C I R E D 18 th International Conference on Electricity Distribution Turin, 6-9 June 2005 CIRED2005 Session No 1 NEURAL NETWORKS RECOGNITION OF WEAK POINTS IN POWER SYSTEMS BASED ON WAVELET FEATURES M. Abdel-Salam 1 , Y. M.Y. Hasan 1 , M. Sayed 2 and S. Abdel-Sattar 1 1 Electrical Engineering Dept., Assiut University, Assiut, Egypt 2 Upper Egypt Electricity Production Company, Giza, Egypt 1 mazen@aun.edu.eg , ymyhasan@aun.edu.eg SUMMARY Early locating and identifying basic weak-points (sharp-edge corona, polluted-insulator "baby arcs" and loose contact arcing) in electrical power systems significantly decrease the imminent failure, outage time and supply interruption. We previously introduced a method for detecting the basic weak- points based on sound/waveform patterns and frequency analysis of their ultrasonic emissions. However, non- stationary patterns of the basic weak-points’ emitted signals and background noise frequently led to confusing discrimination. Therefore, this paper develops an effective pattern recognition scheme, employing wavelet feature extraction and Artificial Neural Network (ANN) classification, to identify the basic weak-points and two weak- point combinations (polluted insulator stressed by a transmission line with a sharp-edge and multiple sharp-edges on the same line), based on their modulated ultrasonic emissions. Extensive testing proved that the proposed scheme achieved average recognition rate of 98% when tested using weak-points underneath 33-kV and 132-kV transmission lines with 2-second detected signals. Moreover, increasing the acquisition time (>30 seconds) and classifying the weak- points based on majority voting over the ANN’s responses of multiple (15) consecutive sections, consistently led to 100% successful recognition of the considered weak-points. INTRODUCTION Electrical transmission lines and insulators are widely used in electrical power transmission and distribution networks that have been serving for many years. The insulation deteriorates under normal operating conditions, and this deterioration is accelerated due to short or long term overloads, lightning or switching surges, moisture condensation, and vibration or other climatic and mechanical stresses. These factors also may lead to other serious weak points such as poor/loose connections with subsequent arcing and micro roughness over line conductors. Hence, the basic weak points in electrical power networks are sharp-edges corona, polluted-insulators "baby arcs" and loose-contacts arcing [1]. Early locating and identifying such weak-points in electrical power systems significantly decrease the imminent failure, outage time and supply interruption that critically reflect the operability and reliability of the electrical power networks/systems [1-7]. In general, the basic weak points (sharp-edges, baby-arcs and loose-contacts) generate audio noise, radio interference complaints and/or ultrasonic noise emissions. Therefore, there are several types of sensors to detect and locate them [2-4]. The ability to locate and identify the weak points guides the maintenance staff to take a proper action such as hot washing of lines and insulators, short-circuiting the gaps by better bonding or tightening the connections, and by smoothening the coronating points to suppress corona activity. Thus, major reduction in the outage time, impending failure, equipment damage and supply interruption can be attained. Several techniques for partial discharge identification [5] using artificial neural network (ANN) , in case of various power components, have been proposed based on various feature extraction methods such a segmented time domain data compression [6] and short duration Fourier transform [7]. Alternatively, the wavelet transform (WT) [8], a mathematical tool developed in the 1980s, has been recently applied to many problems in power systems, such as analysis and visualization of electrical transients [9]. Previously, we introduced a method for detecting the three basic weak-points based on sound/waveform patterns and frequency analysis of their ultrasonic emissions [4]. However, non-stationary patterns of the basic weak-points’ emitted signals and background noise frequently led to confusing discrimination/misclassification that was a strong motivation to automate the discrimination process. Hence, this paper develops an efficient pattern recognition scheme, employing wavelet feature extraction and ANN classification, to identify not only the basic weak-points but also two weak point combinations (polluted insulator stressed by a transmission line with a sharp-edge and multiple sharp-edges on the same line), based on their modulated ultrasonic emissions. The rest of the paper is organized as follows: Background on the WT and ANN is first presented. Next, the proposed methodology is introduced. Then, details on the experimental set-up and data assembly are given. Next, the design of the ANN used is presented. The detailed results and discussion are presented and finally followed by the conclusion. BACKGROUND In pattern recognition, the extracted features should preserve as much of the original information of the signal of interest as possible while eliminating redundant and irrelevant information that could cause extraneous noise [10]. Given a zero-mean, finite energy wavelet mother function h(t), a set of functions h s,τ (t) can be generated from the single wavelet mother function by dilations and translations. The functions h s,τ (t), for all possible scales s and shifts τ, are referred to as the wavelets. The continuous