Note : Les descriptions sont présentées dans la langue officielle dans laquelle elles ont été soumises.
CA 02918679 2016-01-19
Pattern Recognition Method for Partial Discharge of Three-phase Cylinder Type
Ultrahigh Voltage GIS
Technical field
The invention involves the technical field of high voltage discharge
identification, and
specifically involves a pattern recognition method for partial discharge of a
three-phase
cylinder type ultrahigh voltage GIS.
Background technology
Gas Insulated Switchgear (hereafter referred to as GIS) is one of the
important
equipments in the ultra-high voltage grid. After optimization design,
breakers, current
transformers, voltage transformers, lightening arresters, disconnecting
switches, earthing
switches, buses, cable terminals, casings of inlet and outlet line and other
parts are
respectively installed in corresponding sealed compartments, and finally
assembled in an
integral casing taking SF6 as insulating medium.
Main defects influencing the performance of insulating medium in GIS include
serious
installation error, poor contact between conductors, high voltage conductor
protrusions, fixed
particles, insulator defect, steam, etc..
The development of GIS tends to be three-phase cylinder type, compounding and
intelligence. Due to the realization of miniaturization, the general assembly
can be completed
in the factory. After passing the test, it is transported to the site in the
form of interval.
Therefore, it can reduce the installation period, and the reliability is
improved.
In terms of internal structure, electric field distribution and other aspects,
three-phase
cylinder type GIS is obviously different from coaxial type GIS. The existing
technology
research mainly focuses on coaxial type GIS, but there is less research on the
detection
pattern recognition of partial discharge of three-phase cylinder type
ultrahigh voltage GIS.
Itt the process of realization of the invention, the inventor found that the
existing
CA 02918679 2016-01-19
technology at least had fewer functions, small scope of application, poor
accuracy and other
defects.
Content of the Invention
The purpose of the invention is to propose a pattern recognition method for
partial
discharge of three-phase cylinder type ultrahigh voltage GIS according to the
above problems,
so as to realize the advantages of many functions, wide scope of application
and good
accuracy.
In order to achieve the above purpose, the invention uses the technical
solution of a
pattern recognition method for partial discharge of three-phase cylinder type
ultrahigh voltage
GIS, including the following steps:
Step 1: The ultrahigh frequency is used to detect the partial discharge of
three-phase
cylinder type GIS, and UHF sensor is used to sample the partial discharge
signal;
Step 2: The improved wavelet threshold filtering method is used for denoising
processing of the partial discharge signal collected, and the real partial
discharge signal is
obtained;
Step 3: The characteristic parameters of sampling signal are extracted through
the mode
algorithm based on the phase analysis;
Step 4: The improved kernel principal component analysis method is used for
dimension
reduction processing of characteristic space consisting of characteristic
parameters, and the
characteristic parameter matrix after dimension reduction is obtained; and
Step 5: K nearest neighbor algorithm based on the cluster idea is used for
pattern
recognition of insulation defect type of GIS.
Further, in Step 2, the improved wavelet threshold filtering method is used
for denoising
processing of the partial discharge signal collected. The calculation method
of self-adaptive
threshold is used, which is as follows:
2
CA 02918679 2016-01-19
T 1
Media*Jai 1V2IncV.) \ =
exp
0.6745
\. );
=
N.
In which, j is the scale. is the number of
wavelet coefficients on the scale.
"fan(ICA is the median of all the wavelet coefficients on the scale. a' is
called the
signal-to-noise ratio factor and is the signal-to-noise ratio of signal in the
threshold
calculation. A is called scale factor and is the estimated error caused when
the maximum
T
value of wavelet coefficients on the scale corrects the different sampling
sequence length.
is the calculated threshold value.
Further, the characteristic parameters in Step 3 include degree of skewness
(Sk), degree
of steepness (Ku), the number of partial peak points (Pe), cross correlation
coefficient (Cc)
and discharge factor (Q).
Further, the degree of skewness (Sk) is as follows.
N3
Sk E (x, mu) = p I cr3
In which, w is the number of phase window in the half cycle; Xi is the phase
position of
the i phase window;
w
PYiIYj
E ptco, = EP; (01-11)2
In which, =)) is the vertical coordinate of spectrum, representing apparent
discharge
magnitude (q) or number of discharges (n); parameter represents the
central position of
partial discharge map collected, cr represents the steepness of symmetry axis
at the center of
the map. L1X is a parameter related to even distribution of partial discharge
map, and g9i is
the phase position corresponding to a point in the map;
The degree of skewness (Sk) reflects the skewness of spectrum shape relative
to normal
distribution. Sk=0 indicates that the spectrum shape is symmetrical; Sk >0
indicates that the
3
CA 02918679 2016-01-19
spectrum biases to the left relative to normal distribution; Sk<0 indicates
that the spectrum
biases to the right relative to normal distribution.
Further, the degree of steepness (Ku) is as follows.
Ku = - p idx u4 - 3
In which, w is the number of phase windows in half cycle; Xi is the phase
position of the
=th
phase window;
if _______
= E = pi(p, 0-= Epi(9i
i=1
In which, Yi is the vertical coordinate of spectrum, representing apparent
discharge
magnitude (q) or number of discharges (n); parameter represents the central
position of
partial discharge map collected, 0" represents the steepness of symmetry axis
in the center of
the map, AX is a parameter related to even distribution of partial discharge
map, and CI is
the phase position corresponding to a point in the map;
The degree of steepness (Ku) is used to describe the protruding degree of
distribution of
a certain shape relative to the normal distribution. The degree of steepness
of normal
distribution is 0. If Ku > 0, it indicates that the contour of the spectrum is
sharper and steeper
than that of the normal distribution; if Ku<0, it indicates that the contour
of the spectrum is
flatter than that of the normal distribution.
Further, the number of partial peak points (Pe) is used to describe the number
of partial
peak points on the contour of the spectrum; whether there is partial peak at
contour point
needs to be determined with the following difference equation.
(Y1¨ Yi¨i) >0, (Yi4-1¨ Y1) <0;
The more the phase windows isometrically divided from the phase axis are, the
larger the
number of partial peak points is.
4
CA 02918679 2016-01-19
Further, the cross correlation coefficient (Cc) is as follows.
E 7,741 ¨ Eq-,¶E iw
i=1
Cc.
w IV/ W \ 2
Al (41 Eql rw [E(q;:-.)2
\ 1=1 _L " i.1 1
q q
In which, is the discharge
quantity in the phase window i. The superscripts
"+" and "¨" respectively corresponds to the positive and negative semiaxis of
the spectrum; c
reflects the correlation betweem the discharge strength and phase distribution
in the positive
and negative half cycle. If the cross correlation coefficient (Cc) is close to
I, it indicates that
the contours of "61 spectrum of the
positive and negative half cycle are quite similar; if Cc
is close to 0, the contour difference of 9 ¨ ("spectrum is great.
Further, the discharge factor (Q) is as follows:
n, qi E 11 1.+ q
= 1
Q
n n
= =
'7
In which, ni+ and are discharge repetition rate in the phase window i. The
superscripts "+" and "¨" respectively correspond to the positive and negative
half cycle of the
q spectrum.
Further, the improved kernel principal component analysis method in Step 4 is
the
improved kernel principal component analysis method. The kernel function
sampled is as
follows.
2'
k(X. X )= X. > exp ______
2 a
;
CA 02918679 2016-01-19
In which, in (a E R,b N,t7 > 0), parameter a, b and are selected
according to the
value of elements in the characteristic matrix. The parameter (3 is used to
control the range
of action in the radial direction of the kernel function; Xi and Xl represent
different sample
x.
vectors, < ) represents the
vector product of sample vectors, R represents the set of
real numbers of the value range of vectors, N represents the set of integers,
and kxj)
represents a new kernel function obtained by combining the advantages of
polynomial kernel
function and gaussian kernel function.
Further, K nearest neighbor algorithm in Step 5 includes:
Step I : In the training set, first all the partial discharge data is
preprocessed and mapped
into a spatial vector;
Step 2: From the first category, the similarity of each two data among all the
signal data
in the category is calculated, the minimum threshold is set, and according to
the statistics the
clusters with close similarity are obtained;
Step 3: All the signal data in each cluster is merged. And then its central
vector is
calculated; in addition, the number of clusters/categories is calculated. This
value represents
the contributing coefficient of the cluster to the category;
Step 4: The new text is preprocessed, and its vector space is obtained;
Step 5: The distance between the spatial vector of new text and the central
vector of each
cluster generated in Step 3 is calculated. These distances are multiplied by
the contributing
coefficient of corresponding clusters. The calculated results of clusters in
the same category
are added. After comparison, the biggest category obtained is the category of
partial discharge
with typical defect to be classified.
The pattern recognition method for partial discharge of three-phase cylinder
type
ultrahigh voltage GIS of the embodiments of the invention has the following
steps. The
ultrahigh frequency is used to detect the partial discharge of three-phase
cylinder type GIS,
and UHF sensor is used to sample the partial discharge signal; the improved
wavelet threshold
6
CA 02918679 2016-01-19
filtering method is used for denoising processing of the partial discharge
signal collected, and
the real partial discharge signal is obtained; the characteristic parameters
of sampling signal
are extracted through the mode algorithm based on the phase analysis; the
improved kernel
principal component analysis method is used for dimension reduction processing
of
characteristic space consisting of characteristic parameters, and the
characteristic parameter
matrix after dimension reduction is obtained; and K nearest neighbor algorithm
based on the
cluster idea is used for pattern recognition of insulation defect type of GIS.
Therefore,
existing technical defects can be overcome, and the accuracy of the detection
pattern
recognition of partial discharge of three-phase cylinder type ultrahigh
voltage GIS is
improved; the defect of less functions, small scope of application and poor
accuracy in the
existing technology can be overcome, and the advantages of many functions,
wide scope of
application and good accuracy can be realized.
Other features and advantages of the invention will be introduced in the
subsequent
description, and part of them are obvious from the description or known
through
implementation of the invention.
Next, through drawings and embodiments, the technical solution of the
invention is
future introduced in details.
BRIEF DESCRIPON OF THE DRAWINGS
The drawings are used for further understanding of the invention, constitute a
part of
the description, and are combined with embodiments of the invention to explain
the
invention without restricting the scope of the inventionthe invention. In the
drawings:
Figure I is the structure diagram of partial discharge test apparatus of GIS
in the pattern
recognition method for partial discharge of three-phase cylinder type
ultrahigh voltage GIS in
the invention;
Figure 2 is flow diagram of the pattern recognition method for partial
discharge of
three-phase cylinder type ultrahigh voltage GIS in the invention. In Figure 2,
N and K are
both natural numbers;
7
CA 02918679 2016-01-19
Figure 3 is the schematic diagram of the relationship between statistical
distribution, Sk
and Ku in the pattern recognition method for partial discharge of three-phase
cylinder type
ultrahigh voltage GIS in the invention. (a) is positive deflection, (b) is no
deflection, and (c) is
negative deflection. Sk is the deflection of spectrum shape relative to the
normal distribution;
(d) is positive deflection, (e) is no deflection, and (f) is negative
deflection. The degree of
steepness (Ku) is used to describe the protruding degree of distribution of a
certain shape
relative to the normal distribution;
Figure 4 is the schematic diagram of effect of kernel function in the pattern
recognition
method for partial discharge of three-phase cylinder type ultrahigh voltage
GIS in the
invention. (a) is a polynomial function, (b) is a gaussian kernel function ,
and (c) is a new
kernel function;
Figure 5 is the waveform comparison chart before filtering (a) and after
filtering (b) in
the pattern recognition method for partial discharge of three-phase cylinder
type ultrahigh
voltage GIS in the invention. In Figure 5, the angle of horizontal axis is 0-
360 degrees. The
vertical axis is the signal amplitude (Q).
Combined with the drawings, the reference numbers in embodiments of the
invention are
as follows.
1-water resistance; 2-HT testing transformer; 3-transformer; 4-HV bushing; 5-
disk
insulator.
DETAILED DESCRIPTION
Below is the description of preferred embodiments of the invention combined
with the
drawings. It should be understood that the preferred embodiments described
here are only
used to describe and explain the invention, but not to limit the invention.
For the defects in the existing technology, according to the embodiments of
the invention,
as shown in Figure I-Figure 5, a pattern recognition method for partial
discharge of
three-phase cylinder type ultrahigh voltage GIS is provided.
As shown in Figure 1, the partial discharge test apparatus of GIS used in the
pattern
8
CA 02918679 2016-01-19
recognition method for partial discharge of three-phase cylinder type
ultrahigh voltage GIS in
the invention includes disk insulators 5, an HV bushing 4 installed on the
disk insulators 5,
and a water resistance I, an HT testing transformer 2 and a transformer 3
connected to the HV
bushing 4 in turn, and a PDSG connected to the disk insulator 5, and the
common end of the
water resistance 1 and the FTV bushing 4 is earthed after passing through a
voltage divider
consisting of capacitors.
The partial discharge test apparatus of GIS mainly includes a transformer 3, a
voltage
divider consisting of capacitors, an oscilloscope, a three-phase cylinder type
GIS, sensors and
a PDSG; by setting high voltage conductor metal protrusions, free metal
particles, fixed
metals on the surface of insulator, air gaps of insulator and other defects
respectively in GIS,
the corresponding partial discharge signal is detected and the mode is
recognized.
The pattern recognition method for partial discharge of three-phase cylinder
type
ultrahigh voltage GIS in this embodiment includes the following steps:
1) Ultrahigh frequency (UHF) is used to detect the partial discharge of three-
phase
cylinder type GIS, and UHF sensors are used to sample the partial discharge
(PD) signal;
In Step 1), the ultrahigh frequency (UHF) method is used to detect the partial
discharge
of three-phase cylinder type GIS. When UHF method is used to detect the
partial discharge of
GIS, different failure types can be recognized according to the spectral
characteristic of
signals measured and the position of discharge on the power voltage waveform.
In Step 1), the real three-phase cylinder type GIS is used to measure the
partial discharge
map under the typical defect conditions. Among them, the three-phase cylinder
type GIS shall
be provided by a professional high voltage switchgear enterprise. The three-
phase conductor
is energized as follows. Two phases are earthed, and the other one phase is
connected to the
high voltage. In the model there is one HV bushing. According to the test
results of coaxial
type GIS test model, the contrastive typical insulation defects are set. In
the cavity of the
three-phase cylinder type GIS, there are free metal particles. On the surface
of the insulator,
there are fixed metal particles. On the insulator, there is air gap defect.
The physical model of
partial discharge is set in the GIS stimulated device for the measurement of
partial discharge.
9
CA 02918679 2016-01-19
2) The improved wavelet threshold filtering method is used for denoising
processing of
the partial discharge signal collected, and the real partial discharge signal
is obtained;
In Step 2), in the improved wavelet threshold filtering method, from the
calculation
method of Median (Cj, k), it can be seen that the selection of threshold is
closely related to
the length of analyzed signal. In actual application, we cannot ensure that
the proportion and
position of the number of points sampled in the entire section of the
effective signal in the
entire sampling sequence is unchanged. Because the length of signal determines
the number
of wavelet coefficients on each scale after the wavelet changes (1\1j), it
affects the value of
Median ( g, 10, so as to affect the threshold. This influence will lead to the
following
results. The same useful signal sequence (u) is contained in a sampling
sequence (s). If the
length of s is different (i.e. the width of the time observation window is
different), the results
after the wavelet threshold filtering will be quite different. For s of
different lengths
containing the same UHF PD signal, the results of soft threshold filtering
algorithm are
applied. The sampling frequency of the original signal is 20GHz. The long
sampling sequence
with wide time window has 50,000 points. The short sampling sequence with
narrow time
window has 16,000 points. The two sequences both completely contain the useful
UHF PD
signal. Obviously, after soft threshold filtering algorithm is applied, UHF PD
signal,
especially the part of shock attenuation that will end, has obvious
difference. The correlation
coefficient of the two is 0.6677.
The primary cause of these consequences is that the calculation formula of
threshold
completely ignores the amplitude of effective signal and signal-to-noise
ratio. Therefore, after
several analysises, the embodiment considers the amplitude of effective signal
and
signal-to-noise ratio, and proposes a new self-adaptive threshold calculation
method, which
greatly reduces the level of sensitivity of wavelet threshold filtering
results to the sampling
points. The comparison of figures before and after filtering is as shown in
Figure 5. In Figure
5, (a) is the graph before filtering, and (b) is the graph after filtering.
Arledia4Cbk 10 in ______ (N ( \
=
a exp
0.6745
\. i;
CA 02918679 2016-01-19
In which, j is the scale. Ali is the number of wavelet coefficients on the
scale.
tt1edia4C =
is the median of all the wavelet coefficients on the scale, a called the
signal to
noise factor is the signal-to-noise ratio of signal in the threshold
calculation. A called scale
factor is the estimated error caused when the maximum value of wavelet
coefficients on the
scale corrects the different sampling sequence length.
-/ is the calculated threshold value.
The effectiveness of the wavelet denoising method mainly depends on a wavelet
primary
function, a wavelet decomposition scale, a threshold function, threshold
selection and other
aspects. In the embodiment, a large number of simulation experiments,
laboratory simulation
and field measured data are used to analyze and verify the effectiveness of
the method used.
The results show that compared with the denoising method with other threshold
rules, the
wavelet denoising method obviously improves the denoising ability in the
partial discharge
signal processing, and has advantages such as small distortion of signal
waveform after the
processing, more accurate extraction and few influencing factors.
3) Characteristic parameters of sampling signal are extracted with the phase
analysis
mode algorithm. Preferably, the parameters include degree of skewness (Sk),
degree of
steepness (Ku), the number of partial peak points (Pe), cross correlation
coefficient (Cc) and
discharge factor (Q).
In Step 3), for the signal after sampling, characteristic parameters are
extracted with
phase resolved partial discharge (PRPD) mode. Among them, the definition of
degree of
skewness (Sk) is as follows:
Sk = 143 = piAx (73
=
Jr..]
In which, w is the number of phase window in the half cycle; xi is the phase
position of
the VI' phase window;
11.
CA 02918679 2016-01-19
w H1
=ylYi Al =I p ico, = (9i JO2
1=1 V i=1
In which, Yi is the vertical coordinate of spectrum, representing apparent
discharge
magnitude (q) or number of discharge (n); parameter '41 represents the central
position of
partial discharge map collected, .7 represents the steepness of symmetry axis
in the center of
the map, 11X is a parameter related to even distribution of partial discharge
map, and is
the phase position corresponding to a point in the map;
The degree of skewness (Sk) reflects the skewness of spectrum shape relative
to normal
distribution. Sk=0 indicates that the spectrum shape is symmetrical; Sk > 0
indicates that the
spectrum biases to the left relative to normal distribution; Sk<0 indicates
that the spectrum
biases to the right relative to normal distribution.
Definition of the phase window: Construction method of ci"qqlspace curved
surface:
The power frequency phase is divided into 256 sections according to 0-360
degrees. The
discharge pulse amplitude (q) is divided into 128 sections by maximum
amplitude, so that
"(I plane is divided into 128x256 sections; the number of discharge in each
section on
tt) -q plane is counted, and '1)'*(1-11 space curved face is obtained.
The degree of steepness (Ku) is defined as below.
Ku = E(x, -1.1)4 p,Ax 1 a4 ¨ 3
In which, the definition of each variable is the same as that of variable in
the degree of
skewness. The degree of steepness (Ku) is used to describe the protruding
degree of
distribution of a certain shape relative to the normal distribution. The
degree of steepness of
normal distribution is 0. If Ku > 0, it indicates that the contour of the
spectrum is sharper and
steeper than that of the normal distribution; if Ku<0, it indicates that the
contour of the
spectrum is flatter than that of the normal distribution.
12
CA 02918679 2016-01-19
The number of partial peak points (Pe) is used to describe the number of
partial peak
points on the contour of the spectrum; whether there is partial peak at
contour point
needs to be determined with the following formula:
d dYi.1 <0
Y i--, ., 0
do
,
and
The above formula is converted into the following difference equation.
Yi - Yi-i Yi-1-1 - Y=
i
q)/ - Cpi-i >0, Pi+1 - q)i <0;
The difference equation can be simplified as below.
_ v
( .7 v i ) >0, ( Yi i
= -=
Y' ) <0;
In the actual calculation, the number of partial peak points is closely
related to the
number of phase windows of the spectrum. Generally, the more the phase windows
isometrically divided from the phase axis are, the larger the number of
partial peak points is.
The cross correlation coefficient (Cc) is defined as below:
w (ii' w \
Eteqi- - EqTE q7 1W
CC
i..1 k,e--1 itt i I
= _
Eqi) ify E(0 _1( zq- iff,
1
In which, q: ' is the discharge quantity in the phase window i. The
superscripts
"+" and "¨" respectively correspond to the positive and negative semiaxis of
the spectrum; Cc
reflects the correlation between the discharge strength and phase distribution
in the positive
and negative half cycle. If the cross correlation coefficient (Cc) is close to
1, it indicates that
the contour of 9-(1. spectrum of the positive and negative half cycle is quite
similar; if Cc is
13
CA 02918679 2016-01-19
aVe
close to 0, the contour difference of spectrum is great.
The discharge factor (Q) is as follows:
n
n q
Q= , =
ri n
=
n
n
In which, and l are discharge repetition rate in the phase window i. The
superscripts "+" and "¨" respectively corresponds to the positive and negative
half cycle of
the (I) ¨ qspectrum. The discharge quantity factor (Q) reflects the difference
of average
discharge quantity within the positive and negative half cycle of the ¨ '1
spectrum.
According to the above formula of statistical operator, through analysis
spectrum and
spectrum the statistical operator is calculated. The characteristic parameters
degree of
skewness (SK), degree of steepness (Ku), the number of partial peak points
(Pe), discharge
factor (Q) and cross correlation coefficient (Cc) are extracted for the
pattern recognition.
4) The improved kernel principal component analysis method is used for
dimension
reduction processing of characteristic space consisting of characteristic
parameters, and the
characteristic parameter matrix after dimension reduction is obtained;
In Step 4), we do not know which characteristic parameters can construct the
simplest
characteristic space of UHF PD signal in advance, namely the non-redundant
full rank
characteristic parameters matrix, the constructed characteristic space has
large number of
dimensions, and there may be redundant number of dimensions, which is bad for
the
operation and recognition results, so the dimension reduction processing of
characteristic
space is required.
In the application of KPCA, the selection of nonlinear transformation (i.e.
kernel
function) is very important. The commonly used kernel functions include
polynomial kernel
function, Gaussian kernel function and Sigmoid kernel function as shown below:
14
CA 02918679 2016-01-19
k(X= Xj )= ("< Xi, X = > a)b
(
k(x,,xj)= exp XAI2
a"
;
k(Xi , X .)= tanh(< xõ x > +a) (a e R)
In which, < Xi is the sample vector, which is the vector product of
Xiand Xi.
11x,
is Euclidean norm of the two. The polynomial kernel function is the bth power
of
, .
distance ("o(<,%> +a) , and the monotonic increasing function of < x1 X >
After
conversion, if (<xõ..r., >+a)>1, the original distance <x oxl >will be
magnified; on the
contrary, if (<xoxt >+a)<1, the original distance <i> will be reduced. It can
be seen
that the effect of polynomial kernel function is reduction of small distance
and further
increase of large distance. Gaussian kernel function, also known as radial
basis kernel
function, is usually defined as the index monotonic decline function of two
vector Euclidean
distances. It is a kind of scalar function with radial symmetry. Among them, 0
is called the
width parameter used to control the radial range of action of the function,
namely width of
gaussian pulse. But usually the range of action of Gaussian kernel function is
small. The
effect is just opposite to that of polynomial kernel function, namely
increasing small distance
and decreasing big distance.
Actually the effect of kernel function shall be further increase of the
original distance, or
reduction of the distance between similar samples and increase of the distance
between
different samples. Therefore, this embodiment combines the advantages and
disadvantages of
polynomial kernel function and gaussian kernel function, and proposes a new
kernel function,
whose effect is as shown in Figure 4.
In Step 4), the kernel function used by the improved kernel principal
component analysis
method is as follows:
CA 02918679 2016-01-19
2i
- X -
J
k X = X X > +4' exp
20'2
In which, in (a E R,b > 0), parameter
a, b and 0 are selected according to the
value of elements in the characteristic matrix. The parameter 0 is used to
control the range
of action in the radial direction of the kernel function: Xi and Xi represent
different sample
xi xi
vectors, ( represents the
vector product of sample vectors, R represents the set of
kkxj)
real numbers of the value range of vectors, N represents the set of integers,
and
represents the new kernel function obtained by combining the advantages of
polynomial
kernel function and gaussian kernel function. Here, a=5 and b=1, so that the
distance after
conversion changes proportionally with the original distance. The parameter 0
is used to
control the range of action in the radial direction of the kernel function.
The distance between
two vectors in the original characteristic matrix is generally not more than
7, so 0 =7. It can
be seen from Figure 4 that the new kernel function proposed in the invention
increases the
original small distance and appropriately reduces the big distance.
5) K nearest neighbor classification method based on the cluster idea is used
for pattern
recognition of insulation defect type of GIS:
In Step 5), the main idea of K nearest neighbor classification method is as
follows. The
test documents are given. The system searches for K nearest neighbors in the
classified
training set. According to the category distribution of these neighbors, the
categories of the
test documents are obtained. The similarity of these neighbors with test
documents can be
weighed, so as to obtain the good classification effect. The cluster refers to
the set of a
category of documents with similar properties. The invention considers partial
discharge
signal data subset in the training set with the maximum distance between texts
in the same
category as one cluster. Therefore, the algorithm of K nearest neighbor
classification method
can be described as follows:
Step 1: In the training set, first of all the partial discharge data is
preprocessed and
16
CA 02918679 2016-01-19
mapped into a spatial vector;
Step 2: From the first category, the similarity of each two data among all the
signal data
in the category is calculated, the minimum threshold is set, and according to
the statistics the
clusters with close similarity are obtained;
Step 3: All the signal data in each cluster is merged. And then its central
vector is
calculated; in addition, the number of clusters/categories is calculated. This
value represents
the contributing coefficient of the cluster to the category, recorded as C;
Step 4: The new text is preprocessed, and its vector space is obtained;
Step 5: The distance between the spatial vector of new text with each cluster
of central
vector generated in Step 3 is calculated. These distances are multiplied by
the contributing
coefficient of corresponding cluster. The calculated results of clusters in
the same category are
added. After comparison, the biggest category obtained is the category of
partial discharge
with typical defect to be classified.
The basis for this algorithm is to find which texts in the same category
belong to the
same cluster. Below the idea of generated cluster algorithm of finding
clusters of the same
category is given. Assume the category:
c=fdl, d2, ....... dm)
Step 1: The threshold of similarity (a) is set;
Step 2: First a cluster is created, recorded as TO. The number of documents
contained in
the cluster is recorded with Ki. The total number of clusters created is
obtained. The
processed document i=2 is initialized;
Step 3: Start from di;
Step 4: The similarity with the first text in Tn is calculated, and s is
obtained;
Step 5: If S > a, and in Tn there is text not compared with this text, then
continue the
similarity calculation, and update s; if where is no text not compared, then
the data is added to
17
CA 02918679 2016-01-19
Cluster Tn; if S< a and there are other clusters not compared, then n++ and
return to Step 4;
if there is no cluster not compared, then new cluster is created, and recorded
as T++total; the
document shall be in Cluster T++total;
Step 6: If .'rn, then i++; return to Step 3; Otherwise, end.
In order to overcome the defect of high erroneous judgment of the nearest
neighbor
method, the nearest neighbor is expanded to K nearest neighbor. K nearest
neighbor method
does not select the nearest neighbor for classification, but select K
representative points
nearest to the text to be classified. And then according the category
information of the K
representative points, the category of text to be classified is determined.
For the characteristic parameter matrix after dimensionality reduction, half
of the
samples are used for training of K nearest neighbor classifier. The other half
is used to test the
performance of the classifier. For the characteristic parameter matrix after
dimensionality
reduction with KPCA, RST and CCMDR algorithm, K nearest neighbor classifier is
used to
recognize GIS insulation defect type. In this embodiment, under C language
software
environment, program documents are prepared, and design, training and
classification
recognition test of the classifier are realized. Because the output of the
classifier designed in
the embodiment does not have scattered distribution centered on a point like
BP neural
network, but corresponds to 4 GIS defect types, and the output value only
includes 4 results
[1, 2, 3, 4], so the pattern recognition result is only expressed as the
accuracy rate of
recognition, as shown in Table 1.
Table 1: The accuracy rate of K nearest neighbor algorithm pattern recognition
Defect type The accuracy rate of K nearest neighbor method
High voltage conductor metal protrusions 92%
Free metal particles 91.5%
Fixed metals on the surface of insulator 88%
Air gap defect of insulator 90%
18
CA 02918679 2016-01-19
In conclusion, the pattern recognition method for partial discharge of three-
phase
cylinder type ultrahigh voltage GIS of the embodiments of the invention has
the following
steps. The ultrahigh frequency is used to detect the partial discharge of
three-phase in one
enclosure GIS, and UHF sensor is used to sample the partial discharge (PD)
signal; the
improved wavelet threshold filtering method is used for denoising processing
of the partial
discharge signal collected, and the real partial discharge signal is obtained;
the characteristic
parameters of sampling signal are extracted through the mode algorithm based
on the phase
analysis, including degree of skewness (Sk), degree of steepness (Ku), the
number of partial
peak points (Pe), cross correlation coefficient (Cc) and discharge factor (Q);
the improved
kernel principal component analysis method is used for dimension reduction
processing of
characteristic space consisting of characteristic parameters, and the
characteristic parameter
matrix after dimension reduction is obtained; and K nearest neighbor algorithm
based on the
cluster idea is used for pattern recognition of insulation defect type of GIS.
The pattern
recognition method for partial discharge of three-phase cylinder type
ultrahigh voltage GIS at
least can have the following beneficial effects. Existing technical defects
can be overcome,
and the accuracy of the detection pattern recognition of partial discharge of
three-phase
cylinder type ultrahigh voltage GIS is improved.
Finally it should be noted that above is only the preferred embodiments of the
invention,
and not used to restrict the invention. Although the invention is described in
details with the
reference of the above embodiments, those skilled in this field can still
revise the technical
solution recorded in the above embodiments, or equally replace part of the
technical features.
Any revision, equal replacement and improvement within the spirit and
principles of the
invention shall be within the scope of protection of the invention.
