Definição dos principais atributos para identificação do estado operacional de isoladores poliméricos por aprendizado de máquina.

Abstract

Polymer insulators have gradually replaced ceramic insulators in transmission lines in the last decades. However, even though polymer insulators have advantages over ceramic ones, their diagnosis is more complex and affected to a greater or lesser degree by the subjectivity of the evaluator. The use of machine learning methods can mitigate this problem. However, specialized literature lacks on the definition of the most significant input features for these methods, as well as on representative databases. Another problem commonly observed is the high subjectivity during the initial classification (labeling) of insulators. In this context, the present work aims to determine the most significant features arising from the following inspection techniques: infrared thermography, detection of ultraviolet radiation and ultrasonic noise, while proposing a methodology for reducing subjectivity in the classification of insulators. It was also intended to provide a vast and representative database. Therefore, 60 polymer insulators removed from a transmission line were tested in laboratory to extract data from each of the aforementioned techniques. The data were processed and a methodology based on fuzzy clustering was proposed to reduce subjectivity in the labeling of insulators, considering the visual inspection in a conservative conception. Three feature selection methods were used to determine the most significant attributes, namely: minimum Redundancy, Maximum Relevance, Relief F and chi-square tests. From the results obtained, it was found that features from ultrasonic noise, more precisely from the wavelet transform of ultrasonic noise signals, were promising for the diagnosis of polymer insulators, as well as features related to the location of hot spots in the insulator body. Thus, the main attributes listed are the maximum locations for the temperature increase, as well as the temperature variations and variations of the variations; the asymmetry of the detail coefficients of the second level of the Wavelet transform and the first harmonic for the ultrasonic noise. However, the results showed that visual inspection is still an important criterion for classifying insulators. Results also indicate that the proposed methodology is not affected by the environmental characteristics in the tests, but more studies are needed. New works may propose other attributes, especially related to inspection by detection of ultraviolet radiation, as well as new methods of attribute selection.