Classification of Leakage Current Waveforms using Wavelet Packet Transform on High Voltage Insulator A. K. Chaou, A. Mekhaldi, B. Moula and M. Teguar Laboratoire de Recherche en Electrotechnique Ecole Nationale Polytechnique d’Alger 10 Avenue Hassen Badi, B.P 182, El-Harrach, 16200 Algiers, Algeria. khaled.chaou@g.enp.edu.dz, abdelouahab.mekhaldi@g.enp.edu.dz and madjid.teguar@g.enp.edu.dz Abstract: In this paper, the Wavelet Packet Transform (WPT) for Leakage Current (LC) examination on high voltage insulators under pollution conditions is exposed. Based on laboratory experiments under various artificial solution natures (consisting in a mixture of distilled water with NaCl, Kaolin or Kieselguhr), LC acquisition is firstly carried out. After a careful examination, three groups of LC waveforms are constituted depending on their peak values. Then, WPT is used to decompose LC waveforms. From this decomposition, feature extraction by energy calculation is processed. Hence, a feature vector, composed of wavelet coefficients energies values, is used as input for three classification algorithms consisting in K-Nearest Neighbors, Naïve Bayes and Support Vector Machine, to distinguish between three LC groups. Indeed, this paper introduces WPT for LC investigation and classification. I. INTRODUCTION Outdoor high voltage insulators, when subjected to large electrical load, suffer from the accumulation of pollution on insulators surface. In presence of wet, fog or rain, the polluted layer is transformed into a conductive one, generating the flow of a current, commonly known as Leakage Current (LC). LC has been largely used for investigation of surface electrical activity and performance monitoring of high voltage insulators [1-8]. Certain authors proposed to extract features directly on LC temporal representation to monitor polluted insulators performance. For this purpose, Li et al. [1] suggested the mean value (  ௠ ), maximum value (  ௠௔ ) and standard deviation of LC. Also, Jiang et al. [2] proposed LC peak ( ℎ ), the phase difference () between LC and applied voltage and the total harmonic distortion (THD).On the other hand, frequency representation of LC was suggested to monitor insulators performance. Suda et al. [3] concluded that magnitudes of 3rd, 5th and 7th (corresponding to 50 Hz, 150 Hz and 250 Hz respectively) harmonics increase within the flashover process. However, many authors claimed that Wavelet Analysis is very efficient to study LC waveforms under pollution [4-9]. Douar et al. [4] analyzed the flashover process on a plan insulating surface under non-uniform pollution and examined the frequency characteristics of LC. Using Standard Deviation- Multi Resolution Analysis (STD-MRA) representation of LC, authors reported good correlation between the insulator state surface and details of LC obtained through Discrete Wavelet Transform (DWT) decomposition. Recently, Moula et al. [5] used this same technique to detect the eventual presence of partial arcs activity over non-uniformly polluted glass surface. Pylarinos et al. [6-7] carried out two relevant studies investigating performance of high voltage insulators through LC waveforms classification. In the first one, authors proposed 20 features from time and frequency domains, and compared classification performance of LC waveforms using three algorithms (KNN, Naïve Bayes, Support Vector Machine). Frequency domain provides better results compared separately to time and frequency-time ones [6]. In the second study, STD- MRA based on DWT was used to extract features from LC and an Artificial Neural Network (ANN) to classify the degree of hydrophobicity loss on a polymer coated insulator surface [7].Besides, based on LC examination, Li et al. [8] claimed that the progress of the contamination discharge process can be classified into three stages, depending on LC peak value: security stage (under 50 mA), forecast stage (between 50-150 mA) and danger stage (over 150 mA). In this paper, 368 LC waveforms are firstly investigated under various pollution natures (using saline solution constituted by NaCl and distilled water sole or mixed with Kaolin or Kieselguhr), and various applied voltage levels. Then, a proposed algorithm is exposed to classify LC waveforms based on Wavelet Packet Transform (WPT). Simple classification rule, based on LC peak values, is employed to distinguish between three waveforms groups. First, signals are investigated under various pollution natures and multiple applied voltage levels. Second, WPT is used to decompose LC patterns in three levels to obtain eight coefficients. Third, feature extraction step is processed by computing energies of these coefficients. Finally, a feature vector composed of eight WPT coefficients energies values is used as an input vector by three well-known classification methods (K-Nearest Neighbors, Naïve Bayes and Support Vector Machine) to distinguish between three LC classes. This work introduces WPT, as an alternative to DWT, to study and investigate LC waveforms under pollution conditions. II. EXPERIMENTATIONS Experiments were carried out using a High Voltage Test Transformer (300 kV/50 kVA, 50 Hz) supplied by a Regulating Transformer (220/500 V, 50 kVA, 50 Hz). LC waveforms are recorded using a TEKTRONIX digital oscilloscope of 500 MHz bandwidth and are stored in a personal computer. Tests are conducted using the U120B insulator. Before tests, the insulator surface was cleaned by washing with ethyl alcohol and rinsing with distilled water, in order to remove any trace of dirt. Three pollution solutions were used in this paper. The first one consists in saline solution made up of NaCl and 1 l of distilled