Location of Partial Discharges within a Transformer Winding Using Principal Component Analysis M. S. Abd Rahman, L. Hao, P. L. Lewin University of Southampton, Southampton, UK

1
Location of Partial Discharges within a
Transformer Winding Using Principal Component
AnalysisM. S. Abd Rahman, L. Hao, P. L. Lewin
University of Southampton, Southampton, UK
PD Source
Introduction
In order to generate partial discharge inside the
transformer winding, a PD source of an artificial
void in oil has been developed as shown in figure
3 and its corresponding output signal is shown in
figure 4. To simulate PD behavior, a hollow
stainless steel sphere was used as the upper
electrode. A cylindrical air bubble was trapped
inside the sphere of size diameter 4.73mm and
length10.30mm. A rectangular perspex block of
length 139mm, width 200mm and thickness 7.53mm
was inserted between pair of electrodes, the
upper electrode was connected to the high voltage
supply and the lower was grounded. The whole PD
source arrangement was immersed in transformer
oil.
Partial discharge (PD) is a common phenomena
which may occur in high voltage equipment such as
power cables and high voltage transformers. It is
due to ageing processes, operational over
stressing or defects introduced during
manufacture. IEC 60270 defines PD as a localised
electric discharge that only partially bridges
the dielectric insulator between conductors when
the electric field exceeds a critical value. PD
might occur anywhere inside a transformer
particularly along the transformer winding and
the discharge signal can propagate along the
winding to the bushing and neutral to earth
connections. Therefore, the identification of a
PD source as well as its location are essential
in PD monitoring and evaluation processes. The
wavelet decomposition is useful in order to
decompose PD signals into detail levels and an
approximation by using an approach known as
quadrature mirror filtering which splits the
signal into two bands, high pass and low pass
signals. Wavelets can be used to improve signal
to noise ratio in many applications, but in this
case it is used to identify the distribution of
signal energies in both time and frequency
domains. Principal component analysis (PCA) is a
pre-processing algorithm that uses an orthogonal
transformation to convert a large set of data
into set of variables or principal components. In
this case, the use of PCA is to compress this
data into three dimensions, to aid
visualisation. An experiment has been developed
in the Tony Davies High Voltage Laboratory that
can be used to create PD data over hundreds of
cycles of applied voltage in order to investigate
the feasibility of using this approach to
identify PD locations.
Fig. 3. Artificial void in oil PD source
Fig. 4. PD signal generated by artificial source
Results
Signal processing
The Discrete Wavelet Transform The discrete
wavelet transform is formed by passing the signal
through a series of filters. The signal is passed
through a low pass filter (m) where the output is
the approximation coefficients, the signal is
also decomposed simultaneously using a high pass
filter (n) giving the detail coefficients
(figure 1). Principal Component
Analysis PCA solves an eigenvalue problem. Thus,
the data is transformed into a new coordinate
system with a corresponding variance.
Theoretically, PCA is a linear transformation
that projects the set of input data to a new
coordinate system, the principal component that
has the greatest variance is labelled as the
first principal component the second principal
component is the one that has second greatest
variance and the third greatest variance becomes
the third principal component, allowing
projection of the input data into three
dimensions as implemented in this case. However,
PCA can be two or three dimensionsal depending on
the particular application.
Approximation coefficients
mk
xk
a. Bushing tap point
b. Neutral
end point Fig. 5. 3-D f-q-n pattern for
discharge signal from terminal 8
Detail coefficients
nk
Figure 1 Block diagram of Wavelet single
decomposition
a. Detail D3 b.
Detail D7 Fig. 6. Detail levels at neutral
end
Experiment
The experiment is based on a high voltage
transformer winding model BS1481998 class 1 and
a 60 kV transformer bushing, model 60HC755 in
order to study the PD activity inside transformer
winding. The model contains 2 types of winding,
interleaved disc and plain disc winding, in this
case PD signal source being injected to the
interleaved disc winding whilst the plain disc
winding was grounded. The interleaved winding
consists of eight terminals, terminal one of the
winding is connected to the bushing core bar
while the last terminal was grounded. The
artificial PD source was injected at each
terminal of the winding and the discharge current
from the PD source is measured at measurement
points. There are two PD measurement points in
this case as the PD signal can travel in both
directions the first point is located at the
bushing tap and the other is at neutral to earth
connection as shown in figure 2. The discharge
current flowing to both ends can be measured by
using a radio frequency radio transducer (RFCT).
The RFCT used in this experiment is the
clamp-type split core RFCT EMCO model 93686-5,
serial model 9802-50174 which has a measurable
frequency range from 10kHz to 200MHz. A digital
storage oscilloscope, Tektronix DPO7254 with a
bandwidth of 2.5 GHz and sampling rate 40 Gs/s
was used to display, analyse and hence store the
obtained output signals from both ends. A Mtronix
partial discharge MPD 600 and a charge calibrator
type cal542 has been used for detecting and
recording partial discharge events and charge
calibration.
a. Bushing tap
point b. Neutral end point Fig. 7.
Energy distribution
Discussion
The phase resolved partial discharge plot (f-q-n)
for both the bushing and neutral ends (figure 5)
show almost similar PD patterns that are typical
of a void in oil PD source. The magnitudes are
higher at the neutral end due to the closer
location of the PD source along the winding.
Wavelet Analysis (typical examples of detail
levels for a captured signal are shown in figure
6) can be used as a method to identify the
distribution of energies in both the time and
frequency domains. Figure 7 shows the analysis of
600 pairs of signals and clearly demonstrates
different energy distributions depending on
measurement point. Hence this analysis produces a
10 dimensional feature vector for each captured
discharge signal. PCA will be used to compress
the 10d energy vector into a three dimensional
representation.
Conclusions
RFCT sensors located at both bushing tap and
neutral end can detect discharge signals and it
is possible to record and store PD signal data
over 50 cycles of applied ac voltage.
Application of wavelet discrete decomposition is
a useful pre-processing technique that can
produce a feature vector that represents the
distribution of signal energy in both the time
and frequency domains. This information can be
described in 3 dimensions by using principal
component analysis. Further work will consider
analysis of the clusters of data produced using
this approach to determine whether it provides
insight into source location.
M.S. Abd Rahman, msar106_at_ecs.soton.ac.uk Universit
y of Southampton, Highfield, Southampton, SO17
1BJ, UK
Contact details
Figure 2 Experiment for simulating partial
discharges within a transformer winding.