wavelet transform-based impulse fault pattern recognition in distribution transformers
TRANSCRIPT
1588 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 18, NO. 4, OCTOBER 2003
Wavelet Transform-Based Impulse Fault Pattern Recognitionin Distribution Transformers
P. Purkait, Member, IEEE,and S. Chakravorti, Senior Member, IEEE
Abstract—This letter presents a novel approach using thewavelet transform to classify the patterns inherent in differentfault currents during impulse testing of distribution transformers.Results for simulated models of a range of distribution trans-formers are presented to illustrate the ability of this approach toclassify insulation failures under impulse stresses.
Index Terms—Distribution transformer, impulse fault simula-tion, pattern recognition, wavelet transform.
I. INTRODUCTION
ONE of the standard tests carried out on a transformer isthe lightning impulse test after assembly, for assessment
of the ability of its winding insulation to withstand the surgevoltage. Conventional fault diagnosis methods include study ofthe winding voltage and currents in time and frequency do-mains [1]. The nonstationary winding current signals (i.e., sig-nals whose frequency varies with time), can be analyzed in thetime-frequency domain using wavelet transforms such that it ispossible to precisely locate in time all sudden variations in thesignal. This letter presents simulation results of an entirely newapproach, based on the time-frequency analysis of the signalsfor classification of impulse faults from winding current wave-forms. The study has been based on impulse faults simulated inelectromagnetic transient program (EMTP) models of a rangeof distribution transformers.
II. WAVELET ANALYSIS
The Fourier transform (FT), though being widely used forobtaining the frequency spectrum of a signal, only reveals themagnitude of each frequency component in the signal. On theother hand, wavelet transform [2] can depict what spectral com-ponent[s] occurs at what time instant. Wavelet transform is ca-pable of providing the time and frequency information simul-taneously; hence, giving a time-frequency representation of thesignal.
The continuous wavelet transform (CWT) of the square inte-grable function with respect to the mother wavelet isdefined as
(1)
As seen in (1), the transformed signal is a function of twovariables, the scale and translation and parameters which
Manuscript received March 30, 2003.P. Purkait is with the School of ITEE, University of Queensland, Qld-4072,
Australia (e-mail: [email protected]).S. Chakravorti is with the Department of Electrical Engineering, Jadavpur
University, Kolkata 700032, India (e-mail: [email protected]).Digital Object Identifier 10.1109/TPWRD.2003.817845
TABLE ITRANSFORMERDATA FOR EMTP MODEL
TABLE IIDIFFERENTFAULTS SIMULATED
are related to the frequency content and the corresponding timesin the wavelet transform domain, respectively.
III. SIMULATION
EMTP-based high-frequency models of a range of distribu-tion transformers, as summarized in Table I, have been devel-oped for the present study. In all of the EMTP models, thedelta-connected disc winding of the HV sides of all the threetransformers has been represented by a network with lumpedparameters [3].
In the present simulation work, the transformer windingswere subjected to lightning impulses of standard 1.2/50wave shapes.
A. Types of Fault
Insulation failures during impulse tests may occur betweenthe discs/turns (series fault) or between the winding andearthed components like tank, core, etc. (shunt fault). In thepresent study, the entire winding has been divided into threesections—namely the line-end, the mid-winding, and theearth-end sections—each involving 33.33% of the total lengthof the winding and the different faults are simulated as shownin Table II. Due to an inadequate number of discs in the rangeof distribution transformers considered, the study thus has beenlimited to classification of shunts faults at different sections ofthe winding and series faults as a whole.
Fig. 1 shows a sample plot of the transformer winding currentobtained by the capacitively transferred tank-current method.
IV. WAVELET APPLICATION FORFAULT CLASSIFICATION
The winding current waveforms were convoluted with theMorlet mother wavelet to obtain the wavelet transform. Fig. 2represents the three-dimensional (3-D) surface plots obtained
0885-8977/03$17.00 © 2003 IEEE
IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 18, NO. 4, OCTOBER 2003 1589
Fig. 1. Time-domain current waveform of SHM in 300-kVA transformer.
Fig. 2. 3-D surface plot of the CWT of the BIL current waveform of Fig. 1.
from wavelet transform of time-domain BIL current wave ofFig. 1.
For classification of the different fault current waveform pat-terns, relevant parameters were extracted from the wavelet trans-formed 3-D contours. In this letter, three parameters have beenconsidered viz. the most predominant frequency component, itstime of occurrence, and the corresponding wavelet coefficient.The most predominant frequency component is calculated fromthe scale which has the highest wavelet transform coefficientamong all the translations considered.
V. RESULTS AND DISCUSSION
The three pattern classification parameters as mentionedabove are calculated for all the five transformers. Fig. 3 repre-sents the 3-D scatter plots for these parameters in per-unit scalefor the transformers. A close look at Fig. 3 reveals that the threedifferent faults corresponding to all of the five transformersform distinct clusters. It is evident from the clusters of the 3-Dscatter plot, that the different faults have definite pattern ofvalues of the three fault classification parameters. In order toobtain a more precise time-frequency characterization of thedifferent types of faults, a two-dimensional (2-D) scatter plotof predominant frequency versus the corresponding times in
Fig. 3. 3-D scatter plot of parameters used for pattern classification of all kindsof faults in all of the five transformers.
Fig. 4. 2-D scatter plot of time (per-unit) versus frequency (per-unit)corresponding to different faults for all of the five transformers.
per-unit basis, extracted from the 3-D scatter plot is presentedin Fig. 4. The different clusters thus formed by the differentparameters corresponding to all of the transformers are markeddistinctively in Fig. 4.
VI. CONCLUSION
The present letter demonstrates a novel technique based onwavelet analyzes for classification of complex current wave-forms arising due to shunt as well as series type of impulse faultsin distribution transformers. 3-D as well as 2-D scatter plotsusing appropriate classification parameters show that differenttypes of fault in the transformers form clearly delimited sep-arate clusters. These preliminary simulation results presentedhighlight the applicability of the proposed method over conven-tional methods as a possible future tool for impulse fault patternanalysis.
REFERENCES
[1] P. Purkait and S. Chakravorti, “Time and frequency domain analyzesbased expert system for impulse fault diagnosis in transformers,”IEEETrans. Dielect. Elect. Insulation, vol. 9, pp. 433–445, June 2002.
[2] I. Daubechies, “The wavelet transform, time-frequency localizationand signal analysis,”IEEE Trans. Inform. Theory, vol. 36, no. 5, pp.961–1005, Sept. 1990.
[3] F. de Leon and A. Semlyen, “Complete transformer model for electro-magnetic transients,”IEEE Trans. Power Delivery, vol. 9, pp. 231–239,Jan. 1994.