Aplicação de técnicas de inteligência artificial na detecção de defeitos em para-raios através da análise de termografias.

Abstract

Zinc oxide (ZnO) lightning arresters are essential for protecting electrical systems against overvoltages, ensuring greater operational reliability. However, it is crucial that these devices are always in good working order. Given this, there is a growing demand for efficient fault detection techniques, especially non-invasive methods such as external temperature analysis. Thermography stands out as a powerful tool, capable of identifying temperature variations that may indicate lightning arrester failures. This work focuses on the implementation of intelligent classifiers through an SVM and convolutional neural networks for the classification of lightning arrester defects from thermal images. Different neural network architectures, including generic CNN, VGG19 and Inception, were analyzed with the aim of extracting features from thermographic images and identifying good and defective lightning arresters. It was observed that the performance of the networks is influenced by the complexity of the architecture and the restricted size of the database, with some more advanced architectures showing signs of overfitting. The project seeks to improve these networks for better performance in defect classification.