Automatic defect recognition for electrical equipment with artificial neural networks

Local heating is a sign of the development of defects in electrical equipment. It is detected by thermal imaging control. The result is thermal images of working electrical equipment - thermograms, which are analyzed using a visual assessment. However, in order to reduce the human factor in processing a large number of thermograms taken from unmanned aerial vehicles, the development of an automated defect recognition system has been proposed. The paper presents the results of recognizing defects in electrical equipment with artificial convolutional neural networks (CNN). It is proposed to use the learning transfer mechanism, the SqueezeNet architecture to classify thermograms into two classes: with a defect and without a defect. To assess the effectiveness of classification, the following metrics are proposed: sensitivity, specificity, balanced accuracy, Matthews correlation coefficient, F-measure. Testing of the trained CNN on a control sample of thermograms not used in training proves the effectiveness of the CNN in the tasks of recognizing defects in electrical equipment. It is promising to continue research on improving the speed and quality of classification of thermograms.