Aplicação de técnicas de visão computacional no reconhecimento e na detecção de defeito em isoladores poliméricos de suspensão.

Abstract

Electrical insulators play a significant role in the reliability of the power system. In recent decades, glass and porcelain insulators have been replaced by polymeric insulators. Although they have advantages over glass and porcelain insulators, polymeric insulators are more susceptible to defects due to exposure to solar radiation and electrical stress. Additionally, the diagnosis of polymeric insulators is more complex, as the recurrent defects are quite small and more difficult to detect during the inspection process. To make the inspection process faster, many researchers are proposing the use of Computer Vision techniques to detect defects in glass and porcelain insulator strings from aerial images. However, few works have used the same methodology for to inspect polymeric insulators. Therefore, the present work aims to identify defects in polymeric suspension insulators using Artificial Intelligence techniques applied to images. To this end, datasets of polymeric insulator images at five different distances between the camera and the insulator are constructed. Due to the need for a large amount of data to train Computer Vision models, data augmentation techniques are used to obtain new samples of images. Next, fine-tuning models of YOLOv8 detectors are trained to detect four types of defects in insulators: exposed rod, end fitting corrosion, cracks, and damaged sheds. Finally, the impact of the distance between the camera and the insulator, the input size of the images, and the type of model used to detect defects is evaluated. The results showed that YOLOv8 can detect defects in polymeric insulators with precision, sensitivity, and AP above 90% from images captured within the safety distance between the camera and the insulator, as long as the input size is high and there is a large amount of data for training.