WAVELET TRANSFORM ANALYSIS OF PARTIAL DISCHARGE SIGNALS. B.T. Phung, Z. Liu, T.R. Blackburn and R.E. James

Transcription

1 WAVELET TRANSFORM ANALYSIS OF PARTIAL DISCHARGE SIGNALS B.T. Phung, Z. Liu, T.R. Blackburn and R.E. James School of Electrical Engineering and Telecommunications University of New South Wales, Australia Abstract: Partial discharge (PD) signals are short transient pulses which occur randomly. For such non-stationary signals, the wavelet transform (WT) is more suitable than the traditional Fourier transform (FT) as it provides information in both time and frequency domains. This paper investigates the WT technique as applied to PD signals, with particular reference to discharge measurements in high-voltage power cables. Laboratory tests were carried out to obtain the signal characteristics of various insulation defects and compared with on-site measurement results. Also it is shown that the WT method can be used to suppress noise and improve the time resolution of the PD pulse signal.. INTRODUCTION Electrical insulation plays a critical role in the performance of high-voltage apparatus. Power equipment failure is mostly due to the breakdown of the insulation. This in turn is often the consequence of gradual and cumulative damaging effects of partial discharges (PD) on the insulation over the years. The presence of partial discharges indicates some defect within the insulation of HV power equipment. Furthermore, the nature and type of the defects influence the PD pulse charactersitics. There are a number of diagnostic methods, such as statistical characterisation of the phase-resolved patterns, for PD identification [,2]. Recently, the wavelet transform (WT) has emerged as a powerful technique for analysing transient and suddenly changing signals. References [3-5] are some examples of its application in the area of power engineering. In this paper, the wavelet transform is utilised to characterise the PD signals with particular reference to on-line monitoring of high-voltage power cables [6]. The possibility of differentiating between internal discharges in the cable and external discharges such as coronas based on their wavelet characteristics is explored. Comparison with the conventional Fourier analysis is also provided. A number of different insulation defects in cables are tested in the laboratory. The results are compared with those from on-line measurements in the substations. For on-site testing, random background noise can significantly reduce the sensitivity of PD detection. The use of WT to denoise the signal is also investigated. 2. WAVELET TRANSFORM The well-known Fourier analysis decomposes a signal into consituent sinusoids of different frequencies (fundamental and harmonics). Such a transformation of the signal from the time domain to the frequency domain causes its time information to be lost. This is undesirable, particularly when the signal is non-stationary and its transitory characteristics are important. To overcome this problem, the Short-Time Fourier Transform (STFT) analyses the signal over a short time window at a time and the signal is mapped into a two-dimensional function of time and frequency. Wavelet analysis is an extension of STFT which allows variable-sized window and produces a timescale view of the signal (instead of time-frequency) [7]. In essence, the technique decomposes a signal into shifted and scaled versions of an original (or mother) wavelet. Shifting a wavelet means delaying or hastening its onset. Scaling a wavelet simply means stretching or compressing it. There is a correspondence between scaling and frequency. Low scale produces a compressed wavelet suitable for rapidly changing details and thus corresponds to high frequency. Similarly, high scale gives a stretched wavelet characterising slowly changing features and thus corresponds to low frequency.

2 The continuous wavelet transform (CWT) is mathematically expressed as folows: t b W ( a, b) = ψ. f ( t)dt a a where f(t) denotes the signal, ψ(t) the mother wavelet, a the scale (or frequency) parameter and b the time (or position) parameter. Typically, the wavelet used is an oscillatory irregular waveform of limited duration and has an average value of zero. The CWT generates many wavelet coefficients. The consituent wavelets of the original signal are obtained by multiplying each coefficient with the corresponding scaled and shifted wavelet. Computing the coefficient for every scale and position is time demanding. A more efficient but equally accurate scheme is to select nd positions based on powers of two. This is called the discrete wavelet transform (DWT). The Mallat algorithm is an implementation of the DWT using filters. Further mathematical details of wavelet analysis can be found in numerous textbooks, e.g. [7]. The computations in this paper made use of the Wavelet Toolbox [] which is a collections of functions run under the MatLab environment. 3. MEASUREMENT SETUP A number of kv and kv XLPE cable samples, about 2.5m long each, were tested in the laboratory. These cables were properly terminated with stress relief sleeves to prevent surface discharges at the ends. Fig.: Experimental setup in the laboratory. The experimental setup is shown in Fig.. The PD sensor is a current transformer with a soft ferrite toroidal core, suitable for wide-band measurement (0kHz-200MHz). It is inserted around the conductor that connects the cable sheath to Earth. Discharges in the cable will result in HF electrical pulses propagating through the earth conductor and thus can be detected with the clip-on CT. The output signal from the sensor is amplified, captured with a fast digital oscilloscope with long memory and transferred to the PC for processing. This monitoring system is called the CDA2 [2]. The overall detection bandwidth of the system is from 0kHz to 30MHz. Measurement in the laboratory also includes the conventional circuit using a blocking capacitor in series with a measuring impedance and the ERA discharge detector (Fig.). The same HF-CTs are used for on-site monitoring. Attachment and removal of the sensors can be easily carried out while the cables are energised. Fig.2 shows the sensors monitoring the three-phase 32kV XLPE cables in a sub-station. HF-CT sensor 32kV cable Fig.2: On-line monitoring of 32kV cables. 4. SIGNAL DE-NOISING The most challenging problem associated with on-site testing is signal corruption caused by the random background noise and interference from various sources such as corona discharges from sharp metallic protrusions, switching transients, coupling of signals between the three phases and pick-ups from radio transmission. Techniques to overcome some of these interference sources have been explored in [6]. For example, rejection of the coupled pulses from the neighbouring phases can be achieved by using simultaneous multi-channel recording and postprocessing software. The latter comprises phaseposition windowing and signal magnitude comparison. Filtering of the random background noise can be carried out using wavelet analysis. As an example, Fig.3(a) shows the original noisy signal which was recorded on-line at a 32kV sub-station. Here, the CDA2 monitoring system was able to detect the

3 presence of intermittent but very large discharges. After rejecting the inter-phase signal coupling, the CDA2 phase-resolved patterns clearly show the disturbance originating from the Yellow phase. The next step is to locate the source of the disturbance in that phase. This was done using the ultrasonic detector. The fault was found to be at the top of the HV bushing connected to the 32kV Yellow phase cable termination (see Fig.2) barely visible against the continuous random background noise as shown in Fig.3(a). Note that the horizontal axis corresponds to 0000 data points of 2ns sampling interval. To improve the accuracy of the propagation time measurement, wavelet decomposition can be used to remove the high-frequency noise from the signal. Successive approximations become less noisy as more high frequency information is filtered out. Thus this provides a simple method to de-noise the signal. Fig.3(b) shows the de-noised signal using level-5 approximation and Daubechies db3 wavelet. In comparison to the original signal, it is much cleaner and the reflected pulse can be clearly seen. The measured reflection time corresponds to twice the cable length. This indicates the fault is at the cable termination rather than inside the cable and thus agrees with the ultrasonic detector finding (a) (b) One disadvantage of the above method is that the fast changing features of the original signal is lost. Note the smoothing effect on the wavefront in Fig.3(b). This would reduce the accuracy of the measurement of the time delay between the first pulse and its reflection. An elegant alternative to overcome this problem is the technique called thresholding whereby the details are discarded only if the magnitudes exceed a certain limit. The procedure is to examine the details vectors of the wavelet decomposition, select the appropriate threshold coefficients and reconstruct the new details signals. The toolbox provides two calling functions: one to calculate the default threshold parameters and the other to perform the actual de-noising. Applying these functions, the result is shown in Fig.3(c). It retains well the sharp detail of the original but is somewhat noisier. It may be possible to improve the result by trying other thresholds Fig.3: (a) Original and (b,c) Denoised signals. If the reflected pulse can be detected, the PD location can also be confirmed by time-domain reflectometry. Due to signal attenuation by the cable (7m long), the reflected pulse is much smaller in magnitude. It is (c) 5. SIGNAL CHARACTERISATION Cables with external and internal discharges were tested in the laboratory. The results are plotted in Fig.4. Shown for each type of discharge are its timedomain waveform, magnitude and power spectra of the Fourier transform. Fig.4(a) corresponds to external corona discharge caused by a sharp protrusion on the HV conductor. The magnitude is about 70pC at applied voltage of 4kV. Fig.4(b) corresponds to external surface discharges with a magnitude of pc at 4kV. Although these discharge sources are external, they are very close to the cable (less than cm from the cable termination). Similarly, the characteristics of internal discharges

4 are shown in Figs.4(c) and 4(d). These correspond to surface discharges on the cable termination (90pC at 6kV) and internal discharges from a floating needle puncturing the cable insulation (0pC at 4kV). It can be seen that different types of fault generate somewhat different signatures. In particular, the last case appears to have more higher frequency content in the spectra. This characteristic may be utilised by the

5 (a) (b) 5 5 (c) Fig.4: Time domain waveforms and spectra of PD signals (laboratory tests) Fig.5: Time-scale view of the absolute CWT coeffcients for the signals in Fig.4. 5 (d)

6 (a) (b) 5 (c) Fig.6: Time domain waveforms and spectra of signals obtained from on-line monitoring (d) Fig.7: Time-scale view of the absolute CWT coefficients of the signals obtained from on-line monitoring.

7 monitoring system for interference rejection. On the other hand, very little difference can be seen with the other cases and thus it would be difficult to distinguish between external discharges and those occur at the cable termination. Fig.5 shows a different way to view the signals of Fig.4 in the time-scale where the absolute values of the CWT coefficients are plotted. Note that the smaller coefficients appear as darker pixels on the plot and vice versa. The low scale region corresponds to high frequency. In comparison, Fig.5(d) has the brightest band in this region which agrees with its Fourier equivalent shown in Fig.4(d). Fig.6 is a sample of some typical waveforms and spectra from on-line monitoring of 33kV cables in a substation. These cables are about 70m long each, connecting the main 33kV bus bars to a 32/33kV transformer. The largest PD magnitude is of the order of 2nC. It can be seen that Fig.6(b) is very similar to Fig.4(a) and thus one may conclude that it is external interference from corona discharges. Fig.6(a) and Fig.6(d) reveal a larger content of higher frequency components which match Fig.4(d) and thus indicating internal discharges. The case of Fig.6(c) is distinctly different. It is a slow wave, the spectrum peaks at about 4MHz. The corresponding time-scale view of Fig.6 is shown in Fig.7. The slow wave results in a uniformly dark band in the low scale region of Fig.7(c). 6. CONCLUSIONS In this paper, the PD signals were characterised using the traditional Fourier transform and the new technique of Wavelet transform. The time-scale presentation is an interesting and informative way to view the PD signals. Based on the analysis obtained thus far, it is thought that the application of WT results in no significant improvement in highlighting the subtle differences between the different types of PD signals (as compared to FT). On the other hand, it has been shown that the WT technique can be utilised to de-noise the PD signals and thus enhance the detection sensitivity.. REFERENCES [] R.E. James and B.T. Phung, "Development of Computer-Based Measurement Systems for Recording and Analysis of Partial Discharge Patterns", IEEE-DEI Trans., Vol.2, No.5, Oct. 5, pp.3-6. [2] B.T. Phung, Computer-based Partial Discharge Detection and Characterisation, Ph.D thesis, University of NSW, Australia, 7. [3] Y. Quan, N. Gao, G. Zhang and Z. Yan, Wavelet transform applying in partial discharge measurement, IEEE Int. Symp. On Electrical Insulation, Virginia, USA, June 7-0,, pp.42-. [4] P. Kang and D. Birtwhistle, Wavelet transform based envelope extraction for condition monitoring of power switching equipment, Australasian Universities Power Eng. Conf., Hobart, Sep.27-30,, pp [5] C.T. Wai, Q. Li and W.W.L. Keerthipala, Transformer inrush current investigation based on wavelet analysis, Australasian Universities Power Eng. Conf., Hobart, Sep.27-30,, pp [6] B.T. Phung, Z. Liu, T.R. Blackburn and R.E. James, On-line partial discharge measurement on HV power cables, th Int. Symp. on HV Engineering (ISH), London, Aug.23-27, 9. [7] C.K. Chui, An Introduction to Wavelets, Academic Press, 2. [] M. Misiti,, Y. Misiti, G. Oppenheim and J. Poggi, Wavelet Toolbox for Use with MatLab, The Math Works Inc., ACKNOWLEDGMENT The project on on-line condition monitoring of HV power cables is financially supported by Integral Energy. Assistance from its staff during site testings, in particular P. Taylor and J. Hickey, is gratefully acknowledged.