Proceedings of the 7th International Conference on Properties and Applications of Dielectric Materials
lune 1-S 2003 Nagoya
PD Pattern Recognition of Power Capacitors Model Based on Combinational Neural Network *
Shengyou Gao Fuqi Li Canghi Yu Kexiong Tan Department of Electrical Engineering, Tsinghua University Beijing 100084, P.R. of China E-mail: eeagsy@tsinghua,edu.cn
Abstract: Five types of partial discharge (PD) models are designed to represent typical PD phenomena in power capacitors. A computer-based acoustic emission signal detecting instrument that is developed for PD is used to collect a lot of acoustic signals of model discharge. The acoustic signal duration of gas cavity discharge is very short, therefore. it is easy to be differentiated from other discharge patterns. According to the time domain and frequency domain graph of acoustic signals, the characters of oil gap discharge is close to that of discharge along oil impregnated paper surface. However, the character of discharge of surface discharge of bushing. metallic impurity in oil impregnated paper insulation is close to that of Combinational neural network (CNN) is used to recognize the five kinds of representative discharge patterns. The result shows that CNN is effective and the characteristics of acoustic signals could be used to recognize PD patterns in power capacitors.
restricted by capacitance of tested object and is immune to electromagnetic interference. used to locate the failure. widely [I]. Five types of PD models are designed to represent typical phenomena of PD in power capacitors. After obtain some AE signals caused by PD within these models, the waveform characteristic are studied about these AE signals. The gas cavity discharge type could be distinguished from other discharge because of having very short signal duration.
As for other 4 classes, the
Moreover, it could be
Therefore, acoustic emission
method is adopted in PD detection of power capacitor
discharge patterns could be recognized by single artificial neural network and combinational neural network. The results show it is possible to recognize PD pattern of power capacitor, and the combinational neural network is more effective and reliable. AE SIGNALS CAUSED
Key words: power capacitor; partial discharge (PD);
acoustic emission; pattem recognition; combinational neural network (CNN) INTRODUCTION Partial discharge pattern recognition of power capacitor is very important for analyzing insulation status, judging failure type and location, as well as improving designing and manufacturing technology. Having large capacitance, it is difficult to obtain high sensitivity using electrical method to detect PD signal in power capacitor. Acoustic emission (A€) detecting method is not *
Project (No. 5997 701 I) supponcd by NSFC
AE signals caused by PD are sampled by detector based
on computer [ 2 ] . This detector includes AE sensors,
converter card and industrial computer. AE sensor is attached on the tank wall of test model. The captured AE signals are stored in hard disk after being amplified, filtered and sampled in order to analysis and to extract the features. The sampling rate is IO MHz, and the data The electrical signal is length is 524 288 points. reference. The following PD patterns are reflected by 5 kinds of typical models, including (a) gas cavity discharge; (b) oil gap discharge; (c) discharge along the surface of oil
observed at the same time of detecting A € signal as the
impregnated paper; (d) discharge in oil impregnated paper insulation with metallic impurity; (e) discharge along bush surface.
In this paper, a-e represents these
PD types respectively. Figure 1 shows the time domain and frequency domain figures of the 5 types of PO. Based on the waveform characteristics, model discharges finally fall into 2 classes. The first class includes gas cavity discharge, oil gap discharge and discharge along the surface of oil impregnated paper. The signal waveforms of them are shown in (a), (b) and (c) of Fig. I . The characteristic of this type waveform is that the duration time is short (usually less than 2 ms), and the amplitude is high (greater than 200 mV). The waveform rises quickly and falls slowly. Total waveform presents oscillatory wave with attenuation and the spectrum energy centered in the range of high frequency. The second class includes discharge in oil impregnated paper insulation with metallic impurity and bush surface discharge. The signal waves of them are shown in (d) and (e) of Fig. I, The characteristic is that the duration time is longer than the first class (usually 6 - 8 ms), and the amplitude is low (less than 300 mV). The rise time is equal to the fall time and the spectrum energy centers
in the range of low frequency.
PATTERN RECOGNITION TOOLS
Fig.1 Acoustic emission signals and spectra of PD models (a) gas cavity discharge; (b) oil gap discharge; (c) surface discharge along oil impregnated paper; (d) discharge in oil impregnated paper insulation with metallic impurity; (e) hush surface discharge The combinational neural network shown in Fig.2 is adopted in pattem recognition. distinguish class a. Rule 0 is used to Subnet A N N l . ANN2. ANN3 are
If there are many pattems to be recognized or the features of some patterns are similar, it is suitable to adopt combinational neural network as pattem recognition tool . some principle. This kind of neural network
consists of multiple independent networks following To AE signal caused by power capacitor model, the duration time of gas cavity discharge is several microseconds, which could be considered as a rule (ruleO), as shown in Fig. 2 . Other types of AE signals could be classified into 2 classes: b, c and d, e . Different signals in same class have similar features, therefore, combinational neural network is used
used to distinguish type b, c, d and e step by step. Single ANN is used for the purpose of compare. Single hidden layer BP network is used and the number of neurons in hidden layer is 20. To combinational neural network, different subnets use different methods to
recognize PD pattem.
extract features in order to improve recognition ability and efficiency. The feature extracting methods include unite-feature, frequency data compression and classical power spectrum estimation method. Unite feature vector method consists of time domain parameters, frequency domain parameters and AR model parameters. Frequency data compression method uses re-sampling technique and re-samples 1000 data in frequency domain as feature vector. Direct classical power spectrum estimation method is used in this paper.
more other methods, the speed of calculating feature vector and training ANN is slow. When using ANN2 to distinguish type h and c, the result is satisfied. (3) When using features got from power spectrum estimation method as input, single ANN has the recognition rate greater than 88%. ANN is not easy to converge when being trained. When using ANN3 to distinguish type d and e, the result is satisfied and the recognition rate is greater than 98%.
(4) It is important to choose correct structure of
combinational neural network. feature vector.
BP network is used in
each subnet, and different ways are used to extract
Table 1 ANN
Recognitioi ‘ates (“A)of PD Class b c d e ‘ Dim ensi on 16
f or Feature V -
Sketches of CNN’s structure
RECONCNITION RESULT AND ANALYSIS
According to rule 0, the discharge of gas cavity model could be completely recognized.
The other 4 kinds of
discharges are classified by CNN and the result is shown
In this table, the recognition results of using
Extracting method Unite feature
Frequency data 1000 100 :ompression Power spectrum estimation Unite feature Frequency data Power spectrum estimation
single ANN is shown for purpose of compare. The following results could be obtained from the data in Tablel and CNN’s training process:
( I ) when using unite features as network input, single
ANN’S recognition ability to type d and e is not good.
The task of ANNl is to distinguish type b and c from typed and e, therefore it has excellent recognition ability. This result shows the waveforms of type d and e are close, and single ANN probably confuses them. (2) When using compressed data in frequency domain as input, single ANN has the recognition rate greater than 81%. Because the dimension of feature vector is much
Study on partial discharge in typical model shows waveform and spectrum characteristics of AE signal could apply in the field of pattern recognition of PD
in power capacitor. 2. A€ signal cause by PD in power capacitor could be classified into 2 classes. The first class includes gas
cavity, oil gap and oil impregnated paper surface discharge. The other class includes discharge in oil impregnated paper insulation with metallic impurity and bush surface discharge.
3 . The advantage of CNN is that the excellent
Science Foundation of China for the financial support (Project No. 5997 701 1).
[I] L.Ghirelli, W.Koltunowicz, A.Pigini, et al.
recognition' results could be obtained.
At the same
Acoustical method for partial discharge detection in high power capacitors.
lime, the disadvantages of single ANN, such as huge network structure, slow calculation speed and bad convergence effect could be overcomed. 4. The extracting method of feature vector is very important to pattern recognition. The good recognition rate of greater than 98% could be obtained under following conditions: using unite feature to distinguish type b, c and type d, e; then using frequency domain data compressed feature to distinguish b and c, and using power spectrum estimation feature to distinguish d and e
IEE Conf. Publication,
Proc. of the 1nten.Conf. on Partial Discharge, Canterbuy, 1993.09.28-30: 92-93 [21 Shengyou Gao, Lei Zhang, Kexiong Tan, Fuqi Li. Partial discharge acoustic emission detector based
on computer for power capacitors.
Proc. of the
Intern. Symp. on El. Insul. Materials, Himeji Japan, 2001. I I. 19-22: 761-764
 Agganval RK, Xuan QY, Dunn RW. A novel fault
classification technique for double-circuit lines based on a combined unsupervisedlsupervised neural network.
The authors gratefully appreciate the National Natural
IEEE Trans. on Power Delivery,
1999. 14(4): 1250-1255
- 322 -