CN106019090A - Partial discharge electromagnetic wave signal energy feature extraction method - Google Patents

Abstract

The invention discloses a partial discharge electromagnetic wave signal energy feature extraction method, and belongs to the technical field of power equipment partial discharge detection. According to the method, a partial discharge electromagnetic wave signal is decomposed through wavelet packet decomposition, the energy information of the partial discharge electromagnetic wave signal is calculated, and energy features are extracted from the energy information through singular value decomposition for partial discharge recognition. According to the invention, wavelet packet decomposition and singular value decomposition are integrated for the first time in the field of power equipment partial discharge detection, a partial discharge electromagnetic wave signal energy feature extraction method based on wavelet packet decomposition and singular value decomposition is put forward, and the method has the advantage that partial discharge electromagnetic wave energy features can be extracted effectively, and can be used for partial discharge type recognition.

Description

Shelf depreciation electromagnetic wave signal energy feature extraction method

Technical field:

The invention belongs to power equipment Partial Discharge Detecting Technology field, be specifically related to a kind of shelf depreciation electromagnetism Ripple signal energy feature extracting method, 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 due to insulation Material is aging or insulant manufacturing deficiency under highfield effect, in power equipment runs in insulant Portion there will be shelf depreciation, and the development of shelf depreciation can accelerate the aging of insulant, thus causes electric power to set The standby lost of life, so the type of shelf depreciation must be found as early as possible and identify, employing measure slows down electric power and sets Standby is aging.

Pulse current of PD signal is ns level pulse signal, and pulse current is broadband signal, and it excites Uhf electromagnetic wave signal be similarly broadband signal, bandwidth from tens MHz to upper GHz, and based on electricity The Measurement bandwidth of the local discharge superhigh frequency detection of magnetic wave coupling principle can reach tens MHz to upper GHz Frequency, so utilize local discharge superhigh frequency detection can be measured that more horn of plenty shelf depreciation frequency letter Breath.

Owing to the pulse current waveform of different electric discharge types is not quite similar, its electromagnetic wave signal excited comprises Frequency information is the most not all the same, so can be very effectively by extracting these shelf depreciation frequency information features Distinguishing shelf depreciation type, in numerous researchs, the information mainly by different frequency section carries out shelf depreciation class Type identification, also utilizes the mode of wavelet decomposition to extract characteristic parameter, and characteristic parameter has energy feature parameter, Fractal characteristic parameter, also achieves reasonable result.

Type diagnostic is carried out using wavelet decomposition to extract energy feature, but due to the pulse electricity of electric discharge type The difference of stream own is not very big, and the electromagnetic wave signal frequency excited also is more or less the same, in order to obtain effectively The energy feature distinguishing different electric discharge type must carry out multilamellar decomposition, but Decomposition order is many necessarily causes energy Amount parameter becomes geometry multiple to increase.

Summary of the invention:

The invention aims to solve more efficient extraction and be capable of identify that the shelf depreciation of shelf depreciation type The problem of electromagnetic wave energy feature, it is provided that a kind of shelf depreciation electromagnetic wave signal energy feature extraction method, And demonstrate characteristic parameter effectiveness in shelf depreciation type identification.

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

Shelf depreciation electromagnetic wave signal energy feature extraction method, comprises the following steps:

1) the shelf depreciation electromagnetic wave signal collected is carried out WAVELET PACKET DECOMPOSITION, obtain shelf depreciation electromagnetic wave Signal WAVELET PACKET DECOMPOSITION tree；

2) calculate the energy of each node in shelf depreciation electromagnetic wave signal WAVELET PACKET DECOMPOSITION tree, set up local and put The energy matrix E of electricity electromagnetic wave signal；

3) utilize singular value decomposition to calculate the singular value of energy matrix E, be met the singular value of condition to Amount；

4) according to Feature Selection principle, the node energy being more than threshold value in singular value vector is retained, and record Corresponding node, then the characteristic quantity that energy is local discharge signal that node is corresponding；According to the characteristic quantity extracted Shelf depreciation letter is identified, and accuracy of identification has reached more than 90%.

The present invention is further improved by, step 1) in WAVELET PACKET DECOMPOSITION use morther wavelet be ' db2 ', ' db4 ' and ' db8 '.

The present invention is further improved by, step 1) in wavelet packet Decomposition order be 8～10 layers.

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

201) WAVELET PACKET DECOMPOSITION posterior nodal point energy is

$e_x=\frac{1}{N}{x}^{T}x---\left(1\right)$

In formula, x represents the wavelet coefficient of each node, and N represents the length of wavelet coefficient, WAVELET PACKET DECOMPOSITION institute The energy having node constitutes the node energy vector of this signal；

202) energy matrix that wavelet decomposition obtains is carried out for one group of ultra-high frequency signal to be shown below:

$E=\left[\begin{array}{c}\begin{array}{ccccc}{e}_{1,1}& e_{1,2}& e_{1,3}& \mathrm{......}& e_{1,M}\end{array}\\ \begin{array}{ccccc}{e}_{2,1}& e_{2,2}& e_{2,3}& \mathrm{......}& e_{2,M}\end{array}\\ \mathrm{.........................}\\ \begin{array}{ccccc}{e}_{n,1}& e_{n,2}& e_{n,3}& \mathrm{......}& e_{n,M}\end{array}\end{array}\right]---\left(2\right)$

In formula, n represents that each signal decomposition is n node, and M represents the number of UHF signal, e_i,jRepresent The wavelet coefficient energy of the little nodal point of jth ultra-high frequency signal i-th after WAVELET PACKET DECOMPOSITION.

The present invention is further improved by, step 3) concrete methods of realizing as follows:

301) energy matrix E is carried out singular value decomposition, obtain singular value array δ；

302) relative singular value Δ δ is calculated₁₂、Δδ₂₃And Δ δ₃₄, formula is as follows:

$\begin{array}{c}{\mathrm{\Δ\δ}}_{12}=\frac{{\mathrm{\δ}}_1-{\mathrm{\δ}}_2}{{\mathrm{\δ}}_1}\\ {\mathrm{\Δ\δ}}_{23}=\frac{{\mathrm{\δ}}_2-{\mathrm{\δ}}_2}{{\mathrm{\δ}}_2}\\ {\mathrm{\Δ\δ}}_{34}=\frac{{\mathrm{\δ}}_3-{\mathrm{\δ}}_4}{{\mathrm{\δ}}_3}\end{array}---\left(3\right)$

According to relative singular value size, selecting one group or the two groups vector that singular value is maximum, selection principle is:

As Δ δ₁₂During more than 0.7, choose one group of vector that singular value value is maximum；

As Δ δ₁₂Less than 0.7 and Δ δ₂₃During more than 0.7, choose singular value value maximum and second largest two groups to Amount.

The present invention is further improved by, step 4) in, according to singular value feature vector, according to characteristic quantity Selection principle selected characteristic energy, specifically comprise the following steps that

401) will be greater than the amount of maximal eigenvector coefficient 5% as energy feature parameter；

402) owing to, in wavelet decomposition, two child nodes constitute a parent node, in order to eliminate characteristic quantity Redundancy, if a parent node is characteristic parameter, and its two child nodes are also characteristic parameters, The most only retain parent node as energy feature parameter.

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

1. by adding up a large amount of shelf depreciation electromagnetic wave signal features, it is determined that singular value feature vector selects The threshold value of characteristic quantity；

2. WAVELET PACKET DECOMPOSITION is combined with singular value decomposition, extracts shelf depreciation electromagnetic wave from whole small echo seeds The characteristic quantity of signal, it is possible to extract characteristic quantity fully and effectively, having reached the less characteristic parameter of use can Efficiently differentiate the electromagnetic wave signal that different electric discharge type excites.

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

1, contrasting original method extracting characteristic parameter from the wavelet packet tree bottom, the present invention is from the whole tree of wavelet packet Middle selected characteristic parameter, will not omit effective characteristic parameter；

2, utilize singular value decomposition that wavelet packet tree energy matrix is decomposed, it is possible to effectively to reduce redundancy parameter, Improve the effect that shelf depreciation electromagnetic wave characteristics is extracted.

In sum, the present invention effectively solves to extract effective characteristic parameter from substantial amounts of energy parameter Problem, the present invention uses tree complete to wavelet tree to search and two kinds of methods of singular value decomposition carry out Characteristic Extraction

Accompanying drawing illustrates:

Fig. 1 is the schematic diagram of the inventive method WAVELET PACKET DECOMPOSITION posterior nodal point energy；

Fig. 2 is the schematic diagram of the inventive method energy node singular value；

Fig. 3 is the schematic diagram that the inventive method characteristic energy selects, and circular Lycoperdon polymorphum Vitt is the energy feature pair selected The wavelet packet node serial number answered.

Detailed description of the invention:

The basic thought of the present invention is to utilize fusion WAVELET PACKET DECOMPOSITION to realize shelf depreciation electromagnetism with singular value decomposition Effective extraction of wave energy signal, idiographic flow is as follows:

1) local discharge superhigh frequency signal is gathered, the oscillograph of use a width of 100MHz～3GHz of its band, its Sample rate is 5GS/s, and the sensor of use is microstrip antenna sensor, and it carries a width of 100MHz～6000MHz；

2) the ultra-high frequency signal sample of different shelf depreciation types is carried out 8 layers of wavelet decomposition, WAVELET PACKET DECOMPOSITION The morther wavelet used is ' db2 ', ' db4 ', and ' db8 ', the WAVELET PACKET DECOMPOSITION number of plies is 8～10 layers；

3) seek the energy of each node in each signal wavelet decomposition tree, constitute the energy vectors of this signal；

Calculating each wavelet packet component energy is

$e_x=\frac{1}{N}{x}^{T}x---\left(1\right)$

In formula, x represents the wavelet coefficient of each node, and N represents the length of wavelet coefficient, by all wavelet packets The energy of node constitutes the node energy vector of this signal；

$E=\left[\begin{array}{c}\begin{array}{ccccc}{e}_{1,1}& e_{1,2}& e_{1,3}& \mathrm{......}& e_{1,M}\end{array}\\ \begin{array}{ccccc}{e}_{2,1}& e_{2,2}& e_{2,3}& \mathrm{......}& e_{2,M}\end{array}\\ \mathrm{.........................}\\ \begin{array}{ccccc}{e}_{n,1}& e_{n,2}& e_{n,3}& \mathrm{......}& e_{n,M}\end{array}\end{array}\right]---\left(2\right)$

Wherein, n represents that each signal decomposition is n node, and M represents the number of UHF signal, e_i,jRepresent The wavelet coefficient energy of the little nodal point of jth ultra-high frequency signal i-th after WAVELET PACKET DECOMPOSITION；

4) calculate the singular value of shelf depreciation electromagnetic wave signal energy matrix, select according to relative singular value size The singular value vector that singular value is maximum and second largest；

Singular value computing formula relatively is:

$\begin{array}{c}{\mathrm{\Δ\δ}}_{12}=\frac{{\mathrm{\δ}}_1-{\mathrm{\δ}}_2}{{\mathrm{\δ}}_1}\\ {\mathrm{\Δ\δ}}_{23}=\frac{{\mathrm{\δ}}_2-{\mathrm{\δ}}_2}{{\mathrm{\δ}}_2}\\ {\mathrm{\Δ\δ}}_{34}=\frac{{\mathrm{\δ}}_3-{\mathrm{\δ}}_4}{{\mathrm{\δ}}_3}\end{array}---\left(3\right)$

Selection gist is:

As Δ δ₁₂During more than 0.7, choose one group of vector that singular value value is maximum；

As Δ δ₁₂Less than 0.7 and Δ δ₂₃During more than 0.7, choose singular value value maximum and second largest two groups to Amount.

5) obtain singular value vector, select wherein to meet the node energy of condition, then corresponding node energy is little Ripple packet node is characteristic of correspondence amount node, and feature extraction is according to as follows:

A) will be greater than the amount of maximal eigenvector coefficient 5% as energy feature parameter；

B) owing to, in wavelet decomposition, two child nodes constitute a parent node, in order to eliminate the superfluous of characteristic quantity Remaining information, if a parent node is characteristic parameter, and its two child nodes are also characteristic parameters, that Only retain parent node as energy feature parameter；

6) according to the energy feature amount that present invention determine that, utilize the most conventional identification model (support vector machine, Neutral net and minimum neighbours' principle), the highest shelf depreciation type identification accuracy can be obtained, below 4 tables are the different recognition accuracies to four kinds of electric discharge types identifying model.

Table 1 support vector machine test result

Table 2 neutral net test result (Mse=0.05, neurons=75)

Table 3 neutral net test result (MSE=0.01, Mse=0.01；Neurons=350)

The minimum neighbours' principle test result of table 4

Claims (6)

1. shelf depreciation electromagnetic wave signal energy feature extraction method, it is characterised in that include following step Rapid:

1) the shelf depreciation electromagnetic wave signal collected is carried out WAVELET PACKET DECOMPOSITION, obtain shelf depreciation electromagnetic wave Signal WAVELET PACKET DECOMPOSITION tree；

2) calculate the energy of each node in shelf depreciation electromagnetic wave signal WAVELET PACKET DECOMPOSITION tree, set up local and put The energy matrix E of electricity electromagnetic wave signal；

3) utilize singular value decomposition to calculate the singular value of energy matrix E, be met the singular value of condition to Amount；

4) according to Feature Selection principle, the node energy being more than threshold value in singular value vector is retained, and record Corresponding node, then the characteristic quantity that energy is local discharge signal that node is corresponding.

Shelf depreciation electromagnetic wave signal energy feature extraction method the most according to claim 1, its feature Be, step 1) in WAVELET PACKET DECOMPOSITION use morther wavelet be ' db2 ', ' db4 ' and ' db8 '.

Shelf depreciation electromagnetic wave signal energy feature extraction method the most according to claim 1, its feature Be, step 1) in wavelet packet Decomposition order be 8～10 layers.

Shelf depreciation electromagnetic wave signal energy feature extraction method the most according to claim 1, its feature Be, step 2) concrete methods of realizing as follows:

201) WAVELET PACKET DECOMPOSITION posterior nodal point energy is

$e_x=\frac{1}{N}{x}^{T}x---\left(1\right)$

In formula, x represents the wavelet coefficient of each node, and N represents the length of wavelet coefficient, WAVELET PACKET DECOMPOSITION institute The energy having node constitutes the node energy vector of this signal；

202) energy matrix that wavelet decomposition obtains is carried out for one group of ultra-high frequency signal to be shown below:

In formula, n represents that each signal decomposition is n node, and M represents the number of UHF signal, e_i,jRepresent The wavelet coefficient energy of the little nodal point of jth ultra-high frequency signal i-th after WAVELET PACKET DECOMPOSITION.

Shelf depreciation electromagnetic wave signal energy feature extraction method the most according to claim 1, its feature Be, step 3) concrete methods of realizing as follows:

301) energy matrix E is carried out singular value decomposition, obtain singular value array δ；

302) relative singular value Δ δ is calculated₁₂、Δδ₂₃And Δ δ₃₄, formula is as follows:

${\mathrm{\Δ\δ}}_{12}=\frac{{\mathrm{\δ}}_1-{\mathrm{\δ}}_2}{{\mathrm{\δ}}_1}$

${\mathrm{\Δ\δ}}_{23}=\frac{{\mathrm{\δ}}_2-{\mathrm{\δ}}_3}{{\mathrm{\δ}}_2}---\left(3\right)$

${\mathrm{\Δ\δ}}_{34}=\frac{{\mathrm{\δ}}_3-{\mathrm{\δ}}_4}{{\mathrm{\δ}}_3}$

According to relative singular value size, selecting one group or the two groups vector that singular value is maximum, selection principle is:

As Δ δ₁₂During more than 0.7, choose one group of vector that singular value value is maximum；

As Δ δ₁₂Less than 0.7 and Δ δ₂₃During more than 0.7, choose singular value value maximum and second largest two groups to Amount.

Shelf depreciation electromagnetic wave signal energy feature extraction method the most according to claim 1, its feature It is, step 4) in, according to singular value feature vector, according to the selection principle selected characteristic energy of characteristic quantity Amount, specifically comprises the following steps that

401) will be greater than the amount of maximal eigenvector coefficient 5% as energy feature parameter；

402) owing to, in wavelet decomposition, two child nodes constitute a parent node, in order to eliminate characteristic quantity Redundancy, if a parent node is characteristic parameter, and its two child nodes are also characteristic parameters, The most only retain parent node as energy feature parameter.