Get documents about "Building a Transformer Defects Database for UHF Partial
Document Sample
431 1
Building a Transformer Defects Database for
UHF Partial Discharge Diagnostics
P. Agoris, P. Cichecki, S. Meijer, and J. J. Smit, Member, IEEE
In general, for the condition assessment of a power
Abstract— In the case of a defective transformer, when a transformer using the UHF PD measuring technique three
partial discharge is detected and recorded, critical information steps have to be performed. First of all, detection of any
can be deduced from its pattern, such as the type of defect, its
partial discharge activity that indicates harmful insulation
criticality or even information on the level of degradation of the
insulation. This information can help to determine the defects, must be accomplished. After the detection and
remaining life of the transformer and thus provide criteria for recording of the PD has taken place, the analysis of the raw
its maintenance and operation. In this paper different artificial data would give the appropriate information in order to
PD patterns will be recorded in the laboratory, representative of identify the defect in the transformer The identification of the
specific transformer defects, in order to build a database for defect can be done by finding the location of the insulation
comparison purposes when measuring on-line. This can greatly
defect and by comparing its pattern with other known defects
improve the recognition and identification of the defect and thus
help take some important life assessment conclusions on the from a reference database. All information then has to lead to
transformer. a certain assessment of the risk for an insulation failure of the
power transformer.
Index Terms—Diagnostics, Identification, Pattern In this paper we are interested in the third stage of the
Recognition, Partial Discharges, Power Transformer, UHF assessment: that is recognizing the recorded pattern and then
classifying the defect in a defects’ database [1]. Finding the
location of the defect can also hint to its identification,
I. INTRODUCTION
however location requires at least three UHF sensors which is
O NE of the key components in the electrical power
network is the power transformer. A transformer failure
not the case in older equipment. On the other hand, a
recording of a pattern can be performed by only one UHF
implicates considerable costs either due to outage or the sensor, sensitive enough to detect it. Identification is possible
refurbishment or replacement of the transformer. As a result, in all transformer cases as long as a proper pattern is
the reliability of power transformers is considered to be recorded. Still, however, in this case, assessment of the PD
important. A significant number of failures of power nature is very hypothetical if there is no reference pattern of a
transformers occur in the dielectric system, mainly due to known defect to compare it with. It is thus necessary to build
insulation deterioration and defects. To avoid such failures, a UHF pattern database of transformer defects.
periodical assessment of the condition of the transformer Correct analysis of the pattern of the PD, in order to obtain
insulation is needed. One of the ways of monitoring the its discriminating characteristics, is necessary. In this paper
condition of the insulation is by means of partial discharge three different distinct defect groups are examined. The
measurement. Although research until now has not patterns and some statistical features obtained will be
progressed enough so that a partial discharge (PD) presented and discussed and will then be used in a database.
measurement itself could provide information regarding the Some independent data will then be used to test the
remaining life of a transformer, it does give information identification capabilities of the appropriate software.
about the quality of insulation, which can help further
decision making processes. II. EXPERIMENTAL SET-UP AND PROCEDURE
The experiments were performed on a 1MVA transformer
Manuscript received October 9, 2001. (Write the date on which you (see Fig. 1). The purpose of the experiments was to build a
submitted your paper for review.) This work was sponsored by Tennet. database of the most common transformer defects, which can
P. A. Author is with the Delft University of Technology, Delft, 2628CD, The
Netherlands (phone: +31 (0)15 27 89042; fax: +31 (0)15 27 88382; e-mail: later be used in the field for the identification of any detected
p.agoris@ewi.tudelft.nl). defects. For this purpose, a number of patterns generated from
P. C. Author is with the Delft University of Technology, Delft, 2628CD, The
Netherlands (e-mail: p.cichecki@ewi.tudelft.nl).
certain artificial defects inside the 1MVA transformer had to
S. M. Author is with the Delft University of Technology, Delft, 2628CD, The be recorded. Therefore, a small opening was made at the top
Netherlands (e-mail: s.meijer@ewi.tudelft.nl). of the tank were the artificial defects’ setup could be inserted.
J. J. S. Author is with the Delft University of Technology, Delft, 2628CD,
The Netherlands (e-mail: j.j.smit@ewi.tudelft.nl). The artificial models that were used, were built with the
431 2
idea that they should be the most basic and most cavities. The cavity was placed in the middle of two-brass
representative partial discharge defects in power electrodes cast in polyester mass and was mounted in a plastic
transformers. Thus the following partial discharge models frame (see Fig. 2). The diameter of the cavity was 3 mm.
were designed and prepared: cavity discharges, surface With this model two measured voltages were investigated.
The first was at the inception voltage level (at 14kV); In this
voltage pulses the discharge ranged from about 11nC to
15nC. Pulses were stable and did not disappear for the next
20 minutes so a measurement could be carried out. In the next
step the voltage was increased to 18 kV and another
measurement was performed. In this case the PD energy
ranged from 11nC to 20nC; the pulses were quite stable and
random very high peaks occurred. The maximum discharge
Fig. 1. 1MVA transformer with opening on top for artificial defect set-up
insertion. The white circled areas indicate the places where the UHF sensors
where located.
discharges and corona in oil. Patterns for these discharge
models were obtained both with the UHF and the IEC 60270
technique; in this paper only the UHF results will be
presented and discussed.
The measuring procedure was the following:
• Increasing the voltage from 0 kV until PD inception
voltage, where first PD activity is observed Fig. 3. Internal defect model phase resolved patterns. The corresponding supply
• Increasing the voltage to a maximum possible level voltage and measuring frequency are also indicated.
(below the breakdown voltage) to find out how the frequency amplitude was 28nC and it was near the breakdown voltage
response and phase resolved patterns are changing with range of 25 kV. Results of the PD in the cavity model are
voltage.
• Recording the frequency spectra in all appropriate
voltage levels.
• Recording of several phase resolved patterns [1] at
different frequencies of high PD energy to observe if and how
Fig. 4. “Vertical point to plane” (3mm pressboard used for tests) and diagram
indicating the possible path of the discharges model.
presented in Fig. 3.
B. Surface discharges
Fig. 2. The internal defect model. Surface discharges are a result of a strong stress component
the pattern changes with different center frequencies. parallel to the dielectric surface. Such a PD activity results in
the deterioration of the insulating material and produces
III. THE PARTIAL DISCHARGE MODELS different chemicals (e.g. nitric acid and ozone), which are
also harmful for the oil insulation [2]. In transformers,
A. Internal defect surface discharges occur mainly in the composite oil-barrier
The final result of PD activity due to cavities in the insulation but can also be a result of bubbles trapped on the
insulation material, is erosion of the insulation material, insulation or the delamination of the pressboard layers. For
chemical changes in material structure and thus local the laboratory measurement the surface discharge model used
carbonization of the material and generation of electric is shown in Fig.4.
treeing discharges. For this reason cavity discharge model has It was designed such as to create a high tangential field and
been constructed to investigate. a high concentration of electric field at a certain part on the
Since cavities appear due to various reasons (e.g. bad surface of the pressboard. Pulses occurred immediately at 16
impregnation process, mechanical distortion during kV and a measurement was performed. At this voltage range
transporting and voltage overstresses etc) and thus have
the maximum amplitude of the discharges was 2.2 nC and
different characteristics (geometry, location, material) it is
appeared in the negative voltage half of the sinus.
difficult to find one representative model for all types of
431 3
The distance between the two electrodes could be set at
different values.
With this model the inception voltage was 28,5 kV with a
pulse amplitude range of 1nC – 3nC; however the pulses were
not stable and disappeared from time to time. After increasing
Measuring Frequency: 1000MHz
Supply Voltage: 18kV
Number of Voltage Cycles: 1000
Fig. 7. The internal defect model phase resolved patterns. The corresponding
supply voltage and measuring frequency are also indicated.
Measuring Frequency: 1000MHz the voltage level to 35 kV pulses the phenomenon started to
Supply Voltage: 18kV
Number of Voltage Cycles: 300 increase and the PD magnitude was now between 8nC and
11nC. The pulses were stable and in this level measurements
Fig. 5. The surface discharges defect model phase resolved pattern. The were performed. After about 20 minutes the voltage was
corresponding supply voltage, the measuring frequency and the recording time
(in voltage cycles of 50Hz, that is 20 ms) are also indicated. increased to 40 kV. In this range the magnitude of the pulses
was from 12nC to 16nC. Recordings are shown in Fig. 7.
After about 10 minutes the discharges at 16 kV
disappeared and the voltage was increased to 18.4 kV and the IV. IDENTIFICATION
measurement was performed again: the pulses ranged
between 1.4 and 2.4 nC. The voltage was raised up to 25 kV A. Data analysis
and 34 kV. It could be seen that increasing the voltage From the recorded patterns it is possible to perform
generated more pulses but the magnitude of these pulses was
Fig. 6. HV corona discharges model: (left) HV electrode (needle) with 4 cm
distance to the ground electrode and (right) model in plastic frame.
not changing considerably. For example at 34 kV the
difference with the 19 kV measurement was only 0.5 nC. The
patterns obtained are shown in Fig. 5.
C. High voltage corona in oil
In power transformers corona discharges can appear in the
oil due to pollution caused by small conductive particles,
which may result in the decrease of the breakdown stress of
the oil. Such particles originate from the manufacturing
process, from mechanical distortion, or from service
operations like oil changing (small particles from pumps and
other service equipment). Sometimes sharp edges of grounded
parts in the transformer tank can also produce corona Fig. 8. The statistical distribution of a surface discharge phase resolved pattern.
discharges. (Top to bottom): a) the maximum and b) the average magnitude of the
For the production of corona discharges, a model was built discharges, and c) the number of discharges in each phase window
of a sharp high voltage electrode (needle) as seen in Fig. 6.
431 4
recorded data (for each defect) comprised of measurements of
600 sweeps at i) two different voltage levels, ii) two different
center frequencies (250 MHz and 1000 MHz); furthermore
measurements of 1200 sweeps at again two different voltages
and frequencies were also added in the database under the
same defect label. The next step was then to calculate the
statistical operators and then the standard deviation of each
operator for each defect group. The results are shown in Table
&&
It can be seen that some operators are very unstable (their
standard deviation is much higher than 1) either for all three
Fig. 9. The fingerprint is a set of statistical operators, which describe the defects or for a specific defect. For instance the skewness Hq,
distributions shown in Fig. 8. is unstable for the HV Corona defect only, while the Kurtosis
different statistic calculations that can describe the pattern of Hq and Hn+ (although it is mainly unstable for HV
characteristics in more detail. Initially the statistical Corona) it can be said that it is slightly unstable for the other
distributions of a) the maximum and b) the average defects as well (since the standard deviation exceeds 1); in
magnitude of the discharges, and c) the number of discharges addition the Kurtosis Hn- is very unstable for the surface
in each phase window can be determined as it is discharges.
demonstrated in Fig 8. Then several operators are applied to Especially in the case of corona the pattern can really
these distributions to describe their characteristic shape (such change with any increase in the supply voltage or the
as the symmetry of the pattern, the skewness e.t.c.) [3]. These measuring frequency. This happens because in the negative
operators, when collected together will generate the specific cycle of the voltage (second half of the pattern) the discharges
fingerprint of the defect, as seen in Fig. 9. Lastly, these are very low and the higher the measuring frequency the
operators are used to compare the patterns with each other so smaller and less they become (even disappearing completely
that they can be separated into three (or more in other cases) at a certain point due to the limited recording resolution as
distinct groups, which will then be added to the database. well); the same applies for the different voltage levels.
The comparison procedure was performed based on the Therefore the asymmetry (showing symmetry between the two
centourscore algorithm [2]. Since at least three fingerprints in halves of the pattern: i.e. the positive and the negative halves
one group are necessary for the algorithm to do a proper of the voltage cycle) and Hq (represents the number of
comparison, it was necessary to create the groups manually at discharges as a function of the discharge magnitude)
first by copying the appropriate fingerprints to each defect operators will naturally give different results in each case.
group. After this initialization the algorithm can be used to With similar logic the Hn operator (the number of pulses in
compare the rest of the fingerprints and add them to the each phase window as a function of the phase angle) will
proper group. generate big variance for the surface discharge, whose pulse
B. Comparison and classification density increases with voltage. The same would apply for any
other defect whose pulse rate is not constant and thus short
The main purpose of this paper is to see if a correct
recording times (thus a few voltage cycles) would always
identification of the same defect can be performed with
different external factors; these external factors are the supply
voltage, the measuring frequency and the recording time of
each measurement. Each factor affects differently the end
result of a measurement (i.e. the phase resolved pattern). For
example, as it was noticed before, during the surface
discharge measurements, the higher the supply voltage the
more dense the pulses become without however changing the (a)
amplitude of the discharge; in this case any statistical
operator which relates these two characteristics will generate
a different result for each voltage case differentiating the
fingerprints. For these reason the statistical operators have to
be chosen with care to be independent of such cases.
The statistical operators used in the current database
software are listed in the left column of Table &&. Their (b)
Fig. 10. The Hn distributions of the surface discharge model patterns in Fig. 5.
definitions can be found in [3], however if it is required in the For the same voltage and measuring frequency the (a) 1200 sweeps and (b) 600
following discussion their definition will be stated also. The sweeps recording generate different Hn distributions.
431 5
generate a different Hq and Hn value. An example of this case discharge patterns. Using all statistical operators the result
is demonstrated in Fig. 10, where the Hn distributions of the was 20% surface discharge, 10% cavity and 5%HV Corona.
patterns in Fig. 5 are shown. Although the patterns represent Removing all the operators related to the Hn parameter
the same defect at the same supply voltage and measuring resulted in 30% surface, 20% cavity and 15% HV Corona. It
frequency, the difference in the recording time generates seems to increase the likeness to surface discharges but it
different Hn characteristics and therefore discriminates the however increases the likeness percentage of the other defect
two patterns. groups as well. Although the increase is very small, it can be
As far as the rest of the operators are concerned, which are quite decisive in other cases where the likeness is higher.
also more amplitude related, they seem to be more stable. Trying different other 100 cycle surface patterns however
This is mainly because the operators are calculated after a gave similar results.
normalization of the discharge amplitudes (for each recording The main difference with the 100 cycle and 500 cycle
separately) and thus is independent of any change in the measurements was the number of pulses recorded; depending
amplitude due to the voltage or the measuring frequency. on the duration of the recording the pattern can also look
different. Since the shape of the patterns is dependent to the
total pulses recorded and since different defects have different
Statistical
Cavity HV Corona Surface pulse rates – in this case the cavity defect and the surface
Operators
0.418 0.245 0.158 defect had quite a high pulse rate compared to corona as it
Skewness Hqmax+
can be seen in the patterns of the defects in Fig. 3, 5 and 7) –
Skewness Hqmax- 0.375 0.062 0.201
it is difficult to specify a standard measuring time. The pulse
Skewness Hqn+ 0.403 0.322 0.163
rate per cycle has to be considered and then the recording
Skewness Hqn- 0.375 0.047 0.137 time should be set appropriately to record the same total
Skewness Hn+ 0.622 0.474 0.504 pulses as the patterns loaded in the artificial defects’
Skewness Hn- 0.422 0.103 0.535 database.
Skewness Hq 0.734 5.607 0.421
V. CONCLUSIONS
Kurtosis Hqmax+ 0.209 0.49 0.762
• The behavior of the different partial discharge models in
Kurtosis Hqmax- 0.254 0.045 0.324
terms of patterns, repetition rate, and amplitude is
Kurtosis Hqn+ 0.204 0.351 0.498 characteristic for each defect type only for the specific
Kurtosis Hqn- 0.3 0.046 0.13 voltage, measuring frequency and measuring time that they
Kurtosis Hn+ 0.976 2.206 1.258 were recorded.
0.587 0.315 2.424
• The PD measurements should thus be done under similar
Kurtosis Hn-
conditions and duration time in order to improve the
Kurtosis Hq 1.056 113.668 1.251
possibilities for a proper match.
Discharge asymmetry • The statistical operators used for the comparison of the
0.655 0.529 0.194
Q Hqmax
different phase resolved patterns should be independent of the
Discharge asymmetry
Q Hqn
0.678 0.408 0.191 pulse rate and be more involved with the actual shape of the
pattern (location on phase cycle, skewness of max and
Cross Correlation 0.145 0.276 0.183
average normalized amplitudes e.t.c.). Studying the pulse rate
Modified Cross
Correlation
0.522 0.404 0.22 of a PD should require longer period of measurements in
order to observe the pulse sequence in time which is also
Peaks found Hqmax+ 0.311 0.354 0.279
characteristic for each defect.
Peaks found Hqmax- 0.245 0.473 0.223 • For the comparison of the phase resolved patterns it is
Peaks found Hqn+ 0.378 0.415 0.293 necessary that the recording parameters (measuring frequency
Peaks found Hqn- 0.233 0.473 0.283 and time) be the same;. Otherwise results may vary since the
patterns might be different due to the number of pulses
Peaks found Hn+ 0.242 0.333 0.334
recorded which vary the shape of the pattern.
Peaks found Hn- 0.21 0.575 0.404
Peaks found Hq 0.146 0.28 0.14 ACKNOWLEDGMENT
Furthermore, a test of the classification procedure was The authors would like to acknowledge TenneT and Nuon
performed if the unstable operators were removed from the N.V. for their support.
classification procedure. The database was consisted with
mainly 500 cycle patterns of all three groups. For this REFERENCES
example a surface discharge pattern of 100 cycles was to be [1] P.D.Agoris, S.Meijer and J.J.Smit: “Evaluation of On-Line Insulation
classified; the pattern was obtained at the same voltage and Condition Assessment Techniques for Power Transformers” International
Symposium on High-Voltage Engineering (ISH), Beijing, China, (2005)
measuring frequency as the rest of the 500 cycle surface
431 6
[2] F. Massingue, S. Meijer, P.D. Agoris, J.J. Smit, J. Lopez-Roldan: “Partial
Discharge Pattern Analysis of Modeled Insulation Defects in Transformer
Insulation” IEEE International Symposium On Electrical Insulation, (ISEI
’06), Toronto, Canada, 2006
[3] A Krivda, “Recognition of discharges discrimination and classification”
Delft University press, 1995, Delft
[4] F. H. Hreuger, E. Gulski, A. Krivda, “Classification of Partial
Discharges”, IEEE Transactions on Electrical Insulation, Vol. 28 No. 6,
December 1993.
