RECOGNIZING PARTIAL DISCHARGE FORMS MEASURED BY THE

Document Sample
RECOGNIZING PARTIAL DISCHARGE FORMS MEASURED BY THE Powered By Docstoc
					Molecular and Quantum Acoustics vol. 26, (2005)                                             35



RECOGNIZING PARTIAL DISCHARGE FORMS MEASURED BY THE
  ACOUSTIC EMISSION METHOD USING THE SPECTRUM POWER
     DENSITY AS A PARAMETER OF THE ARTIFICIAL NEURON
                                       NETWORK

       Tomasz BOCZAR, Sebastian BORUCKI, Andrzej CICHOŃ, Marcin LORENC

     Technical University of Opole, Faculty of Electrical Engineering and Automatics,
    Institute of Electric Power Engineering, ul. Sosnkowskiego 31, 45-272 Opole, Poland

        tboczar@po.opole.pl, sborucki@po.opole.pl, acichy@po.opole.pl, lem@op.pl


     The subject matter of the paper refers to the next stage of the research work
     connected with the improvement of the acoustic emission (AE) method used for
     the evaluation of partial discharges (PDs) generated in paper-oil insulation
     systems of high-voltage power appliances. The paper presents research results
     referring to use of artificial neuron networks (ANN) for recognizing basic PD
     forms which can occur in paper-oil insulation weakened by aging processes.
     Describe systems of spark-gaps for modeling basic PD forms and a system for
     registration of acoustic signals generated by the assumed PD forms. Next, based
     on the signals registered and using the power spectrum density (PSD), the
     analysis of the effectiveness of recognition of the particular PD forms by the
     implemented neuron network was carried out.



1. INTRODUCTION
     The research work on the use of the AE method for the evaluation of the condition of
insulation systems of power appliances began in late 60s of the twentieth century [12, 13]. At
present this method constitutes a significant complement of measurement methods used in
diagnostics of insulation systems. It provides information on the existence, size, and first of
all on the place of PD occurrence, which cannot be obtained by using other methods. In recent
years the development of the AE method has been caused mainly by the improvement of the
systems that enable the measurement and analysis of the AE signals and in which the latest
achievements of digital electronics and computer technology are used [3, 14]. Currently the
problem is more and more often not the measurement taking of the AE generated by PDs, but
36                                                  Boczar T., Borucki S., Cichoń A., Lorenc M.

a proper analysis and interpretation of the results obtained, and in consequence, a proper
evaluation of the phenomena under study.
       The subject matter of this paper refers to one of the aspects of the analysis of the signals
registered, that is a correct and effective recognition of the AE signals generated by basic PD
forms. The AE signals registered can be closely connected with basic PD forms presented in
the literature [1, 2]. The particular PD forms can be identified with the type and the degree of
damage of paper-oil insulation. Therefore, thanks the correct process of recognizing the AE
signals registered, coming from the particular PD forms, it is possible to identify the type of
damage to an insulation system and to evaluate preliminarily a damage degree of this
insulation.
       The research work carried out so far on the correct recognition of basic PD forms has
been based mainly on comparing graphic representations of the selected parameters of the AE
signals registered (characteristics of the amplitude spectrum, characteristics of the power
spectrum density) and on the analysis of descriptors that represent them ( shape coefficient,
peak coefficient, median frequency). It caused a significant time extension of the
measurement result interpretation and the evaluation of the type and degree of damage to
insulation. The application of artificial neuron networks in the process of recognizing basic
PD forms, through a parallel data processing, caused a considerable acceleration of this
process. The research work carried out constitutes the next step in building a diagnostic
system based on the AE method which enables a correct evaluation of the paper-oil insulation
condition.

2. PD FORMS UNDER STUDY
       Due to a big complexity of processes connected with generation and propagation of the
AE signals emitted by PDs an experimental procedure was applied and the experiments were
carried out in laboratory conditions using spark-gaps that enable modeling of basic PD forms
[6, 7, 8]. Based on both literature information and their own research work, the authors of the
paper isolated the following basic PD forms:
     1. discharges in the point-point system in oil, which can model PDs that occurred due to
        insulation damage of two neighboring windings of the transformer winding,
     2. discharges in the point-point system in oil with gas bubbles, which can model PDs in
        gassy oil and are caused by insulation damage of two neighboring transformer
        windings
     3. discharges in the point-plane system in oil, which can model PDs in occurring
        between a damaged part of a transformer winding insulation and grounded flat parts
        (elements of the tub),
Molecular and Quantum Acoustics vol. 26, (2005)                                             37

   4. discharges in the surface system of two flat electrodes with paper-oil insulation
       between them; the most common PD form occurring in the so-called triple point, in
       which the electrode surface touches solid and liquid dielectrics,
   5. discharges in the surface system of one flat electrode and the other multipoint
       electrode with paper-oil insulation between them; different distribution of the electric
       field intensity compared with discharges in the surface system with two flat electrodes,
   6. discharges in the multipoint-plane system in oil, which can model PDs occurring
       between a multipoint insulation damage of a transformer winding and grounded flat
       parts (elements of the tub),
   7. discharges in the multipoint-plane system in oil with gas bubbles, which can model
       PDs occurring between a multipoint insulation damage of a transformer winding and
       grounded flat parts (elements of the tub), but in oil with gas particles,
   8. discharges on particles of an indefinite potential that move in oil, which can model
       PDs occurring in oil containing particles of cellulose fibres formed in the process of a
       gradual degradation of paper-oil insulation caused by aging processes.

3. REGISTRATION SYSTEM OF THE AE SIGNALS GENERATED BY PDS
     The diagram of the system for PD generation and registration of the AE signals is
shown in Fig. 1. For generation of the assumed PD forms there were used spark-gaps that
modeled them and which were placed in a transformer tub filled with electroinsulation oil.
The spark-gaps were supplied with alternating voltage of power frequency and rms voltage
equal to 0.8 Up (breakdown voltage) of each of the systems. A generator (GP) was used to
produce gas bubbles. It’s nozzle generating repeatable in respect of shape and size bubbles
was placed under the spark-gap generating the assumed PD form in such a way that the
bubbles emitted every 0.1 s on the average were directed into the space between the
electrodes of the spark-gap. To model PDs on particles of an indefinite potential a modeling
multipoint-plane spark-gap was used, which was immersed in electroinsulation oil containing
cellulose fibers. The volume density of the cellulose fibers contained in the oil used for
examination was 10 mg/dm3 on the average.
38                                                Boczar T., Borucki S., Cichoń A., Lorenc M.

                                       1




                 WNZ          WN



                                   4                                                  6
                                                          5
                  2


            3




Fig. 1. Diagram of the measuring set-up (1 – transformer tub filled with electroinsulation oil,
2 – spark-gap modeling one of the assumed PD forms, 3 – generator of gas bubbles (GP), 4 –
measuring transducer, 5 – amplifier and a measuring filter, 6 – computer with a measuring
card)

     The AE signals generated by PDs were measured with a piezoelectric WD AH17
transducer by the PAC firm, which was attached to the tub. The transducer used is
characteristic of a good sensitivity (55 dB ± 1.5 dB in reference to V/ms-1) and of a wide
transfer band from 100 kHz to 1MHz in the range of ± 10 dB [11]. In order to amplify the
measuring signal the outputs of the WD AH17 transducer were connected with the differential
inputs of the AE Signal Conditioner amplifier by the firm AE System. The amplifier has a
stable 40 dB amplification and the transfer band (0 ÷ 1.5) MHz. Additionally the system is
equipped with a band-pass filter of the cut-off frequencies of 10 kHz and 700 kHz. The
application of the above-mentioned filtering band is necessary due to the elimination of the
disturbing signals occurring in the lower and upper frequency band and also the elimination of
the phenomenon of aliasing [5].
     For observation and registration of the AE signals measured a computer equipped with a
measuring card type NI 5911 by the firm National Instrument and a specialized Virtual Bench
Scope software were used. The sampling frequency equal to 2.56 MHz, which translated into
a 14-bit resolution of the A/C transducer, was assumed for measurements. The time of a
single measurement was 20 ms.

4. RECOGNIZING THE PARTICULAR PD FORMS BY THE ANN USING THE POWER
   SPECTRUM DENSITY
     The Matlab environment was used for implementing, teaching and testing the neuron
network, used in the process of recognizing the particular PD forms. In view of the literature
[9, 10] on the use of neuron networks as classifiers and the tools recognizing models, a
Molecular and Quantum Acoustics vol. 26, (2005)                                               39

unidirectional three-layer network of the type Feed-Forward Backpropagation recognizable
Network (F-F BP) was suggested. For each neuron occurring in the network structure a
sigmoid activation function was determined. As the AE signal parameter during teaching and
testing the network, one of the parameters of the frequency analysis signals – PSD was
suggested. For each of the assumed PD forms a 100 measuring files were registered, of which
part of the files were vectors of the learning set (LS), and the remaining part were vectors of
the test set (TS) – teaching with a teacher. In the process of teaching and testing the network a
series of simulations was carried out, the aim of which was to obtain the best recognition
effectiveness possible of the assumed PD forms. The paper presents research results on the
use of PSD as the parameter characterizing the assumed for analysis basic PD forms and its
use for determining the recognition effectiveness of the particular PD forms by the ANN
implemented.
      In order to determine the recognition effectiveness of the assumed PD forms by the
network created, the concept of a ‘class’ was introduced, which, in this case, defines the
assumed PD forms. Accepting for analysis the eight PD forms listed in the previous chapter,
eight classes were defined: class 1 – discharges in the point-point system in oil, class 2 -
discharges in the point-point system in oil with gas bubbles, class 3 - discharges in the point-
plane system in oil, class 4 - discharges in the surface system of two flat electrodes with
paper-oil insulation between them, class 5 - discharges in the surface system of one flat
electrode and the other multipoint electrode with paper-oil insulation between them, class 6 -
discharges in the multipoint-plane system in oil, class 7 - discharges in the multipoint-plane
system in oil with gas bubbles, class 8 - discharges on particles of an indefinite potential that
move in oil.
      Figure 2 shows the recognition effectiveness of the PD forms under study depending on
the number of recognizable classes (Ikr) and LS size (Rcu) at a constant number of neurons of
the hidden layer.
40                                                         Boczar T., Borucki S., Cichoń A., Lorenc M.

     a)
     S t r [%]                                              b)




                                                             S t r [%]
                                        Ikr                                           Ikr
                        Rcu                                              Rcu


     c)                                                     d)
     S t r [%]




                                                             S t r [%]
                                        Ikr                                           Ikr
                        Rcu                                              Rcu




Fig. 2. Recognition effectiveness of PD forms (Skr) by the network applied depending on the
number of recognizable classes (Ikr) and LS size (Rcu): a) two neurons in the hidden layer, b)
6 neurons in the hidden layer, c) 10 neurons in the hidden layer, d) 14 neurons in the hidden
layer

                 It results from the recognition effectiveness of the particular PD forms shown in Figs
2a, 2b, 2c, 2d that with the increase of the number of recognizable classes and at a constant
number of vectors in LS (number of vectors LS = Rcu x Ikr) the recognition of effectiveness
of the network tested drops. This dependence is shown most clearly in the figure on which the
structure of the hidden layer of the ANN contains 2 neurons (Fig. 2a). In this case the
improvement of effectiveness can be achieved through increasing the number of LS vectors
(the increase of the LS size). This effectiveness, however, is not sufficient from the point of
recognition correctness of a particular PD form, as it does not exceed 85% (8 recognizable
classes and 2 neurons of the hidden layer) The other way of increasing the recognition
effectiveness of the particular forms is to increase the number of neurons of the hidden layer,
which can be observed while analyzing the consecutive diagrams in Fig. 2. In order to
visualize better the analysis results of the influence of the number of neurons of the hidden
layer on the recognition effectiveness of the basic PD forms, Fig. 3 shows the dependence of
the recognition effectiveness of the particular forms on the number of neurons of the hidden
layer and the number of recognizable classes. The LS size was assumed as a constant
parameter of this analysis.
Molecular and Quantum Acoustics vol. 26, (2005)                                                    41

   a)
    Skr [%]                                               b)




                                                           Skr [%]
                                      lnwu                                           lnwu
                      lkr                                            lkr


   c)                                                     d)
    Skr [%]




                                                           Skr [%]
                                      lnwu                                           lnwu
                      lkr                                            lkr




Fig. 3. Recognition effectiveness of PD forms (Skr) by the network applied depending on the
number of recognizable classes (Ikr) and the number of neurons in the hidden layer (lnwu):
a) Rcu = 10 b) Rcu = 20 c) Rcu = 30 d) Rcu = 40

              Based on the recognition effectiveness of the particular PD forms, presented In Figs 3a,
3b, 3c, 3d, it can be again observed that with the increase of the number of recognizable
classes and a constant number of neurons of the hidden layer, the recognition effectiveness
decreases. The improvement of the recognition effectiveness for the particular number of
recognizable classes can be achieved by increasing the number of neurons of the hidden layer,
e.g. for the eight classes from Fig. 3b the recognition effectiveness, at two neurons, is about
70%. However, for the same number of classes but at ten neurons the achieved effectiveness
was about 93%, which is a satisfying result.
              From the recognition effectiveness dependence by the network of the particular PD
forms, presented in Figs 2 and 3, it results that for a constant number of neurons in the hidden
layer and for a constant number of vectors in LS ( constant LS size) there exists such a point
after exceeding of which there takes place a sudden drop in the recognition effectiveness by
the network of the particular PD forms. It happens so because with the increase of the number
of LS vectors a gradual saturation of the particular neuron weights of the network neurons
takes place, which leads to the convergence loss of the network teaching process and
manifests itself directly with a significant drop in recognition effectiveness. From the
characteristics presented it also results that increasing the number of neurons in the hidden
layer we can use a smaller LS size to obtain similar values of recognition effectiveness.
              The consecutive graphic interpretation of the research results obtained refers to
determining the influence of the determination accuracy of the power spectrum density
42                                                            Boczar T., Borucki S., Cichoń A., Lorenc M.

(number of points averaging PSD) on the recognition effectiveness of basic PD forms. Fig. 4
shows the obtained recognition effectiveness values of PD forms depending on the LS size
(number of LS vectors) and a changing number of points averaging the power spectrum
density (lpu). The number of neurons of the hidden layer was assumed as a constant
parameter of this analysis.
     a)                                                       b)
      S t r [%]




                                                              S t r [%]
                                              lpu                                             lpu


                            Rcu                                            Rcu


     c)                                                       d)
     S t r [%]




                                                              S t r [%]




                                               lpu                                            lpu


                            Rcu                                            Rcu




Fig. 4. Recognition effectiveness of PD forms (Skr) by the network used depending on the LS
size (Rcu) and a changing number of points averaging the PSD (lpu) for 8 recognizable
classes: a) 2 neurons in the hidden layer, b) 6 neurons in the hidden layer, c) 10 neurons in the
hidden layer, d) 14 neurons in the hidden layer

                  The diagrams of the simulations carried out, presented in Fig. 4, confirm the results
obtained earlier and made the authors more convinced that many factors have influence on
obtaining the highest values possible of the recognition effectiveness of the PD forms
assumed. The number of neurons in the hidden layer proved to be significant, as either too
small or too big numbers of them cause that the effectiveness obtained is not at a satisfying
level (below 90%). Also the number of LS vectors (LS size) is significant in the process of
teaching as its too big size causes weight saturation of particular neurons, which is manifested
directly with a sudden loss of stability of the teaching process and decrease of the recognition
effectiveness.
                  The results shown in Fig. 4 lead to the conclusion that 128 PSD points are sufficient for
satisfying recognition effectiveness (above 90%) for 8 classes passed simultaneously on the
input layer of the ANN. For a lower number of the points averaging PSD the effectiveness is
lower than 90%, and the increase of the PSD points above 128 insignificantly increases the
recognition effectiveness, but it causes a significant elongation of the teaching process and the
process of recognition of the particular PD forms.
Molecular and Quantum Acoustics vol. 26, (2005)                                             43

5. CONCLUSION
     The analysis of the research work results carried out confirms the possibility of applying
an ANN for recognizing PD forms measured by the acoustic emission method. The kind of
the neuron network adopted – F-F BP of a three-layer structure made it possible to recognize
very well the particular PD forms. This is confirmed by the recognition effectiveness results
of the AE signals coming from PDs presented in this paper.
     The research work carried out also proved the usefulness of the PSD – the parameter
representing an AE signal as a criterion of teaching and testing the neuron network applied. In
order to obtain the recognition effectiveness of the particular PD forms at the level exceeding
90% (at 8 recognizable classes) the number of neurons of the hidden layer should be at least
10 and the number of vectors in LS should be from 240 to 320 (LS size from 30 to 40).
In the further stage of research the works on the improvement of the effectiveness of the
recognized PD forms by the implemented neuron network will be carried on. The research
methodology will be based on the change of the network elements and on the selection of a
different parameter determining the AE signals registered.

  The research work is co-finances by the European Social Fund and the state budget

REFERENCES

1. Boczar T., Widma emisji akustycznej generowanej przez wyładowania niezupełne w
   izolacji olejowej, SiM, z. 114, Politechnika Opolska, 2000.
2. Boczar T., Obiektywizacja wyników akustycznej metody oceny wyładowań
   niezupełnych przy zastosowaniu do opisu sygnałów analizy statystycznej i cyfrowych
   metod przetwarzania, Of. Wyd. Politechniki Opolskiej, 2003 r.
3. Boczar T.: Możliwości zastosowania do opisu sygnałów emisji akustycznej od wyładowań
   niezupełnych analizy statystycznej i cyfrowych metod przetwarzania sygnałów,
   Politechnika Opolska, Of. Wyd., 2003 r.
4. Boczar T., Borucki S., Cichoń A., Lorenc M. „Rozpoznawanie wyładowań elektrycznych
   rejestrowanych metodą emisji akustycznej przy wykorzystaniu sieci neuronowych”,
   Przegląd Elektrotechniczny –konferencje 1’2005, Jubileuszowe X Sympozjum EUI’2005,
   „Problemy eksploatacji układów izolacyjnych wysokiego napięcia”, Krynica, 27 – 30
   września 2005, str. 30-33
5. EA System, AE Signal Conditioner – instrukcja obsługi, Warszawa 2001
6. Fleszyński J.: Badania mechanizmu przebicia elektrycznego cieczy izolacyjnych. ZN
   Polit. Wrocławskiej. Wrocław 1986.
7. Florkowska B., M. Florkowski, R. Włodek,. P. Zydroń, „Mechanizmy, pomiary i analiza
   wyładowań niezupełnych w diagnostyce układów izolacyjnych wysokiego napięcia”,
   Wydawnictwo IPPT PAN, Warszawa, 2001.
8. Gacek Z., „Wysokonapięciowa technika izolacyjna”, Wydawnictwo Politechniki Śląskiej,
   Gliwice, 1996.
44                                              Boczar T., Borucki S., Cichoń A., Lorenc M.

9. Nałęcz M.,Duch W.,Korbicz J. Rutkowski L., Tadeusiewicz R.,
   Biocybernetyka i inżynieria biomedyczna 2000, tom 6 Sieci neuronowe, Akademicka Of.
   Wyd., 2000.
10. Osowski S., Sieci neuronowe w ujęciu algorytmicznym, WNT, Warszawa, 1996
11. PAC, Acoustic emission broad-band transducer type WD AH 17, 2004
12. Skubis J.: Wybrane zagadnienia z techniki i diagnostyki wysokonapięciowej, Politechnika
    Opolska, Of. Wyd. 1998 r.
13. Skubis J., Gronowski B., Ranachowski J., Zalewski J ., Pomiary tłumienia
    emisji akustycznej od wyładowań elektrycznych w materiałach elektroizolacyjnych, IPPT
    PAN Warszawa 1986r.
14. Zydroń P.: Wybrane zagadnienia analizy czasowej i częstotliwościowej wyładowań
    niezupełnych, Akademia Górniczo-Hutnicza - Wydaw., 2001.