The application ò fuzzy entropy to selecting features of fantial deschange high voltage cable joents
ISSN 1859-1531 - THE UNIVERSITY OF DANANG, JOURNAL OF SCIENCE AND TECHNOLOGY, NO 12(97).2015, VOL 33 THE APPLICATION OF FUZZY ENTROPY TO SELECTING FEATURES OF PARTIAL DISCHARGE IN HIGH VOLTAGE CABLE JOINTS Nguyen Tung Lam The University of Danang, University of Science and Technology; tunglam87@gmail.com Abstract - Partial discharge (PD) measurement is one of the most important diagnostics methods of insulation systems in high voltage equipment.PD activities may stem from various kinds of defects, and its characteristics correspondingly behave differently In this study, 104 features of partial discharge are collected through a series of experiments in laboratory, which are large dimensionality data set.However, not all of features are useful for classification and recognition, so the problem needed to solve is the selection of the relevant features and elimination of nonimportant features.The fuzzy entropy algorithm was applied to find out features owning characteristics for distinguishing the defects in high voltage cable joints Key words - high voltage cable joint, partial discharge, feature selection, Fuzzy entropy, recognition Partial discharge data acquisition and analysis Partial discharge could be defined as an electrical pulse or discharge in a gas filled void or on a dielectric surface of a solid or liquid insulation system This pulse or discharge only partially bridges the gap between phase insulation to the ground, and phase to-phase insulation A full discharge would be a complete fault between line potential and ground These discharges might occur in any void between the conductor and the ground The pulses occur at high frequencies; therefore, they attenuate quickly as they pass through a short distance The discharges are effectively small sparks occurring within the insulation system Therefore, it can deteriorate the insulation and can eventually result in complete insulation failure A set of PD measurement tests were carried out at the High Voltage Laboratory of National Taiwan University of Science and Technology (NTUST) based on the standard IEC60270 [3] Limiting resistor Cable HV Voltage divider capacitors Introduction Underground cables are a key link in metropolitan power grids Hence, any cable accident can lead to serious economic losses and disruption of service to customers Despite the strict quality controls for the cable production process in a plant, potential defects can occur in cable accessories during installation [8] Although the degradation mechanism and identification process of cable joints have not been fully cleared yet, it is deserving to conduct an investigation into the prevention of unexpected failure of cable systems [6] Power cable system basically consists of cables themselves and their accessories Cable accessories include joints and terminations.Statistically, the accidents caused by partial discharge mostly occur at cable joints [7] Compared to many protection methods in power system, partial discharge is considered as one of the most promising measures for monitoring and detecting possible faults in the system before they occur.One of the undoubted advantages of a computer-aided measuring system is the ability to process a large amount of information and transform this information into an understandable output [4] In this study, phase-resolved data are acquired from digital PD measurement systems during tests The phase resolved data consists of a 3D discharge pattern: phase angle – discharge magnitude - discharge rate (q-φ-t) at a specified test voltage There are many kinds of defects in cable joints and each defect own specific characteristics Different kinds of defects create different partial discharge signs and the extents of damage are not the same Based on the investigation into partial discharge from defects, the type of defects could be recognized, and from that the states of cable joints can be evaluated appropriately.In this study, 104 features of partial discharge are collected through a series of experiments in laboratory, which are large dimensionality data set However, not all of features are useful for classification and recognition, so the problem needed to solve is the selection of the relevant features and elimination of non-important features In addressing this problem, different methods of data reduction have been used and managed to eliminate the redundancy and non-important features present in the data sets Among them feature selection has been shown to be a powerful approach of dealing with high dimensional data by selecting relevant features from data set and at the same time removing irrelevant and redundant features that harm the quality of the results, and therefore builds a good learning model A good feature selection techniques will be able to detect and model the noisy and misleading features from the domain problem and help to get minimal feature subsets but still keep the important information present in the original data [5] This research proposes Fuzzy entropy method to evaluate the contribution of each feature to classification It shows that not all the features have one and the same discriminatory power [1] As a result, the crucial features are identified by using fuzzy entropy V Voltage signals Attenuator MD Circuit protector Partial discharge signals Figure Experimental setup for PD measurement 34 Nguyen Tung Lam 2.1 Partial Discharge Data Acquisition PD measurements were generated and recorded from laboratory tests During the experimental process, all of the measuring analog data was converted to digital data in order to be stored in computer After that process, these data was transformed into q-φ-t format (discharge evaluation pattern) This data format is called as Phase resolved partial discharge data Figure illustrates general PD data acquisition scheme The most basic quantities of PD activities are apparent charge, q, apparent charge number n, and phase position of PD pulses with respect to the applied test voltage, φ, interpretation purposes of the under test insulation system by a defect present in the insulation on a 3D phase-resolved pattern, representing a one second (40 cycles) snapshot of PD activity This is achieved by plotting each pulse, or in the case of the IEC data the peak amplitude of the apparent charge, on a three-dimensional axis consisting of the pulse’s relative amplitude, the cycle number on which the pulse appears and the phase position of the pulse on the voltage cycle An example of a phase-resolved pattern, which represents three kinds of defect PD activities, can be seen in Figure The pattern is in the form of a 40x600 matrix of floating points that represent the PD activity in 40 consecutive cycles across 600 phase windows of the voltage cycle; with the positive half cycle appearing first, between 0° and 180° and then the negative half cycle between 180° and 360° Figure A general PD data acquisition scheme However, the above mentioned quantities cannot be sufficient for a perfect diagnostics So, heuristically there have been introduced lots of features derived from basic quantities termed as deduced and statistical operators, which can be used for defect identification and evaluation Therefore acquiring PD data and extracting statistical feature from acquired data benefit us for reliable PD monitoring - Basic quantities, which are quantities observed during one voltage cycle - Deduced quantities, which are integrated values of basic quantities from the first group observed throughout several voltage cycles - Statistical operators, which are operators for the statistical analysis of the deduced parameters This process data analysis can be a good indicator for ambiguous PD patterns to diagnose as it presents distinctive features of each PD defect pattern which has been accumulated for a longer time than PD real-time data For the convenience of statistical analysis, the 3D patterns were decomposed into two 2D distributions by projecting it into the two axes - phase and magnitude Statistical analysis is performed separately for those two distributions Also, statistical analysis is performed separately for phase angles from to 180° (“positive” PDs), for phase angles from 180° to 360° (“negative” PDs), and on the difference between positive and negative PDs For each of the distributions, two types of statistics, names amplitude statistics and shape statistics, are calculated The statistical descriptors are mean, standard deviation, skewness and kurtosis In addition, overall maximum magnitudes of positive and negative PDs and discharge phase region PD patterns are also calculated as features To diagnose a fault from the PD data it first needs to be transformed into a generic workable format One way of displaying the data is to plot consecutive pulses generated Figure Partial discharge signal measurement Figure 3D q-φ-t transformation Figure Defect type A 3-D q-φ-t pattern Using data in this form, the knowledge-based system offers an automated approach to defect classification and offers an explanation of the reasons for its conclusion This ability offers a physical explanation for the automatic classification sets.For further statistical analysis, the 3-D patterns are decomposed into two 2-D distributions by projecting it into the two axes - phase and magnitude Figure shows 3D q-φ-t pattern decomposition that is ISSN 1859-1531 - THE UNIVERSITY OF DANANG, JOURNAL OF SCIENCE AND TECHNOLOGY, NO 12(97).2015, VOL transformed into 2-D distribution pattern that includes two axes - phase and magnitude Figure 3-D q-φ-t decomposition Having basic PD quantities at hand, through statistical operators, 104 statistical features (positive and negative) were extracted from the four major PD quantities related to phase and height distributions These 104 statistical features are also called PD fingerprints The PD-fingerprint in this work is a histogram combination of statistical features of a PD signal The shape of the histogram provides information about the nature of the PD signal [9] The features of a histogram are statistical characteristics, where the histogram is used as a model of the probability distribution of a pattern These statistical features provide us with the characteristics of a PD pattern Statistical methods for extracting PD features are based on phase resolved PD patterns By applying statistical computation on PD patterns, different distributions can be characterized as statistical features The mean value, standard deviation, skewness and kurtois values are calculated according to statistical formulas Discharge phase region is also calculated 2.2 Experimental setup In this research, the experimental objects are cable joint defect models Two types of relevant models are well designed, based on investigations of numerous power equipment failures Two defect types are described, as follows: Defect type A: remove a part of insulation According to the criterion, the length of insulation of two cable sides is the same and complies with standard In this type, a part of one insulation side was cut out Defect type B: gap between insulation, a hole made in insulation belong to the part of cable inside the joint Feature selection based on fuzzy entropy Feature selection is a process of choosing small subset of features out of a set of candidate features based on certain criteria Feature selection plays an important role in classification for several reasons First it can simplify the model and in this way computational cost can be reduced and also when the model is taken for practical use fewer inputs are needed which means in practice that fewer 35 measurements from new samples are needed Second by removing redundant features from the data set one can also make the model more transparent and more comprehensible, providing better explanation of suggested diagnosis Feature selection process can also reduce noise and in this way enhance the classification accuracy The key of PD classification problem like in any classification systems is a set of high quality features These features should represent the characteristics of PD signals More importantly, these features must possess strong discriminant power so that the classifier designed based on those features can give desired performance Since PD is a stochastic process, namely, the occurrence of PD depends on many factors, such as temperature, pressure, applied voltage, and the test duration, and since PD signals contain noise and interference, PD measurements corresponding to different insulation conditions are almost indistinguishable, i.e., PD diagnosis is a complex classification problem Thus finding a set of high quality features that give more accurate and reliable classification is even more critical in design of PD diagnostic systems In this paper, Fuzzy entropy method is proposed to reduce dimension of partial discharge features For a classification system, the most important procedure is partitioning the pattern space into decision regions Once the decision regions are decided, we can apply these partitioned decision regions to classify the unknown patterns The partition of decision regions is part of the learning procedure or training procedure since the decision regions are decided by the training patterns In fuzzy entropy classifier, decision regions are enclosed by the surfaces produced from each dimension The surfaces are determined by the distribution of input patterns Entropy is a measure of the amount of uncertainty in the outcome of a random experiment, or equivalently, a measure of the information obtained when the outcome is observed [2] In this paper, a fuzzy entropy measure which is an extension of Shannon’s definition will be proposed The fuzzy entropy can discriminate the actual distribution of patterns better By employing membership functions for measuring match degrees, the value of entropy not only considers the number of patterns but also takes the actual distribution of patterns into account The fuzzy entropy reflects more information in the actual distribution of patterns in the pattern space Since the fuzzy entropy can discriminate pattern distribution better, we employ it to evaluate the separability of each feature Intuitively, the lower the fuzzy entropy of a feature is, the higher the feature’s discriminating ability is The procedure for computing the fuzzy entropy of each feature is described as the flowchart in Figure This process includes four main parts: - Determine the number of intervals - Determine the interval locations - Assign a membership function for each interval - Compute the fuzzy entropy of each feature via summation of the fuzzy entropy of all intervals At first, assume the number of interval I equal to which 36 Nguyen Tung Lam is the smallest number of interval Then increase I until the total fuzzy entropy of I intervals is less than that of I - intervals The final fuzzy entropy is computed with I-1 intervals Set initial number of interval I=2 Set initial centers of intervals c Assign interval label to each element Yes Recompute the cluster centers Check: Does any center change? Table Values of Fuzzy Entropy A1, B1 No Feature Value Value 83 0.69525 30 0.72506 29 0.72590 0.75572 14 45 Value 0.67481 0.75072 0.74844 35 0.89264 104 0.75031 11 0.94105 14 0.78717 46 0.95920 0.75999 13 0.79021 47 0.96313 0.76027 45 0.79145 45 0.96674 82 0.76313 63 0.81147 38 0.98958 27 0.76987 23 0.81163 0.99492 16 0.80721 46 0.81554 33 0.99593 46 0.81381 11 0.82831 34 0.99628 A2, B1 Assign membership function for each interval I>2 the total fuzzy entropy of I intervals is less than that of I intervals I=2 I=I+1 True The number of interval I=I-1 & The fuzzy entropy is computed with I-1 intervals False Figure Flowchart of calculating value Fuzzy entropy Results As mentioned in the previous chapter, each defect was tested on cable joints Table Set of PD data’s class PD Defect Types Remove a part of insulation Gap between insulations Class A (A1, A2, A3) B (B1, B2, B3) After finishing all the tests, partial discharge data of defects was collected In this study, data obtained in the results of experimental works was considered and transformed into statistical features and these considerations are explained in PD Data Acquisition and Analysis section These statistical features including skewness, kurtosis, standard deviation, mean, DPR, 〖 Q〗_sum,〖 Q〗_num,〖 Q〗_max,〖 Q〗_ave… are all calculated based on the PD signals Statistical features consist of 104 features numbered from to 104 In this study, 34 kV PD experiment data that include 120 sample data for each defect model is used Apply Fuzzy entropy theory to all features of partial discharge, with inputs to Matlab program as values of all features and types of defects corresponding As a result, the fuzzy entropy value of each feature is computed; features with higher fuzzy entropy are less relevant to classification goal Totally pairs of defects (each pair includes defect type A and defect type B) were conducted to calculate fuzzy entropy values Table shows the features owning smallest values of fuzzy entropy of each case A1, B3 No Feature No Centers of intervals are determined Compute the total fuzzy entropy of all intervals A1, B2 No Feature A2, B2 No Feature Value No Feature Value 83 0.43104 97 46 0.48254 98 0.55353 0.55498 A2, B3 No Feature Value 0.78922 0.91810 0.79779 46 0.93049 92 0.95837 45 0.93576 21 0.96971 0.94505 0.55773 22 0.97513 0.94838 88 0.55843 45 0.97986 10 0.95489 45 0.56206 46 0.98135 18 0.96062 61 0.56995 101 0.98343 17 0.96299 60 0.56995 0.98379 15 0.96787 58 0.57423 19 0.98403 41 0.97356 A3, B1 No Feature A3, B2 Value No Feature 83 0.69525 30 29 A3, B3 Value No Feature Value 47 0.15372 35 0.14152 0.72506 50 0.20041 46 0.14490 0.72590 46 0.21575 31 0.14748 0.75572 40 0.22609 45 0.14959 14 0.75999 0.22911 53 0.15461 45 0.76027 44 0.24992 0.15624 82 0.76313 42 0.25441 10 0.16715 27 0.76987 45 0.27248 0.17647 16 0.80721 61 0.27675 38 0.19347 46 0.81381 93 0.28323 34 0.20328 As a result, it can be clearly seen that the feature number 2, 45 and 46 are always in the top of features having the smallest values of fuzzy entropy As mentioned in the theory, those features impact significantly on classifying defects in cable joints We can use three features instead of all 104 features to recognize not only more accuracy but also less time of computing Table Selected features No 45 Features Total number of partial discharge in all circles Height distribution average partial discharge ISSN 1859-1531 - THE UNIVERSITY OF DANANG, JOURNAL OF SCIENCE AND TECHNOLOGY, NO 12(97).2015, VOL 46 magnitude – kurtois values Height distribution average partial discharge maximum– kurtois values Conclusion Based on the preprocessing stage, it is necessary to gather the database from conducted PD tests with two high voltage cable joints including prefabricated defects Phaseresolved PD data was successfully evaluated and processed The data gathered from the selected databases is connected to MATLAB software where the data is processed The second part of the analysis system is based on feature selection algorithm This takes Fuzzy entropy method as a medium to process the PD data Feature selection techniques were applied to the PD data and the characteristic points of interests are being selected by computing fuzzy entropy value of each defect The data provides meaningful information for the classification of PD defects The measure of uncertainty is adopted as a measure of information Hence, the measures of fuzziness are known as fuzzy information measures The measure of a quantity of fuzzy information gained from a fuzzy set or fuzzy system is known as fuzzy entropy In this study, the fuzzy entropy algorithm was applied to find out three features owning most useful characteristic for distinguishing the defects in high voltage cable joinst 37 REFERENCES [1] G MacLachlan, "Discriminant Analysis and Statistical Pattern Recognition", Willey-Interscience, pp 389-398, 2004 [2] Hahn-Ming Lee,Chih-Ming Chen, Jyh-Ming Chen, Yu-Lu Jou, "An Efficient Fuzzy Classifier with Feature Selection Based on Fuzzy Entropy", IEEE Transactions On Systems, Man, And Cybernetics— Part B: Cybernetics, Vol 31, No 3, pp 426-432, 2001 [3] IEC 60270, "High-voltage test technique-Partial discharge measurements", 2000 [4] N Sahoo, M Salama and R Bartnikas, "Trends in partial discharge pattern classification: a survey", IEEE Transactions on Dielectrics and Electrical Insulation, Vol 12, No 2,pp 248 - 264, April 2005 [5] Q S Jensen Richard, "Computational intelligence and feature selection: Rough and Fuzzy Approaches", IEEE Press Series on Computational Intelligence, 2008 [6] Tokunaga S, Tsurusaki T, "Partial Discharge Characteristics till Breakdown for XLPE Cable Joint with an Artificial Defect", Proceedings of the 7th International Conference on Properties and Applications of Dielectric Materials, Vol.3, pp 1206-1209, Nagoya, 2003 [7] Wenhu Yang,Yanqun Liao, Yang Xu, Xiaolong Cao "Analysis of Partial Discharge Measured on Field for Cable Joint", International Conference on High Voltage Engineering and Application, pp 408411, Chongqing, China, November 9-13, 2008 [8] Wu Ruay-Nan, Chang,Chien-Kuo, "The Use of Partial Discharges as an Online Monitoring System for Underground Cable Joints", IEEE Transactions On Power Delivery, vol 26, pp 1585-1591, 2011 [9] Yu-Hsun Lin, Ruay-Nan Wu, I-Hua Chung, "Novel trend of "l" shape in PD pattern to judge the appropriate crucial moment of replacing cast-resin current transformer", IEEE Transactions on Dielectrics and Electrical Insulation, vol 15, no 1, pp 292-301, 2008 (The Board of Editors received the paper on 07/10/2015, its review was completed on 09/27/2015) ... fingerprints The PD-fingerprint in this work is a histogram combination of statistical features of a PD signal The shape of the histogram provides information about the nature of the PD signal [9] The features. .. insulation belong to the part of cable inside the joint Feature selection based on fuzzy entropy Feature selection is a process of choosing small subset of features out of a set of candidate features. .. the number of interval I equal to which 36 Nguyen Tung Lam is the smallest number of interval Then increase I until the total fuzzy entropy of I intervals is less than that of I - intervals The