PATTERN RECOGNITION FOR PARTIAL DISCHARGE DIAGNOSIS OF POWER TRANSFORMER

Transcription

1 PATTERN RECOGNITION FOR PARTIAL DISCHARGE DIAGNOSIS OF POWER TRANSFORMER PO-HUNG CHEN 1, HUNG-CHENG CHEN 2, AN LIU 3, LI-MING CHEN 1 1 Department of Electrical Engineering, St. John s University, Taipei, Taiwan, 25135, R.O.C. 2 Department of Electrical Eng., National Chin-Yi University of Science and Technology, Taichung, Taiwan, 411, R.O.C. 3 Department of Computer Science and Information Engineering, St. John s University, Taipei, Taiwan, 25135, R.O.C. phchen@mail.sju.edu.tw Abstract: This paper presents a novel pattern recognition approach based on a four-layer artificial neural network (ANN) for the partial discharge (PD) diagnosis of power transformer. A precious PD detector is used to measure 3-D (φ-q-n) signals and PD-fingerprints of four experimental models in a shielded laboratory. This work has established a database containing 160 sets of 3-D patterns and PD-fingerprints. The database is used as the training data to train a back-propagation neural network. The training-accomplished neural network can be a good diagnosis system for the practical insulation diagnosis of epoxy-resin power transformers. The proposed approach is successfully applied to practical power transformers field experiments. Experimental results demonstrated that the proposed pattern recognition approach has high recognition rate. Keywords: Pattern recognition; Artificial neural network; Partial discharge; PD-fingerprints; Power transformer 1. Introduction Breakdown of a power apparatus such as a power transformer can cause an interruption in electricity supply and result in a loss of considerable profits [1]. Therefore, detecting insulation defects in a power apparatus as early as possible is of priority concern to a power apparatus user. Partial discharge (PD) pattern recognition is an important tool for power apparatuses insulation diagnosis. More than half of the breakdown accidents of power apparatuses are caused by insulation deterioration. The reliability of a power apparatus is affected significantly by the presence of insulation defects. PD pattern recognition has been regarded as an important diagnosis method to prevent power apparatuses from malfunction of insulation defect. PD phenomenon usually originates from insulation defects and is an important symptom to detect incipient failures in power transformers. Classification of different types of PD patterns is of importance for the diagnosis of the quality of power transformers. PD behavior can be represented in various ways. Because of the randomization of PD activity, one of the most popular representations is the 3-D (φ-q-n) distribution [2], [3], i.e., the PD pattern is described using a pulse count N versus pulse height Q and phase angle φ diagram. Previous experimental results have adequately demonstrated that φ-q-n distributions are strongly dependent upon the PD source. Therefore, the 3-D patterns can be used to characterize insulation defects [4]. This provides the basis for pattern recognition techniques that can identify the different types of defects. In recent years, a biologically artificial intelligence technique known as artificial neural network (ANN) has emerged as a candidate for the feature identification problems. The basic conception of ANN is intended to model the behavior of biological neural functions. An ANN is generally modeled as a massively parallel interconnected network of elementary neurons. The original desire for the development of ANN is intended to take advantage of parallel processors computing than traditional serial computation. In our previous efforts, an ANN-based pattern recognition approach has been proposed [5]. Various pattern recognition techniques such as fuzzy clustering [6], expert system [7], extension theory [8], [9], and PD-fingerprints [10], [11] have been proposed. These techniques have been successfully applied to PD pattern recognition. However, these conventional approaches not only require human experiences but also have some difficulties in acquiring knowledge and maintaining the database of decision rules. The ANN can directly acquire experience from the training data to overcome the shortcomings of the conventional approaches. The ANN can quickly and stably learn to categorize input patterns and permit adaptive processes to access significant new information [12]. This work has established a database which contains /10/$ IEEE 2996

2 sets of 3-D patterns and PD-fingerprints. Then, this database is used as the training data to train an ANN. Finally, the training-accomplished ANN can be a good diagnosis system for the practical insulation diagnosis of epoxy-resin power transformers. Experimental results show that different types of insulation defects within power transformers are identified with rather high recognition rate by the proposed recognition approach. 2. Practical PD field measurement PD is a forced phenomenon occurring in insulation parts in a power apparatus. When the intensity of electric field exceeds the breakdown threshold value of a defective dielectric, PD occurs and results in a partial breakdown in the surrounding dielectrics. PD is a symptom of insulation deterioration. Therefore, PD measurement and identification can be used as a good insulation diagnosis tool to optimize both maintenance and life-risk management for power apparatuses. The new standard IEC60270 [13] for PD measurement has been published in 2001, which establishes an integral quality assurance system for PD measurement instead of the old standard IEC published in The standard IEC60270 ensures accuracy of measuring results, comparability and consistency of different instruments and measuring methods. Moreover, the new standard provides digital PD measuring recommendations as well as the analog measuring. In this work, all PD experiments are based on the new standard IEC A PD experiment laboratory, including a set of precious instrument (Hipotronics DDX-7000 Digital Discharge Detector), has been set up in the St. John s University according to the standard IEC60270 recommendations. The constitution of the PD experiment laboratory includes a PD analyzer, a high-voltage control panel, an isolation transformer, a high-voltage generator (step-up transformer), a calibration capacitor, and a coupling capacitor. Figure 1 shows the practical PD field measurement in the shielded room. The experimental objects are epoxy-resin power transformers in this work. The rated voltage and capacity of the transformers are 12 kv and 2kVA, respectively. For testing purposes, four experimental models of power transformers with typical insulation defects were purposely manufactured by a power apparatuses manufacturer. These typical PD models include a low-voltage coil PD (Type A), a high-voltage coil PD (Type B), a high-voltage corona discharge (Type C), and a healthy transformer (Type D). The voltage step-up procedure of the PD experiment, according to the standard IEC60270 recommendations, is shown in Figure 2. First, the high-voltage generator generates a rising-voltage from 0V to 18kV, which is the 1.5 times of the rated voltage, in 50 seconds. This high-voltage will be maintained for 1 minute to trigger discharges. Then, the voltage will descend from 18kV to the rated voltage in 20 seconds. The 12kV rated voltage will be kept and the PD detector starts to measure and record the PD signals for 2 minutes. During the experimental process, all of the measuring analog data are converted to digital data in order to store them in a computer. Test Voltage (kv) Figure 1. Practical PD field measurement 18kV 12kV Test Time (second) Figure 2. Voltage step-up procedure 3. 3-D patterns and PD-fingerprints D patterns In this paper, an ANN-based pattern recognition approach using 3-D (φ-q-n) patterns and PD-fingerprints is proposed for epoxy-resin power transformers insulation diagnosis. Four cast-resin transformers with typical insulation defects, which were purposely made by a manufacturer, are used as the models of the PD examination. The DDX-7000 Digital Discharge Detector is then used to measure the 3-D (φ-q-n) signals and PD-fingerprints of these transformers. Figure 3 shows the 3-D (φ-q-n) patterns of four typical defects. Obviously, the problem of transformer insulation diagnosis is essentially a PD pattern classification problem. 2997

3 Therefore, the ANN is an applicable solution tool to the problem of transformer insulation diagnosis PD-fingerprints (a) Low-voltage coil PD (Type A) (b) High-voltage coil PD (Type B) (c) High-voltage corona discharge (Type C) (d) Healthy transformer (Type D) Figure 3. 3-D (φ-q-n) patterns of typical defects The PD-fingerprint in this work is a histogram combination of statistical parameters of a PD signal. The shape of the histogram provides information about the nature of the PD signal. 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. In this work, these statistical parameters including skewness, kurtosis, cross correlation, peaks, asymmetry, and phase factor are all calculated based on the PD signals. The skewness, kurtosis, and cross correlation are defined as follows [10], [11]. A. Skewness (SK) In statistics, skewness is a measure of the degree of asymmetry of a distribution. Positive skewness indicates the distribution is skewed to the left, with a longer tail to the right of the distribution maximum. Negative skewness indicates the distribution is skewed to the right, with a longer tail to the left of the distribution maximum. The skewness shows how symmetry the distribution. It measures the asymmetry about the mean in a PD signal. It is defined as N 3 xi µ SK( x1,..., xn ) = Pi (1) i=1 σ where x i is the value of the i-th variable, P i is the probability of appearance for the i-th value, µ is the mean value,σ is the variance, and N is the number of sampling points of a PD signal. B. Kurtosis (KU) In statistics, kurtosis is a measure of whether the value of a variable are peaked or flat relative to a normal distribution. Higher kurtosis means more of the variance is the result of infrequent extreme deviations, as opposed to frequent modestly sized deviations. It is defined as N 4 xi µ KU ( x1,..., xn ) = Pi 3 (2) σ i= 1 C. Cross Correlation (r) In pattern recognition, cross correlation is a measure of similarity of two waveforms as a function of a time-lag applied to one of them. This is also known as a sliding dot product or inner-product. It is commonly used to search a long duration signal for a shorter known feature. The PD signal is divided into positive and negative half cycles in according to phase. The cross correlation describes the difference in distribution shape between these two cycles. It is defined as r = M p M M a p M M a p M a x y x y i= 1 j= 1 i= 1 j= 1 i= 1 j= 1 p a 2 M p M a M p M a M M M M 2 x x / i= 1 j= 1 i= 1 j= 1 i= 1 j= 1 i= 1 j= 1 ( M M ) / (3) 2 p a p a 2 ( M pm a ) y y /( M pm a ) 2998

4 where x is the value at -th position of the positive half cycle and y is the value in the corresponding negative half cycle. M p is the number of pixel of each half cycle at x axes and M a is the number of pixel at y axes. The coefficient ranges from -1 to 1. The value of r represents the degree of symmetry between positive and negative half cycle of a distribution. (a) Low-voltage coil PD (Type A) (b) High-voltage coil PD (Type B) (c) High-voltage corona discharge (Type C) (d) Healthy transformer (Type D) Figure 4. PD-fingerprints of typical defects According to these definition, the average values of the statistical parameters corresponding to the four types of field-test PD patterns are calculated and shown in Figure 4. Starting with PD patterns on different types of specimens, a suitable set of statistical parameters are determined and then used as input variables to an ANN for the purpose of classifying the insulation defects. From the literature survey, several models and learning algorithms of ANN have been proposed for solving the patterns classification problems [14]. In this paper, we establish a four-layer feed-forward ANN for solving the PD patterns classification problem. The number of neurons in the output layer is set at the number of defect types. In this work, the transfer function in the hidden layer and output layer is the hyperbolic tangent function and the sigmoid function, respectively [14] Learning algorithm In this paper, a faster back-propagation learning algorithm named RPROP algorithm is used as the learning rule. Riedmiller and Braun [15] showed that both convergence speed and memory requirement of the RPROP algorithm are better than traditional gradient-descent learning algorithms. In the RPROP algorithm, the update-values for each weight are modified according to the behavior of the sequence of signs of the partial derivatives in each dimension of the weight space, not according to the gradient value. The modified procedure of a weight of the RPROP algorithm can be mathematically formulated as follows: = + η η ( t 1) ( t 1),, ( t 1), if if W W ( t 1) > 0 W ( t 1) < 0 W else, if ( w = +, if ( t w 0, else t) > 0 (4) W ) < 0 (5) W ( t + 1) = W + W (6) where W () t : weight back propagated from neuron j to neuron i η + /η : learning velocity, where 0<η <1<η + Et (): error function 2999

5 1 1-1 Proceedings of the Ninth International Conference on Machine Learning and Cybernetics, Qingdao, July 2010 () t : update value of W () t 4. Experimental results The proposed ANN-based pattern recognition approach was implemented on a MATLAB software and executed on a Pentium IV personal computer. 160 sets of field measuring 3-D patterns and PD-fingerprints are used to test the recognition system. The recognition system will randomly choose 80 sets of data as the training data, and the rest as the testing data. Two example cases are studied to illustrate the identification ability of the proposed approach. Throughout the study, an existing method [5] (Example 1) is used as the main benchmark of comparison for the proposed approach (Example 2) Example 1: existing method Our previous effort [5] proposed a triple-layer feed-forward ANN recognition system as shown in Figure 5. The input data for the recognition system is the field measuring 3-D patterns. Table 1 shows the test results of the existing method with different number of neurons in the hidden layer from H=20 to H=100. Test results show that H=60 has the highest recognition rate of 90%, 90%, 93%, and 80% for defect type A, B, C, and D, respectively. The existing method achieves a average recognition rate of 88.25%. 3-D PD Pattern Translate to 1600x1 Matrix Recognition Result A B C 80 88% 80% 85% 77% 82.5% % 80% 77% 73% 79.5% 4.2. Example 2: proposed approach Figure 6 shows the topology structure of the proposed novel recognition system which is a four-layer feed-forward ANN. The input data for the novel recognition system is the field measuring 3-D (φ-q-n) patterns and PD-fingerprints. The number of neurons in the output layer (O) depends on the defect types to be identified, which is O=4 in this work. To fit the form of the input layer and accelerate convergence, each original 3-D pattern is pre-translated into a 1600x1 matrix. Therefore, the number of neurons in the input layer for the 3-D pattern (I) is set at I=1600. Since each PD-fingerprint is a histogram combination of 29 statistical parameters of a PD signal in this work, the number of neurons in the input layer for the PD-fingerprint is set at 29. Then, for a multi-layer ANN, the main control parameter is the number of neurons in the hidden layers (H 1, H 2 ). Table 2 shows the test results of the novel recognition system with different number of neurons in the two hidden layers from H 1 =60, H 2 =40 to H 1 =80, H 2 =60. Test results show that H 1 =80, H 2 =40 has the highest recognition rate of 95%, 98%, 100%, and 93% for defect type A, B, C, and D, respectively. The proposed novel recognition system achieves a high average recognition rate of 96.5%. From the two example cases, notably, the proposed novel recognition system has better average recognition rate than the existing methods. 3-D (φ-q-n) patterns D I= Input Layer H Hidden Layer O=4 Output Layer PDfingerprints Figure 5. Structure of existing method [5] TABLE 1. RECOGNITION RATE OF EXISTING METHOD [5] Neurons in Defect Type hidden layer (H) A B C D Average 20 80% 88% 85% 80% 83.25% 40 88% 88% 90% 80% 86.5% 60 90% 90% 93% 80% 88.25% Figure 6. Structure of proposed novel approach TABLE 2. RECOGNITION RATE OF PROPOSED NOVEL APPROACH Neurons in Defect Type hidden layers (H 1, H 2 ) A B C D Average 3000

6 H 1 =60, H 2 =40 88% 90% 93% 90% 90.25% H 1 =60, H 2 =60 85% 85% 93% 90% 88.25% H 1 =80, H 2 =40 95% 98% 100% 93% 96.5% H 1 =80, H 2 =60 93% 98% 100% 90% 95.25% 5. Conclusions This paper presents a new pattern recognition methodology based on a four-layer feed-forward ANN for solving the insulation diagnosis problem of power transformers. The input data for the recognition system is the field measuring 3-D (φ-q-n) patterns and PD-fingerprints. The proposed approach is considered a general tool because it can be easy implemented on the popular MATLAB software. Moreover, the approach is considered a flexible tool because it can be also applied to other high-voltage power apparatuses such as current transformer, potential transformer, cable, and rotation machine. Experimental results demonstrated that the proposed novel recognition system has better average recognition rate than the existing methods. Acknowledgements Financial support to this research by the National Science Council of the Republic of China, Taiwan, under grant No. NSC E is greatly appreciated. References [1] R. Feinberg, Modern Power Transformer Practice, John Wiley & Sons, [2] E. Gulski, H. P. Burger, G. H. Vaillancourt, and R. Brooks, PD Pattern Analysis During Induced Test of Large Power Transformers, IEEE Transactions on Dielectrics and Electrical Insulation, Vol. 7, pp , [3] C. S. Kim, T. Kondo, and T. Mizutani, Change in PD Pattern with Aging, IEEE Transactions on Dielectrics and Electrical Insulation, Vol. 11, pp , [4] A. Krivda, Automated Recognition of Partial Discharge, IEEE Transactions on Dielectrics and Electrical Insulation, Vol. 2, pp , [5] P. H. Chen, and H. C. Chen, Application of Back-Propagation Neural Network to Power Transformer Insulation Diagnosis, Lecture Notes in Computer Science, Vol. 4493, pp , June [6] K. Tomsovic, M. Tapper, and T. T. Ingvarsson, A Fuzzy Information Approach to Integrating Different Transformer Diagnostic Methods, IEEE Transactions on Power Delivery, Vol. 8, pp , [7] L. Satish, and B. I. Gururaj, Application of Expert System to Partial Discharge Pattern Recognition, Proceeding of CIGRE Study Committee 33 Colloquium, Leningrad, Russia, [8] M. H. Wang, and C. Y. Ho, Application of Extension Theory to PD Pattern Recognition in High-Voltage Current Transformers, IEEE Transactions on Power Delivery, Vol. 20, pp , [9] M. H. Wang, Partial Discharge Pattern Recognition of Current Transformers Using an ENN, IEEE Transactions on Power Delivery, Vol. 20, pp , [10] A. Cavallini, G. C. Montanari, F. Puletti, and A. Contin, A New Methodology for the Identification of PD in Electrical Apparatus: Properties and Applications, IEEE Transactions on Dielectrics and Electrical Insulation, Vol. 12, pp , [11] J. Fuhr, M. Haessig, B. Fruth, and T. Kaiser, PD-fingerprints of some high voltage apparatus, Proceeding of IEEE International Symposium on Electrical Insulation, pp , [12] M. M. A. Salama, and R. Bartnikas, Determination of Neural Network Topology for Partial Discharge Pulse Pattern Recognition, IEEE Transactions on Neural Networks, Vol. 13, pp , [13] IEC, High-Voltage Test Techniques-Partial Discharge Measurements, IEC 60270, [14] P. D. Wasserman, Neural Computing, Theory and Practice, Van Nostrand Reinhold, [15] M. Riedmiller, and H. Braun, A Direct Adaptive Method for Faster Back Propagation Learning-The RPROP Algorithm, IEEE International Conference on Neural Networks, Vol. 1, pp ,