Classification of distribution power grid structures using inception v3 deep neural network

To maintain the supply of electrical energy, it is necessary that failures in the distribution grid are identified during inspections of the electrical power system before shutdowns occur. To automate the inspections, artificial intelligence techniques based on computer vision are proposed. Due to the low number of visible faults, it is difficult to train deep learning models based on images of electrical power system inspections. In this paper, it is proposed to use segmentation and edge detection techniques to increase the database, making classification possible using the Inception v3 deep neural network model. From a pre-processing using the Gaussian filter to smooth the image, the techniques of the threshold with binarization, adaptive binarization, and Otsu and riddler-calvard are used for segmentation; and for edge detection, the sobel and canny techniques are used. The Inception v3 had better results than VGG-16 and ResNet50, considering mean squared error, root mean square error, accuracy, precision, recall, F-measure, and speed to convergence in this application.