Análise de imagens termográficas no diagnóstico de manutenção preditiva

Abstract

Thermographic analysis of electrical items and systems is a non-intrusive predictive maintenance technique, so it can be applied without interrupting the operation of the system. Currently the diagnosis condition from thermographic analysis is performed by a specialist, be it an electrical engineer or a technical expert. Therefore, it is a costly task and requires skilled workers. In this paper, we propose a system that automatically classifies thermographic images of fuses, based on features extracted from the image through a pre-trained convolutional neural network, and aims to assist professionals in decision making. Specifically, the classifier parameters are tuned from K-means produced for each class of NH- and D-type fuse thermographic images. The database was acquired from thermographic analysis reports and categorized in order to be applied to classification tasks, then the database was used in the process of training and validating a variety of convolutional network architectures in conjunction with classification methods, which were evaluated using cross-validation. The best results indicate that the VGG16 network in conjunction with the SVM classifier, properly parameterized, can diagnose the condition of NH-type and D-type fuses with a hit rate of approximately 93%.