identification of partial discharge...
TRANSCRIPT
University of Wisconsin, Madison
ECE/CS/ME 539: Introduction to Artificial Neural Networks and Fuzzy Systems
Identification of partial
discharge signals
Marcus Vinicius Marques de Paula
Madison 12/20/2013
Background
1. Partial Discharges
The definition of partial discharges have been contradictory and generated some
discussion. The greater consensus on its definition is: partial discharge (PD) is a
localized dielectric breakdown of a small portion of a solid or fluid electrical insulation
system under high voltage stress. The term "partial" refers to the fact that the ionization
process occurs only in part of the space between the electrodes responsible for
generating the field, thus not getting to form a complete rupture of the insulation system
(breakdown).
Figure 1. Schematic of partial discharge
It is considered that there are three requirements for the occurrence of partial
discharges in the insulation system: the presence of a gas is required; the presence of an
electric field that exceeds the limit of isolation by the gas is needed; and the presence of
at least one free electron is required for the process to be triggered. Once the first
discharge occurred, free electrons typically become abundant and the occurrence of new
discharges becomes easier.
The occurrence of partial discharges in an isolation material can’t lead to any
serious consequence or can lead to complete rupture of the same. Obviously, the
presence of partial discharges which do not compromise the insulation performance
over time is perfectly acceptable. However, and unfortunately, in most cases the
occurrence of discharges cause degradation of the material and may cause disastrous
effects. The cavities found, arise due to imperfections in the manufacturing process of
the equipment or due to chemical reactions.
It is observed that, during the occurrence of partial discharges, the walls of the
dielectric material are constantly bombarded with electrons and positive ions, which
have high kinetic energy. This bombardment causes deterioration of the dielectric
material, due to energy transfer, to the atoms of the dielectric material.
The phenomenon is a "snow-ball" and can lead to the growth of the degradation
of the material to the point of complete loss of insulating capacity and consequent
failure in the electrical system. The growth of the electrical treeing (figure 2) may take
years, weeks or even days, which is why regular checking of high-voltage equipment is
needed and the phenomenon is considered so dangerous to electric power systems.
Figure 2. Electrical treeing
Therefore, partial discharge (PD) analysis is a highly pursued feature due to the
economy related to scheduled shutdowns, disassembling and transport, and for that
reason, the generating and transmission of energy companies keeps monitoring
programs based on a set of techniques believed to be reliable. Nevertheless, the PDs
measurements are frequently limited by interferences found in high-voltage facilities, a
situation that imposes the continued development of PD signal processing methods.
2. Filtering
The measurement systems are characterized by their PDs band-pass, which
depends on the equipment under test and evaluation objectives. Although there is no
consensus on the classification, the literature reports measurements at close range (units
of kHz), wide (300kHz - 1MHz), ultra-wide (10 - 20MHz) and very high band-pass
(units to tens of GHz). Narrowband and broadband systems are commonly used for the
detection of distributed parameters, such as transformers and electric machines.
The measurement of partial discharges in the field usually requires the use of
techniques to eliminate noises, among which the most common is the limitation of the
band-pass of the measuring equipment. This technique may prove satisfactory in cases
where the noise spectrum is limited (eg, radio signals), but other types of interference
may hinder the measurement.
Most times, the measurement signal has noise whose frequency range is very
similar to the frequency range of partial discharges and for that reason, digital filtering
methods are commonly used.
Goal
In the last years the wavelet transform (WT) has been recognized as a powerful
technique for PD processing due to its capacity to process localized, non-stationary
signals. Several authors have reported good results of its use and, more recently, new
WT-based approaches have been developed specifically to improve PD processing.
The method used by Dr. Hilton Mota [2], is a new technique for the processing
of partial discharge signals, based on the wavelet transform and a spatially-adaptive
coefficient selection procedure. Spatially-adaptive selection is an excerption approach
that aims to explore the localized processing capabilities of the WT as a way to improve
the separation of coefficients related to the signal and noise. This approach frequently
allows a better processing for time-localized signals, like the PDs, when compared to
traditional, threshold-based techniques.
In his work, the spatial correlations were characterized by the local modulus
maxima propagation theory. Coefficients selection was performed by the
characterization of maxima lines shapes and classification by a deterministic rule and a
pattern classifier. The procedure relies on the Translation-invariant Wavelet Transform
as a way to avoid PD pulse losses and improve the signal reconstructions.
The process basically consists of six steps:
1. Decomposition of the signal into six levels using WT.
2. Extraction of each decomposition.
3. Construction of the maxima lines.
4. Classify lines, separating the ones related to the signal from the noise lines.
5. Delete rows associated with noise.
6. Rebuild signal using the remaining lines.
Thus, the goal of this project is to implement a classifier to accomplish the step 4
of the process above. Also, this work shows the efficiency of the Support Vector
Machine and of the Multilayer Perceptron.
Training Set
To implement the whole process of filtering PDs signals, a bunch of data was
needed to test, improve and compare results. Therefore, to test the method proposed, we
built an equipment to simulate partial discharges inside the laboratory (figure 3) and a
measuring system [1] to be able to collect digital samples of the phenomenon.
Figure 3. PDs simulation system
The system consist basically of an electric transformer to create the high voltage
stress (bottom right), a capacitor of high capacitance to store energy (on the middle) and
a third instrument (on the left) that consists of a solid metal cylinder with a needle that
is positioned perpendicularly to the flat surface of the cylinder in order to simulate the
occurrence of a partial discharge.
That way we were able to collect a large number of data regarding some
experiments.
But also, Eletrobras Furnas, a company that generates and transmits electricity in
Brazil, provided us with some real PDs signals obtained tests on components, which
help us a lot in terms of testing the methodology with those samples.
Implementing the classifier
It is not the purpose of this work, deal with the other steps of the filtering
process (more information can be taken from the references), so only a few comments
will be made about these steps.
A partial discharge signal looks like the figure below.
Figure 4. Partial discharge signal
When you take the Wavelet Transform of a signal like this, parts of major
discontinuity as the peak are preserved by the WT. And as the PD have such feature, it
helps to identify them. In the figure below is shown what happens to these signals when
we use four and six levels of decomposition.
Figure 5. Decomposition by WT of a PD signal using 4 levels
Figure 6. Decomposition by WT of a PD signal using 6 levels
And when we have a signal with more than one partial discharge spatially
distributed, we can see their peaks clearly as we go over the levels. One example of this
is shown below.
Figure 7. Maxima lines
Analyzing the last figure, it becomes clearer the idea of maxima lines and what
they represent. They are lines that go over the levels of the WT picking the maximum
value as shown in the figure 7.
But these signals don’t have any kind of noise and it is easy to see a partial
discharge. However, when we deal with real signals with some noise, the task of
separate one from anther becomes difficult. But we know that the PDs are spatially
distributed along the signal which is different from the noise that exists all over the
signal. So, to do this task of separate things, we implement a classifier to deal with the
problem.
Then, to build the classifier, I used some labeled signals. The training data
consist of six columns (related to the six levels of the WT) and seventh column for the
label (1 for PD signals and 0 for noise).
Figure 8. Training data
SVM classifier
For train the Support Vector Machine, there are some data with different types
of noise so we can analyze the efficiency of the work.
For the harmonic noise test:
• Confusion matrix [
]
• Classification rate
For the pulse noise test:
• Confusion matrix [
]
• Classification rate
And for a sample obtained from Furnas:
• Confusion matrix [
]
• Classification rate
We can see that the classification rate for the pulse noise test was the worst of
all. After some thinking, I think the reason for that is because pulse noises are easily
mistaken with PDs.
These results have a good classification rate value but it is really good just when
we rebuild the signal deleting the noise part and be able to see the PDs along the
samples.
For that reason, I did the step 5 and 6 of the process and the result is shown
below.
Figure 9. Filter
Looking more closely we can see the work done by the filter and note that the
classifier did a great job indeed.
Figure 10. Filter zoomed
MLP classifier
For the Multilayer Perceptron, I used the professor’s code bp.m to train the
network and to get some test results.
I did some modifications on the code (mainly the bp.m file and the bpconfig.m
file) so I could run it in a loop and identify the best configuration for the job. Then, after
some tests, there was one configuration with good and stable results: using 2 layers
(excluding the input one), 5 neurons on each layer, a α = 0 and a momentum of 0.1.
The results for the same tests did for the SVM classifier are shown below.
For the harmonic noise test:
• Confusion matrix [
]
• Classification rate
For the pulse noise test:
• Confusion matrix [
]
Classification rate
And for a sample obtained from Furnas:
• Confusion matrix [
]
• Classification rate
And we can see that the classification rate for the pulse noise was again the
worst of all tests.
Although the classification rates weren’t that bad, I noticed (while I was run the
code) that the MLP classifier had very different results (with very different
classification rates) for the same problem what made think that this method is not very
stable, which makes the SVM a good choice since it always have the same result for the
same problem.
And also, the SVM classifier present really good results, up to 93.5 % while the
MLP classifier on got to 87 %, which is worse than the worst result from the SVM.
Therefore, I could conclude that the SVM classifier is the best choice for this
filtering problem and it was easier to use it too.
References
• [1] MOTA, H., Sistema de aquisição e tratamento de dados para monitoramento
e diagnóstico de equipamentos elétricos pelo método das descargas parciais
(Acquisition system and data processing for monitoring and diagnostic of
electrical equipment by the method of partial discharges). Universidade Federal
de Minas Gerais (UFMG), Electrical Engineering Graduate Program. Belo
Horizonte, Minas Gerais, Brazil, March of 2001.
• [2] MOTA, H., Processamento de sinais de descargas parciais em tempo real
com base em wavelets e seleção de coeficientes adaptativa espacialmente
(Signal processing of partial discharges in real time based on wavelets and
selection of spatially adaptive coefficients). Universidade Federal de Minas
Gerais (UFMG), Electrical Engineering Graduate Program. Belo Horizonte,
Minas Gerais, Brazil, November of 2011.