Some Problems with Partial Discharge Measurement in On-line Mode
K. Zalis, L. Beranova, L. Prskavec and R. Teminova
Czech Technical University in Prague, Faculty of Electrical Engineering, Department of Electroenergetics
Technicka 2, CZ- 16627 Prague 6, Czech Republic
Phone: +420-2-2435 2369, fax: +420-2-3333 7556
E-mail: {zalis, beranol, xprskave, teminor}@ fel.cvut.cz
Abstract  This paper deals with the problems
discussing the transition from off-line diagnostic methods
to on-line ones. Based on the experience with commercial
partial discharge measuring equipment a new digital
system for the evaluation of partial discharge
measurement including software and hardware facilities
has been developed at the Czech Technical University in
Prague. Problems with the elimination of disturbances,
recognition of patterns of partial discharges and
calibration of partial discharge equipment and measuring
circuit are also discussed.
Index Terms  Dielectric diagnostics, measurement, online measurement, partial discharges.
I. INTRODUCTION
Diagnostic methods are usually used for the
determination of actual state of high voltage (HV)
insulating systems, for the estimation of their residual
lifetime, their behavior estimation and the risk
assessment in the future operation [1]. Diagnostics of
HV insulating systems in off-line mode, i.e. during the
putout period or overhauling of the machine, is worked
up sufficiently and it is broadly executed. However,
both the price and power of the newly installed HV
equipment in the power engineering branch grow up,
and that is why the operator’s attention is focussed on
the operational reliability of their equipment at first. The
tendency of all operators is to monitor the state of their
equipment continuously, i.e. using on-line methods.
However, the application of some ‘classical’ methods
for on-line diagnostics is inappropriate, sometimes even
impracticable (e.g. direct current methods, loss factor
measurement, overvoltage tests). On the other hand,
suitable methods for on-line diagnostics are methods for
the observation of a discharge activity, which are
usually based on the monitoring of secondary effects
accompanying partial discharges (PD) in dielectric
materials ([2], [3]). One of these PD methods,
applicable on HV grounded objects, is the galvanic
method with parallel connection of the HV coupling
capacitor and the measuring impedance with a lowpass
filter. This method is broadly expanded due to its high
sensitivity, a good resolution of individual types of PD
impulses, and due to the fact that its using is not limited
in the capacity of measured object or the quantity of
used testing voltage. The advantage of this method is
also in the possibility in using it directly during the
machine operation by means of permanently installed
probes.
In the transition process from off-line diagnostics to
on-line one (monitoring) it is not possible to take over
original methodologies of the evaluation of diagnostic
parameters ‘automatically’. Some of diagnostic
parameters of off-line diagnostics can not be measured
in on-line diagnostics, some lose their sense and, on the
other hand, it is necessary to develop new on-line
diagnostic methodologies regarding of new conditions.
For example, in an off-line PD measurement, the
evaluation methodology is based on the dependence of
basic diagnostic parameters (apparent charge, PD
current, PD frequency) on applied testing voltage. In online measurement, the value of voltage is constant, but
new dependencies appear, e.g. changes of basic
diagnostic parameters in operational time. That is why
is necessary to develop new methodologies based on the
monitoring of time shift of diagnostic parameters.
Recently, some suspects about calibration process
accuracy in the case of measurement of large capacity
objects have appeared in technical community.
Although these problems are not still satisfactorily
solved, in this case the transition process from the offline diagnostics to the on-line one offer a quick and nontroubleshooting solution.
As regards the evaluation of a diagnostic
measurement, the quality of the evaluation and the
reproducibility of results are stigmatized by relatively
complicated methodologies and complicated (frequently
artificially made) diagnostic parameters, what leads to
the necessity of the consultation of top human experts
for the high-quality evaluation. However, the
complexity of the decision-making mechanisms
(frequently on the edge of the intuitive decisionmaking) leads very often to the ambiguous or opposite
evaluation of the actual state of the tested machine and
estimation of its behavior in further operation. This
practice is not acceptable for on-line diagnostics and it
is one of reasons why it is not so reliable and why the
on-line diagnostics is not so wide spread these days. In
connection with the on-line methodology development
it is necessary to reduce a number of diagnostic
parameters to the essential minimum, even at the cost of
wasting partial information about the machine actual
state. However, this disadvantage is entirely
compensated by the fact that the changes in diagnostic
parameters in an on-line measurement are indicated at
once, and the damage evolution can be monitored
permanently. The impossibility of using top human
experts for the routine on-line evaluation because their
temporary inaccessibility leads to the necessity of
developing such an instrument which compensates the
human expert view without decreasing the decisionmaking process quality. Expert systems based on the
elements of the artificial intelligence are the best
solution of this problem. Their knowledge bases
containing experience of the top specialists in the
decision-making process are able to solve standard
situations reliably. In addition, during an on-line
measurement, the expert system indicates a defect of the
insulating systems immediately and, at the same time,
offers a solution with respect to the safety and reliability
of the machine or equipment in further operation. There
are different types of expert systems based on
production rules (rule-based expert systems, framebased expert systems), neural networks, genetic
algorithms, fuzzy logic, etc., as well as expert systems
of a mixed type. At present, the rule-based expert
systems and neural networks are used the most often,
whereas their application depends on the type of
processed information. For the information of the type
„assumption-hypothesis“, i.e., the „if-then“ type of the
decision, production rule-based expert systems provide
the best solutions. On the other hand, neuron expert
systems (neural networks) are usually used for
complicated or intuitive decisions ([4], [5]), and in cases
when the algorithm development of the knowledge base
is very complicated, e.g. in PD patterns recognition ([6]
- [9]). For the evaluation of the state of HV insulation
systems, the rule-based expert systems are very
effective as a base of the evaluating or decision-making
systems.
II. DEVELOPMENT OF PD EQUIPMENT
The on-line measurements need a high reliability of
the measuring apparatus. Unsuitability and complicity
of existing measuring PD measurement apparatuses is
one of problems, which must be solved by the
diagnostic staff. ‘Classical’ PD measuring apparatuses
were developed no only for the data collection, but also
for the direct process evaluation of the diagnostic
parameters. They are usually based on analog data
processing. That is why commercial PD devices have
several disadvantages:
• The equipment is too expensive.
• The PD devices are single-purposed and usually they
can not be modified according to the specific
conditions of a PD measurement, tested equipment,
an operational interference and according to the
actual state of the research and development in this
branch.
• Analog components of PD devices change their
quality parameters in time. The calibration of these
PD devices must be usually done in the original
workplace of the producer, and thus the operational
costs increase.
• The commercial PD devices are too complicated and
mechanically sensitive and there is no guarantee of
their reliable operation during a long-time
measurement under the operational conditions.
These disadvantages result from the analog data
processing and a stable system conception of the data
processing. Both of them have negative influence on
data processing quality: time shifts of tolerances and
measuring ranges, low frequency ranges of analog
amplifiers, displacement of operating points, apparatus
sensitivity to disturbances etc. One of the most effective
possibilities to reduce the disadvantages mentioned
above is the consequent digitization of the PD impulses
immediately after their detection at the beginning of the
evaluation process (the best, directly after the indication
of the PD impulses) and subsequent processing digitized
data only. In addition, this data processing system
enables to apply an ‘arbitrary’ data filtration, data
processing by the classical computer programs, expert
systems etc.
A new principle of PD device has been developed at
the Czech Technical University in Prague in
cooperation with the Developing Laboratories at
Podebrady town ([10], [11]). The stable-measuring
equipment (a stand, a measuring workplace) for the PD
measurement and evaluation under the operational
conditions in on-line (non-interruptive) mode has been
developed, too.
Measuring unit for the measuring, digitizing and
processing PD data including a calibration equipment
has been developed within. Detected analog PD
impulses are digitized in the measuring unit by a special
analog-digital converter and they are saved in a special
memory block. The connection (via standard serial line
RS232) between the measuring unit and the computer
enables to transfer digitized PD impulses into a
computer for their further processing.
The proper detection and digitization of the input data
(measured values of diagnostic PD parameters) is
performed in the measuring unit, where the PD impulses
enter. These impulses are detected on the classical
measuring impedance. Like in case of the classical PD
measurement, the surface of the individual current PD
impulse is converted into a voltage value in a standard
capacitor, which is then discharged in a discharging
circuit. In contrast to classical PD devices, the
discharging time is set by a built-in digital clock in this
case, which is advantageous in exact countdown of the
discharging time and in the possibility of its further
digital processing. The discharging circuit was
developed and set in such a way that the discharging
time of the maximal charged standard capacitor (in case
of the maximal value of the current impulse in the input
amplifier), including resetting, should not take longer
time than 50 µs (it is adequate to 256 levels in a digital
form). It has a sufficient accuracy for reading the
apparent charge value as well as it has a sufficiently
high speed for processing PD signals (to the limit 200
signals during the one period of supply voltage, i.e.
during 20 ms). A phase shift of the PD impulses is
distinguished with the accuracy 1.8 °el., which is
sufficient enough.
Only two diagnostic parameters, an apparent charge
and phase shift of each impulse, are processed. These
two data (information about every PD impulse) of
10 periods of the supply voltage are saved in the
memory data block of the measuring unit and, after the
request from the computer, are, with the help of a
standard serial RS232 line, transferred into the
computer, where a special software processes them
further. The central computer automatically controls the
gain of the amplifier of the measuring unit, also via
RS232 line.
After the value digitization of diagnostic PD
parameters, the crux of the further activity lies in the
software data processing by a special software [12].
Before a proper evaluation of a PD activity, the
statistical processing is applied on measured data for
removing both random data and characteristic
disturbances (radio interference, thyristor disturbances,
etc.). ‘Cleaned up’ data are further processed and
modified for the input into the expert systems.
The evaluation of the diagnostic parameters and
monitoring of the insulation system in operation are
performed not only by standard classical methods
(according to the criteria values and alarm systems), but
also by the expert systems with the elements of artificial
intelligence, which enable to include the experience of
the human experts in this branch as well.
The developed evaluating system also uses two
independent expert systems for proceeding measured
PD data. These expert systems work simultaneously and
special software controls their coordination. The rulebased expert system performs the amplitude analysis of
PD impulses to specify the extent of the damage of the
insulating system. The neural expert system (a neural
network) has better ability of the abstraction, and
therefore it is used for the phase analysis of PD
impulses (the recognition of PD patterns), which enable
not only to specify the kind of PD activity, but even to
localize PD resources.
The outputs and visualization are performed userfriendly; i.e. all results are displayed on a virtual panel
of a standard measuring device. Except of the classical
visualization of PD impulses within one period of the
supply voltage (the visualization of PD impulses on a
sine curve, an ellipse or on an abscissa), the results of
evaluations the by expert systems, the mode of data
filtering, and the results of statistical processing, the
actual levels of diagnostic parameters, levels and the
alarm state activity are also continuously displayed on
the monitor screen.
III. ELIMINATION OF DISTURBANCES
In case of the PD measurement we usually want to
find the magnitude of partial discharges, especially we
made an integration of the current impulses. However,
great problems in the evaluations of PD data are in
various sorts of interferences.
We can divide disturbing signals into internal
disturbing signals and external ones. The internal
interferences are dependent on the actual voltage of the
measured subject; external interferences are not
dependent on it.
A lot of disturbing signals we can eliminate by using
of a suitable circuit modification. E.g. the four-capacity
bridge is usually used for small capacitance objects due
to its technical simplicity. We can also convert the PD
signal into digital (discrete) form.
The staff of the High Voltage Laboratory of the
Czech Technical University in Prague, Faculty of
Electrical Engineering, is interested in problems with
interference for several years, particularly in the frame
of the development of the new digital PD measuring
equipment. This equipment uses special software to
eliminate the interferences. It is then possible to
recognize and delete disturbing signals from processing
data sets.
We developed several algorithms for the elimination
of disturbing signals of digitized PD data [13]. Mainly,
we can eliminate randomize signals and thyristor
disturbances. The program applies math theorems to all
measured samples; it is for 2000 data from ten sets
(periods).
For the detection of randomize interferences we use
the following theorem:
n
q mm /(( ∑ q mi − q mm ) / n − 1) ≥ k
i =1
n ∈ 3;20 ,
where qmi is a maximal value from i-set, qmm is a
maximal value from all sets, n is a number of sets, k is a
constant and k∈ ⟨2;10⟩.
For the detection of thyristor interferences, we use
two theorems: At first, we find six highest charge
magnitudes in each set and we sort them (from m1 to
m6). Each of these elements has the own charge
magnitude qmi and the own phase shift ϕmi. Thyristor
interference impulses must perform two following
conditions :
•
Amplitude condition:
(( ∑ q mi ) / 6) /(( ∑ q i − ∑ q mi ) / (n − 6 )) ≥ k ,
6
n
6
i =1
i =1
i =1
where n is a number of sets (n = 200), qmi is an ielement from the six highest values of the actual set, qi
is an i-element of the set, k is a constant and k∈ ⟨2;10⟩.
•
Time condition:
n
≤k,
6
where n is a number of sets (n = 200), ϕmi+1 is a
position of an (i+1) element, ϕmi is a position of an ielement, k is a constant and k∈⟨1;5⟩.
A new way in the evaluation of PD signals is the
mathematical analysis, which makes possible a better
data processing. According to the frequency spectrum
of the signal we can recognize steady signals and
unsteady signals. In case of steady signals (stochastic
immediate broadband signals) we are interested in the
frequency spectrum only. It prefers Fourier Transform
(FT), which transform a view of the signal from timebased region to a frequency-based one. Mathematically:
ϕ mi +1 - ϕ mi -
∞
X ( f ) = ∫ x(t )e − jωt dt ,
−∞
where t is time, f is frequency, ω = 2πf and j = − 1 , x
represents signal in the time region and X represents
signal in the frequency region. The signal does not
change in time.
In the case of the unsteady signals, the frequency is
variable. We must adapt FT to analyze signal in the
small section of the time only. A this technique is called
the windowing of the signal, Short-Time Fourier
Transform (STFT). Mathematically:
∞
STFT (τ , f ) = ∫ [ x(t ) w(t − τ )]e − jωt dt ,
−∞
where w is a window size function and τ is a time
shift. STFT represents a sort of compromise between
the time- and frequency-based views of a signal. The
precision is determined by the size of the window.
The Wavelet Transform (WT) does not process of
signal in the time-frequency region as STFT, but it
processes it in the time-scale region. One of the main
advantages of this processing is the fact, that it is
possible to perform a local analysis. That is, to analyze a
localized area of a larger signal. In comparison with the
STFT the window size could be changed during
analysis. Mathematically:
∞
C (τ , s ) = ∫ f (t )Ψ (τ , s, t )dt
,
−∞
where C are coefficients of CWT, s is scale, t is time,
τ is time shift and Ψ represents a wavelet function.
The wavelet functions are unstable signals (wavelets)
and the coefficients represent their modifications. We
use several families of wavelets (Haar, Daubechies,
Biorthogfonal, Symlets and the others). Their variability
is very advantageous for the processing of unsteady
signal, e.g. for evaluation of PD signals. It could detect
signal changes and other variability. That is why the
WT is the best for PD signal analysis.
IV. RECOGNITION OF PD PATTERNS
The neural network [15] based on the empty expert
system named Neurex 5.1 was developed for the
recognition PD-patterns. This neural network was
trained by means of the software generator named GCV
(PD-generator). Two kinds of training set were created
by means of the GCV. The first one was the training set
with fixed values and the second one was with interval
values of input parameters. Both training sets have eight
final elements in the output layer, according to the basic
PD-patterns. We developed three versions of neural
network and six neural networks for training (three
versions with two kinds of training sets).
Testing set with sixteen elements was also created by
the GCV. None element from the training set was
accepted in the testing set.
After training procedure, all developed neural
networks were tested. Four training sets was acceptable,
two training sets (A-A and B-A) were unacceptable for
the neural network in the consultation mode. Variants
A, B and C are different by the count of neurons in
inner layers. Variant marked A-B has 25 neurons in
inner layer and it has the training set marked B (fixed
values). Training set A is with interval values. Variant B
has 100-25 and C has 100 neurons in the inner layer. All
our developed neural networks have three layers only.
After testing procedure we selected variant C-B for
the using in the neuron expert system, because it has the
best testing results. Training set A has not so acceptable
for this purpose.
Measuring data was then put into graph (see Fig. 1 and
2), where we can compare characteristics of tested
calibrators.
V. PD CALIBRATION
Calibration levels of 5000 and 10000 pC
Calibration levels of 500 and 1000 pC
1000
G6-6
G6-7
G3
TETTEX
BIDDLE
800
400
200
20 000
10 000
5 000
2 000
500
1 000
200
50
0
100
G6-6
G6-7
8000
G3
TETTEX
6000
4000
2000
20 000
5 000
10 000
2 000
500
1 000
200
50
100
20
0
10
0
Loading capacitance [pF]
Fig. 2. Calibration levels of 5000 and 10000 pC.
Alternative method: The output of calibrator is loaded
by the resistance Rm. The output voltage um(t) measured
on Rm is monitored by the digital scope. This scope must
have the bandwidth greater or equal than 50 MHz. The
value of Rm should be chosen between 50 Ω and 200 Ω.
Charge q generated by the calibrator then is:
1
q = ∫ i (t ) dt =
∫ u m (t ) dt ,
Rm
where i(t) is current impulse generated by the
calibrator and um(t) is voltage impulse measured by the
scope.
In the High Voltage Laboratory of CTU FEE in
Prague we measured the characteristics of commercial
calibrators by using the alternative method [15]. The
digital scope LeCroy 9350 with bandwidth 500 MHz
was used, which satisfied ČSN EN 60270. The values of
Rm used for loading of the calibrators output were 5000,
1000, 500, 200, 100, 50, 20, 10 and 5 Ω.
ACKNOWLEDGMENT
REFERENCES
600
20
10000
This research is financially supported by the
Department of Electroenergetics of the CTU and by the
Grant Agency of the Czech Republic (grant No.
102/02/0105).
1200
0
•
10
•
Following calibrators were tested:
Calibrator TETTEX, type 9216, range 10 pC –
10 000 pC.
Calibrator developed in the Research Laboratory of
the CTU, Faculty of Electrical Engineering in
Podebrady, type G6-6 (1993), range 5 pC –
10 000 pC.
Calibrator developed in the Research Laboratory of
the CTU, Faculty of Electrical Engineering in
Podebrady, type G6-8 (1998), range 10 pC – 25 000
pC, impulse frequency 50 Hz, 100 Hz, 1 Hz or
5 kHz switchable.
Calibration charge [pC]
•
12000
Calibration charge [pC]
PD measurement is the comparative measurement
and that is why that result of whole measurement
depends considerably on the calibration of measuring
devices and measuring circuits. Nevertheless at the
moment calibration is unjustly underrated in practice.
The newly released IEC 60270 introduces many new
requirements for PD measuring systems and related
calibrators. Calibrator consists of generator producing
pulses of voltage of the amplitude U0 connected in
series with capacitor C0. Charge of calibration pulses q0
then equals q0=U0C0.
Accuracy of measurement depends on the accuracy of
calibrators and we have to carry out operational test.
These tests have to be carried out because of
determination and keeping of characteristics of
calibrators.
Operational test from the newly released IEC 60270
consists of:
• Charge of calibrator q0 for all setting of calibrator,
with an uncertainty within 5% (or 1 pC – whichever
the greater).
• Rise time tr of voltage U0, with an uncertainty within
10%. Rise time must be less than 60 ns (tr < 60 ns).
Loading capacitance [pF]
Fig. 1. Calibration levels of 500 and 1000 pC.
[1] A. Šandrik and J. Zlatovský, „Diagnostics and
construction of the insulating system of electrical
machines”, Proceedings of the 12th International
Conference Dielectric and Insulating Systems in
Electrical Engineering DISEE ’98, 1998, Bratislava and
Častá-Píla (Slovakia), pp. 62-65.
[2] D. König and Y.N. Rao Y.N.: “Partial discharges in
electrical power apparatus”, VDE-Verlag GmbH, Berlin
and Offenbach, 1993.
[3] K. Marton, “Imperfect point in insulating system as a
radiator of electromagnetic signal”, Proceedings of the
12th International Conference Dielectric and Insulating
Systems in Electrical Engineering DISEE ’98, Bratislava
and Častá-Píla (Slovakia), 1998, pp. 30-33.
[4] M. Kolcun, “Optimalization of operation of power
stations in electrification system”, FEI TU Košice, 1999.
[5] R. Cimbala, “Utilization of pattern recognition
principles for diagnostic purposes”, Journal of Electrical
Engineering (Slovakia), Vol. 45, No. 8, 1994, pp. 315317.
[6] E. Gulski and F.H. Kreuger, “Computer-aided
recognition of discharge sources”, IEEE Trans. of
Electrical Insulation, Vol. 27, 1992, No 1, pp. 82-92.
[7] F.H. Kreuger, E. Gulski and A. Krivda, “Classification
of partial discharges”, IEEE Trans. on Electrical
Insulation, Vol. 28, 1993, pp. 917-931.
[8] I. Kolcunová, “Diagnostics of high voltage and extra
high voltage systems by partial discharge method”, FEI
TU Košice, 1999.
[9] I. Kolcunová and I. Kršňák, “Types of partial discharges
and their recognition in insulation of high voltage
rotating machines”, Proceedings of the Scientific
Conference, Oradea (Romania), 1997, pp. 67-70.
[10] J. Říha and K. Záliš, “Charge calibrator G 6-8”, CTU
FEE, Prague, 1999.
[11] Patent No. 284558 “Final stage of Partial discharge
calibrator”. 1997, CZ.
[12] J. Karas, “Evaluation System for Partial Discharge
Measurement”, Proceedings of the 3rd International
Student Conference on Electrical Engineering “POSTER
99”, Prague, 1999, ref. EE1.
[13] K. Záliš and L. Beranová, “Sort of disturbances
influencing partial discharge measurement and their
elimination”, Proceedings of the DIAGNOSTIKA ’01
Conference, 2001, Nečtiny (Czech Republic), pp. 57-60.
[14] K. Záliš and L. Prskavec, “Mathematical models of
partial discharge patterns”, Proceedings of the
DIAGNOSTIKA ’01 Conference, 2001, Nečtiny (Czech
Republic), pp. 61-64.
[15] K. Záliš and R. Těmínová, “Calibrators for partial
discharge measurement”, Proceedings of the DIAGNOSTIKA ’01 Conference, 2001, Nečtiny (Czech
Republic), pp. 65-67.