Análise de coeficientes wavelet para classificação de sinais acústicos de descarga parciais

Abstract

Insulation degradation in substation electrical equipment is a problem that needs to be identified before the equipment is broken, in order to avoid causing an unexpected interruption that affects the energy supply. Another important point is to ensure the safety of employers who work in the operation, inspection and maintenance of this equipment. Partial discharges are a major cause of this failure, and there is an increasing concern to identify in a time effective manner, in order not to compromise the equipment, in addition to being of great interest the classification of the severity level of the partial discharge, allowing thus, preventive maintenance. On the other hand, the acoustic signals present in substations tend to be very polluted, since it’s a place containing numerous electrical equipment that generate acoustic noise, signals at high frequencies and magnetic fields that can affect acoustic measurements. For this reason, it is necessary to carry out signal processing to filter the important characteristics, that must be analyzed during the identification and classification of partial discharge. In this work it was proposed to use the Wavelet transform to perform the signal decomposition in several levels of approximation, detail and signal energy. With the support of the literature, the Daubechies family was identified as the most promising to work with acoustic signals. With the support of the Extra High Voltage Laboratory – Federal University of Pará, it generated a population of acoustic signals from partial discharge tests in glass insulators, which formed the database that supported the studies of this work. Using an exploratory data analysis in conjunction with principal components analysis, the results obtained from the Wavelet transform, 75% of signals were classified correctly, without using machine learning techniques, using a rule for graphical analysis, which compared the first and second principal components with each other. The possibility of creating a second rule to classify the rest of the population was also found, which could increase the effectiveness of the method, however the population decreases to a point where it affects the validation of the method, requiring more signals to reach a conclusion about its assertiveness. The coefficients obtained from the Wavelet transform can be easily modeled to work with machine learning which can improve the efficiency of the method.