Power Quality Disturbance Detection and Visualization Utilizing Image Processing Methods

Transcription

1 Proceedings of the 4th International Middle East Power Systems Conference (MEPCON ), Cairo University, Egypt, December 9-2, 2, Paper ID 57. Power Quality Disturbance Detection and Visualization Utilizing Image Processing Methods Hussain Shareef and Azah Mohamed Faculty of Engineering and Build Environment Universiti Kebangsaan Malaysia Bangi, Selangor, Malaysia Abstract - This paper presents a novel technique to visualize and detect various power quality disturbance events. It is based on the image processing methods known as grayscale images and binary images. Gray image created from recorded disturbance voltage waveform is first represented as a transverse wave having compressions and rarefactions. Then using image enhancement techniques, the unique features of the disturbance waveform are visualized. Furthermore, the patterns obtained for a pure sine signal and the signal with disturbances are compared for identification of the signal with disturbance. The decision regarding the disturbance type is made using binary image analysis techniques. Evaluation studies for verifying the accuracy of the method are presented. Index Terms - Power quality. Image processing. Grayscale patterns. Binary images. Disturbance classification. I. INTRODUCTION Due to the stringent demands from the microelectronics industry nowadays, the need for improved power quality (PQ) is also increasing. Poor PQ could cause failure or malfunction of certain equipment and processes. Hence, to guarantee high quality of power and to identify PQ problems, a large number of smart power quality meters have been used in electric power systems and industrial premises. These devices capture and store a huge number of the PQ disturbance events every day. However, sophisticated software and algorithms are required to analyze the captured data. In the past, PQ data was analyzed manually, which is very time consuming and also requires special expertise. Many automatic PQ disturbance identification methods have been proposed in the last few years. Some of the earliest methods that were employed in the characterization of PQ events are based on root mean square (rms), fast Fourier transform and short time Fourier transform (STFT) [-2]. These methods are useful in providing information for signals that are stationary. However, most of the PQ data captured are non-stationary and hence the techniques cannot properly track the signal dynamics. The alternative algorithm of STFT is the wavelet transform. This technique has the capability in extracting information from non-stationary signals. In [3-6] the authors utilize wavelet transform to extract the unique features for fast-changing signals such as switching transients and impulses. Even though the wavelets provide a variable window for low and high frequency currents and voltage waveforms, the performance of the wavelet transform gets affected because of the effect of the noises [7]. To improve the performance of the wavelet, an alternative technique called the S-Transform was developed. The S-transform is equivalent to phase corrected continuous wavelet transform. It is fully convertible from the time domain to the two dimensional frequency translation domain, and to the familiar Fourier frequency domain. Researchers [8-] have utilized S-transform to extract features such as amplitude factor, frequency factor, etc., from the PQ disturbance signals. In [2], the power signal disturbances in time time transformation (TT-transform) are derived from the S- transform. TT-transform is the two dimensional time time representation of a one dimensional time series. TT-transform helps in the interpretation of the S-transform. Based on the features extracted from the aforementioned signal processing techniques, a variety of methods has been adopted for the decision-making stage of automatic classification of PQ disturbances. The authors in [3-5] have utilized the Support Vector Machines (SVMs) to classify the disturbance types of PQ. Similar to SVM, artificial neural network (ANN) approaches have found applications in predicting the type of PQ disturbance [6]. Fuzzy and Rulebased expert system have also been employed for the decision-making step in the process of classifying PQ disturbance types [,, 7-9]. In this paper, a novel method for visualizing and classifying the different types of PQ disturbances is proposed. The proposed method is based on the image processing methods known as grayscale images and binary images. Gray image created from recorded disturbance voltage waveform is first represented as a transverse wave having compressions and rarefactions. During a disturbance such as a transient, the gray patterns of these images reveal the information about the type disturbance, as will be shown in the paper. A brief overview on the application of grayscale and binary imaging methods to represent PQ disturbances is given first. II. CONCEPT OF IMAGES FOR VISUALIZING PQ EVENTS Image processing techniques have been successfully applied in many areas, including medicine, radar, sonar, robotics, 238

2 material science, and indentation to name a few. It addresses binary and grayscale images, as well as color and 3 dimensional imagery. A grayscale image is a data whose values represent intensities within some range. Generally, the intensity represents black and the intensity represents white. Each element of the image data corresponds to one pixel of the image. For PQ disturbance visualization, an image can be created by converting the recorded single phase voltage information from the PQ monitor for each time step by using the following transformation function: f : [,] () The transformation function shown in () involves a twostep procedure. First, the voltage values are normalized and shifted in the positive direction to make them bounded between and. Fig. shows grayscale image obtained for a pure voltage waveform. As seen in Fig., a grayscale image file consists of pixels on 2 dimensional (2D) space. The 2D coordinates of one pixel is composed by treating x-axis as time and y-axis as the pixel intensity. This transformation causes the time varying sinusoidal waveform to appear somewhat like a longitudinal wave with compressions and rarefactions. The compressions can be thought of high intensity regions and the low intensity regions as the rarefactions, as shown in the Fig.. Observe that the transition in gray intensity throughout the image is regular and smooth. This is a unique feature for a pure sinusoidal voltage waveform. However, if there is any disturbance such as a transient in the waveform, then there will be a drastic change in intensity in the disturbance area of the generated image, as shown in Fig. 2. Pixel intensity Pixel intensity (a) pure voltage waveform (b) generated pattern from actual signal Fig. Contraction of grayscale pattern from actual signal. D is tu rb an ce a re a (a) actual disturbance waveform (b) generated pattern from disturbance signal Fig. 2 Contraction of grayscale pattern for a disturbance signal. In the second step of the procedure, the features of the disturbance can be further vivified by applying image enhancement techniques such as gamma corrections, contrast adjustments and other morphological operations. Although the gamma correction and contrast adjustment of the generated grayscale disturbance image improve the vividness of the features, it is not uncommon to use binary images for image analysis. In a binary image, each pixel assumes one of only two discrete values; or. Zeros and ones are interpreted as black and white, respectively. Fig. 3 shows the binary images generated for the pure voltage waveform depicted in Fig. Patterns in Fig. 3b and Fig.3c corresponds to negative and positive part of the actual waveform respectively. A combination of these patterns is shown in Fig. 3d. Observe the equidistance black and white pixels having same cross sectional areas in this pattern. Based on the connectivity of the pixels in a binary image that forms the objects, much information required for the automatic classification of PQ disturbances can be obtained. This information includes area, centroid, Euler number, extreme points, intensity statistics, pixel list of connected objects, etc. III. METHODS USED FOR ENHANCING PQ DISTURBANCE CHARACTERISTICS This section illustrates the different image processing techniques that are utilized in enhancing various types of PQ disturbances from a pure signal. These types of PQ disturbances include sag, swell, notch, voltage flicker, short interruptions, and impulsive and oscillatory transients. There are many image enhancement and analysis methods that can be used for this purpose. However, for best performance, three main methods are used to boost the features of PQ disturbances mentioned above (d) image showing combination of positive and negative patterns (e) scaled waveform Fig. 3 Patterns describing a pure voltage waveform. 239

3 First method mainly uses gamma correction for classifying sag, swell, and flicker, respectively. Secondly, edge detection technique is used for identifying transients, and lastly a morphological valley finding process is used to detect the notches and the short interruptions. A. Gamma Correction Gamma correction is a non-linear adjustment to individual pixel values. While in image normalization, linear operations carried out on individual pixels, such as scalar multiplication and addition or subtraction, gamma correction carries out a non-linear operation on the source image pixels. For a more mathematical analysis, consider any arbitrary pixel to be visualized which has the grayscale intensity value defined as x, and the intensity value to appear on the screen to be y. Then the relationship between the input x and output intensity values can be described as: y x (2) Fig. 4 shows the effect of gamma correction for varying values of gamma. In the ideal case, input and output intensity values match perfectly as shown at the Fig. 4a. As the gamma value deviates from unity, significant differences can be noticed in the input and output intensity values. As gamma approaches zero, the output pixels become brighter, and as gamma approaches infinity, the pixels become darker. This is a very simple and straightforward algorithm to implement for detecting PQ disturbances namely, sags, swell, and flicker. For example, if one wants to detect a sag in the disturbance signal, the generated grayscale image can be subjected to gamma correction to darken up the low intensity pixels of the image. This will cause only the pixels that represent the positive peaks of the actual waveform to remain brighter, while all other pixels eventually become relatively darker. The same procedure can be applied to the complimentary image of the original image to observe pixel variation that corresponds to the negative values of the original disturbance waveform Normal image intensity values- Gamma Brighter image- Gamma = Darker image- Gamma = (a) Gamma = (b) Gamma = (c) Gamma = 2 Fig. 4 Effect of varying gamma value on the pixel intensity. B. Edge Detection The purpose of edge detection in general is to significantly reduce the amount of data in an image, while preserving the useful properties to be used for further processing. Several edge detection algorithms exist. However, in this work the edge detection technique developed by John F. Canny [2] is used for visualizing and detecting high frequency transients in the PQ disturbance signal. The Canny algorithm basically finds edges where the grayscale intensity of the image changes the most. However, to prevent noise being mistaken for edges, the image, I is first smoothed out by applying a Gaussian filter, G as: J I G (3) Gradients at each pixel in the smoothed image J are then determined by applying what is known as the Sobel-operator in the x- and y-direction respectively. The gradient and its magnitudes can be determined as: J ( J, J ) ( J / x, J / y) (4) x 2 y 2 J J x J y (5) where: J x and J y are the gradients in the x- and y-directions of the smoothed image J respectively. Although the gradient magnitudes often indicate the edges quite clearly, the edges are typically broad and thus it is difficult to determine exactly where the edges are. To make it possible to determine exact edges, the direction of the edges must be determined. The direction of edge, θ is given by the following equation: arctan( J y / J ) (6) The edge directions are used in the next step called nonmaximum suppression. The purpose of this step is to convert the blurred edges in the gradient image to sharper edges. Basically this is done by preserving all local maxima in the gradient image, and deleting everything else. The edge-pixels remaining after the non-maximum suppression step are still marked with their strength pixel-bypixel. Many of these will probably be true edges in the image, but some may be caused by noise. The simplest way to discern between these would be to use a threshold, so that only edges stronger that a certain value would be preserved. The Canny edge detection algorithm uses double thresholding. Edge pixels stronger than the high threshold are marked as strong. Edge pixels weaker than the low threshold are suppressed and edge pixels between the two thresholds are marked as weak. Further tracking is done by using hysteresis where strong edges are interpreted as certain edges, and can immediately be included in the final edge image. Weak edges are included if and only if they are connected to strong edges. The complete mathematical derivation of the Canny s edge detection algorithm can be found in [2]. x 24

4 C. Peaks and Valley Detection Grayscale images can be thought of in three dimensions where the x- and y-axes represent pixel positions and the z- axis represents the intensity of each pixel. In this interpretation, the intensity values represent elevations where regional minima and maxima appear as shown in Fig.5. Determining the regional maxima (minima) of a grayscale image is relatively easy and several algorithms have been proposed in literature [2]. One of the most efficient methods makes use of grayscale reconstruction and is based on the following proposition. The binary image J(I) of the regional maxima of I is given by: J( I) I pi ( I ) (7) where, p I is the individual pixels in grayscale image I. Fig. 6 illustrates the extraction of the regional maxima of a grayscale image I by reconstruction of I from I-. By duality, a similar technique can be derived, enabling to extract regional minima through dual grayscale reconstruction. This algorithm is used in visualization and identification of notches and short interruptions. IV. RESULTS This section describes the results obtained by applying different image processing techniques to visualize and detect the PQ disturbances. First the patterns obtained for different PQ events are illustrated. Then to evaluate the performance of the proposed method, an additional 5 recorded PQ events are assessed. It is done by comparing the output of the proposed method with the results based on the visual inspection. A. Patterns for Sag Fig. 7 illustrates the patterns obtained by applying gamma correction followed by transformations to binary images. This procedure causes the positive and negative peaks in the sag region of the original disturbance waveform to disappear in the generated binary images as shown in Fig. 7b and Fig. 7c. The dark and white objects (bars) in Fig. 7b and Fig. 7c respectively, shows non-sag positive and negative parts of the original waveform. Then by combining both positive negative half patterns, the exact sag region(s) can be identified as shown in Fig. 7d with a large dark rectangular object. Furthermore, the first cycle shown in Fig. 7e and its corresponding patterns show the compared pure wave patterns. The pure wave patterns are important in the final decision-making stage of the automatic detection. To detect the sag, the procedure first checks for objects in pure wave regions of both positive and negative half patterns. If there are objects in both those patterns, then a sag is registered whenever there are some abnormally large objects. Otherwise, it will register as a non sag event. B. Patterns for Swell A swell is defined as an increase to between. and.8 per unit in rms voltage or current at the power frequency for durations from.5 cycle to min [22]. Therefore it is possible to detect a swell by applying the same gamma correction followed by transformation to binary images as depicted in Fig. 8. This procedure causes the positive and negative peaks in the swell region of the original disturbance waveform to emerge in the generated binary images as shown in Fig. 8b and Fig. 8c, while all the other objects vanishes. The dark and white objects in Fig. 8b and Fig. 8c respectively, shows the negative and positive parts of the original waveform where swell in registered. Observe that the objects that usually emerge in pure wave patterns images died out in this case and this is the main criterion in the final decision-making part of the detection. Fig. 5 Representation of peaks and valleys in a grayscale image. 28 (d) image showing sag location Fig. 6 Extracting the regional maxima of a grayscale image I by reconstruction of I from I-. 28 Fig. 7 Patterns describing a sag in a disturbance waveform. 24

5 28 28 (d) image showing swell location Fig. 8 Patterns describing a swell in a disturbance waveform. C. Patterns for Voltage Flicker The same gamma correction and binary imaging can also be used to visualize and detect voltage flickers in the disturbance waveform. A flicker is a time varying sinusoidal voltage waveform and it is sometimes referred as a cyclical variation of the voltage envelope [22]. Fig. 9 illustrates the patterns obtained from the procedure mentioned above. From the patterns generated to represent the negative and positive part of the actual waveform as shown in Figs. 9b and 9c, it is obvious that the dark and white objects in these patterns are also undergoing a cyclic variation. Therefore, by calculating the area of each of these objects in the corresponding negative and positive part of the waveform patterns and tracking the sizes of objects, it is possible to distinguish a voltage flicker from other forms of disturbances. However, for proper detection, initially it is important to subject the disturbance waveform to a high pass filter to remove the high frequency noise. Moreover, the objects areas of compared pure wave are also again vital for proper detection. Fig. 9d shows the image obtained by combining negative and positive parts of wave patterns for proper visualization (d) image showing complete flicker pattern Fig. 9 Patterns describing a voltage flicker in a disturbance waveform. D. Patterns for Transient To visualize and detect the transients, the Canny s edged detection method is used in this paper. Fig. shows an impulsive transient disturbance detected by the Canny s algorithm. Since an impulsive transient is a sudden, nonpower frequency change in the steady state condition of voltage or current that is unidirectional in polarity, the object(s) that emerge in the binary image have very small area(s) as shown in Fig. b. Furthermore, in cases where multiple impulses exist, objects that correspond to each impulsive transient have large distance between each other. Therefore, the small size of object areas and the distance between the objects in the binary pattern are used in the automatic detection of impulsive transient. The same image processing technique is used for detecting the oscillatory transients. The only difference in this case is that the intervals between the identified edges are very close to each other as shown in Fig. b. This feature is used to differentiate between the impulsive and oscillatory transients. E. Patterns for Notches A notch is described as a switching or other disturbance in the normal power voltage waveform, lasting less than a half-cycle, which is initially of opposite polarity than the waveform and is thus subtracted from the normal waveform in terms of the peak value of the disturbance voltage [22]. According to the description above, the notching initially causes a valley in the waveform and therefore it is possible to apply the algorithm that locates the regional maxima and minima for detection of notches (b) binary pattern showing the impulsive transient (c) scaled disturbance waveform Fig. Patterns describing an impulsive transient in a disturbance waveform (b) binary pattern showing the oscillatory transient (c) scaled disturbance waveform Fig. Patterns describing an oscillatory transient in a disturbance waveform. 242

6 However, the success of notch visualization and detection again depends on eliminating the very high frequency noise such as transients. To remove this noise in the notch detection procedure, the generated grayscale image is smoothed out before introducing it to regional minima and maxima identification algorithm. Then for better detection, the objects that characterize the regional minima and maxima of pure signal must be removed from the disturbance waveform. It can be done by comparing the object areas in the reference pure wave signal pattern and the object areas in the disturbance pattern. Fig. 2 shows the regional minima and maxima corresponding to the areas of notch in the disturbance image. Note that in the pattern shown in Fig. 2b, there are always two objects that correspond to the notch in the negative part of the original disturbance waveform. Similarly, for notches in the positive part of the original disturbance signal, again two nearby objects always appear as shown in Fig. 2c. This is the unique feature that detects the notch in the disturbance waveform. F. Patterns for Short Interruptions For detecting short interruptions in the disturbance waveform, the same regional minima and maxima identification algorithm can be used. The grayscale pattern and the binary patterns for a disturbance waveform where there is short interruption are shown in Fig.3. Note that the location of the interruption area in these patterns is very obvious and it can be clearly seen even in the grayscale image. When regional minima and maxima detection algorithm is applied, there will be a large rectangular object that represents the regional minima in the interruption region as depicted in Fig. 3b. However, there will not be any object that represents the regional maxima in the interruption region as depicted in Fig. 3c. Then by combining the two patterns shown in Fig 3.b and Fig. 3c, the exact location of the interruption area can be obtained as illustrated in Fig. 3d. This algorithm is better than the gamma correction method for detecting the interruptions since it can detect even small peaks and valleys easily (b) binary pattern showing the regional minima in the notch area (c) binary pattern showing the regional maxima in the notch area (d) scaled disturbance waveform Fig. 2 Patterns describing notches in a disturbance waveform (b) binary pattern showing the regional minima (c) binary pattern showing the regional maxima (d) pattern showing both the regional minima and maxima Fig. 3 Patterns describing a short interruption in a disturbance waveform G. Assessment Studies One hundred and fifty unknown waveforms of various types of disturbances have been obtained and utilized in the assessment studies. They include field data from Power quality monitoring system in Malaysia, and simulated data generated by employing PSCAD /EMTDC software. The results are shown in Table. In this table, the first column represents the type of PQ disturbances studied. The following two columns, 2 and 3, represent the total number of cases studied of each disturbance and the number of cases that are correctly identified, respectively. Column 4 indicates the number of cases of each disturbance that are mistakenly identified. The last column shows the correct identification rate. The last row of this table represents the aggregate results. It can be seen that the new system yields a correct identification rate of 99.33%. These studies show that the proposed methods for feature extraction and decision-making are quite efficient for PQ disturbance classification. TABLE I CLASSIFICATION RESULTS YIELDED BY THE PROPOSED METHOD Type of PQ disturbances Number of disturbances Number of cases mistakenly identified Sag 3 Swell 3 Interruption Oscillatory transient Impulsive transient 5 5 Flicker Notch 3 Pure wave Correct identification rate (%) Sum: % 243

7 V. CONCLUSION This paper proposed a novel method to visualize six distinctive power quality disturbance waveforms employing image processing techniques. The pattern obtained for various disturbance signals were unique from the patterns obtained for the pure wave signal. Based on the distinctive features of the patterns, a method was designed for detecting the types of the disturbances, which has the advantage of simplicity. The results obtained from the designed algorithms are quite promising and hence will help in accurate identification of the power signal disturbances. Furthermore, this technique can be applied to other areas in power engineering such as fault location and sag source location. ACKNOWLEDGMENT This work was carried out with the financial support from the Universiti Kebangsaan Malaysia under the research grant UKM-GUP-BTT REFERENCES [] A.M. Gargoom, N. Ertugrul, and W.L.Soong, A comparative study on effective signal processing tools for power quality monitoring, th European Conf. on Power Electronics and Applications, Dresden, Germany, pp. -, -4 September 25. [2] A. Moussa, M. El-Gammal, E.N. Abdallah, and A.A. El-SLoud, Hardware - software structure for on-line power quality assessment, The 24 ASME/IEEE Joint Rail Conf., Baltimore, Maryland, USA,, pp.47-52, April 6-8, 24. [3] E.M. Tag Eldin, A new algorithm for the classification of different transient phenomena in power transformers combining wavelet transforms and fuzzy logic, The 46th IEEE International Midwest Symposium on Circuits and System, Cairo, Egypt, pp. 6 2, 27-3 December 23,. [4] H. He, and J.A. Starzyk, A self-organizing learning array system for power quality classification based on wavelet transform, IEEE Trans. Power Deliv., vol. 2, no., pp , January 26. [5] H. Kim, J. Lee, J. Choi, and G.B. Chung, Power disturbance classifier using wavelet based neural network, The 37 th IEEE Power Electronics Specialists Conf., Jeju, South Korea, pp. -6, 8-22 June 26. [6] M. Gauda, M.A. Salam, M.R. Sultam, and A.Y. Chikhani, Power quality detection and classification using wavelet multiresolution signal decomposition, IEEE Trans. Power Deliv., vol. 4, no. 4, pp , October 999. [7] H.T. Yang, and C.C Liao, A de-noising scheme for enhancing waveletbased power quality monitoring system, IEEE Trans Power Deliv., vol. 6, no. 3, pp , Juyl 2. [8] F. Zhao, and R. Yang, Power quality disturbance recognition using S- transform, IEEE Trans. Power Deliv., vol. 22, no. 2, pp , April 27. [9] M.V. Chilukuri, and P.K. Dash, Multiresolution S-transform-based fuzzy recognition system for power quality events, IEEE Trans. Power Deliv., vol. 9, no., pp , January 24. []T. Nguyen, and Y. Liao, Power quality disturbance classification utilizing S-transform and binary feature matrix method, Electr. Power Syst. Res., vol. 79, no. 4, pp , April 29. []M.F. Faisal, Identification of multiple power quality disturbances using S transform and rule base classification technique, J. Appl. Sci., vol. 9, no. 5, pp , 29. [2]S. Suja, and J. Jerome, Pattern recognition of power signal disturbances using S Transform and TT Transform, Int. J. Electr. Power Energy Syst., vol. 32, no., pp , January 2. [3]P. Janik, T. Lobos, and P. Schegner, Classification of power quality events using SVM networks, 8 th IEE International Conf. on Developments in Power System Protection, Amsterdam, Netherlands, pp , 5-8 April 24. [4]M.F. Faisal and A. Mohamed, A new technique to predict the sources of voltage sags using support vector regression based S-transform, IASTED International Conf. on Modelling, Simulation and Identification, Beijing, China, pp. -6, 2-4, October 29. [5]G. LV, X. Wang, H. Zhang, and C. Zhang, PQ disturbances identification based on SVMs classifier, International Conf. on Neural Network and Brains, Beijing, China, pp , 3-5 October 25. [6]I. Monedero, C. Leon, J. Ropero, A. Garcia, J.M. Elena, and J.C. Montano, Classification of electrical disturbances in real time using neural networks, IEEE Trans. Power Deliv., vol. 22, no. 3, pp , July 27. [7]Y. Liao, and J.B. Lee, A fuzzy-expert system for classifying power quality disturbances, Electr. Power Syst. Res., vol. 26, no. 3, pp , March 24. [8]P.K. Dash, R.K. Jena, and M.M.A. Salama, Power quality monitoring using an integrated Fourier linear combiner and fuzzy expert system, Electr. Power Syst. Res., vol. 2, no. 7, pp , October 999. [9]M.B.I. Reaz, F. Choong, M.S. Sulaiman, F. Mohd-Yasin, and M. Kamada, Expert system for power quality disturbance classifier, IEEE Trans. Power Deliv., vol. 22, no. 3, pp , July 27. [2] J. Canny, A computational approach to edge detection, IEEE Trans. on Pattern Anal. Machin. Intell., vol. pami-8, no. 6, November 986. [2]L. Vincent, Grayscale area openings and closings: their efficient implementation and applications, EURASIP Workshop on Mathematical Morphology and its Applications to Signal Processing, Barcelona, Spain, pp , May 993. [22]R.C. Dugan, M. F. McGranaghan, S. Santoso, and H.W. Beaty, Electrical Power Systems Quality, 2 nd ed., McGraw-Hill Professional,