It is divided into test set and training set；

502) method using supervised learning, utilizes training set to be identified model；

503) being taken to by test set in identification model, inspection identifies the effect of model, and according to accuracy of identification, adjusts model ginseng Number；

504) 502 are repeated) and 503) step, until it reaches meet the accuracy of identification required, i.e. obtain the shelf depreciation superelevation of correspondence Frequently single waveform recognition model.

Local discharge superhigh frequency single waveform frequency domain character the most according to claim 3 extracts and recognition methods, its feature It is, step 501) in, FreMatrix being normalized in the range of [-1,1], formula is as follows:

$y=(ymax-y\min )*\frac{(x-xmin)}{(xmax-x\min )}+y\min ---\left(1\right)$

Wherein, ymax is 1, and ymin is-1, and x is characterized parameter, and xmin is characterized the minima of parameter, and xmax is characterized parameter Maximum.

Local discharge superhigh frequency single waveform frequency domain character the most according to claim 3 extracts and recognition methods, its feature It is, step 502) in, utilize support vector machine to be identified the training of model according to training set, obtain the identification model of SVM, The kernel function of support vector machine is as follows:

K (x, y)=exp (-gamma × | x-y |²) (2)

Wherein, x is characterized parameter, and y is the value that electric discharge type is corresponding, and gmmma is default value 1/ (eigenvalue number).

Local discharge superhigh frequency single waveform frequency domain character the most according to claim 5 extracts and recognition methods, its feature Being, the value of gmmma is 1/2000；Punishment parameter is default value 1.