ANÁLISE DE DESEMPENHO DOS MÉTODOS DEAPRENDIZAGEM PROFUNDA PARA DETECÇÃO DEISOLADORES E SUAS ANOMALIAS

Abstract

Monitoring and performing maintenance on electrical network insulators is essential since failure in this component can cause numerous damages. However, the identification of faults in insulators is challenging since these are complex contexts. Currently, this monitoring is mostly done manually and the methods proposed to make this work automated are directed only to one country, if only one type of failure is retained in the insulators or the component is not in the real scenario. In order to provide an assertive and agile survey, contribute to a diversified database and identify a method with good performance for this objective, first a database was built with varied samples, captured in real inspection that include data from different countries, backgrounds, types of insulators and defects, to provide a robust detection with wide variability and obtain a method that reaches the most varied backgrounds and types of insulators. Second, two-stage deep learning methods that are part of the state-of-the-art were applied, which were Faster R-CNN, Dynamic R-CNN and Cascade R-CNN, both for the detection of insulators and for their defects. Finally, they were compared and Cascade R-CNN had a higher average of 𝐴𝑃 with 92.1 % for the class of defects and 83.6 % for the class of insulators and demonstrates good detection under various conditions, such as in backgrounds that resemble the object or when the object is too small in relation to the image.