Monitoramento automatizado de isolamentos por imageamento ultravioleta empregando aprendizado profundo

Abstract

The presence of surface discharges or corona discharges in the vicinity of equipment and components of the electrical system is, in general, an indication of the occurrence of some undesirable phenomenon. In many cases, it potentially indicates a process that can lead to the failure or physical degradation of materials. One of the most promising techniques of corona effect monitoring is the use of specialized cameras for the detection of ultraviolet radiation. This dissertation presents an innovative algorithm for classifying the criticality of insulation based on attributes extracted from videos recorded with a camera capable of detecting ultraviolet radiation. The proposed methodology is based on extracting three attributes from each detected facula origin: maximum persistence, area and minimum distance between facula origin and isolation. To obtain this last attribute, we proposed a methodology for segmenting insulation in images using a combination of a deep convolutional neural network model and an adaptive thresholding method based on the mean. The deep learning model achieved 85.5% precision in detecting insulation on a validation dataset consisting of 1985 images and 8730 instances. The classification results showed that the distance between insulation and facula origin is an essential attribute for video analysis. This variable provides context for recorded discharges and allows differentiation between cases where ultraviolet radiation originates from insulation and those where discharge location is less critical.