Automatic Recognition of Electrical Grid Elements using Convolutional Neural Networks

Due to the extensive proportions of Brazilian railways, there is a high demand for remote and automatic diagnose tools. This work proposes a scene selection method using Deep Learning techniques, namely Convolutional Neural Networks (CNN), to recognize the poles, which gathers objects of interest to be inspected in the railway power network. Videos were obtained through the railway and the data divided and preprocessed for the network training and testing. A VGG network architecture served as a starting point, and after exhaustive search and comparisons of many techniques, two network topologies are presented and compared in field tests. The results yield more than 93% efficiency for both proposed topologies.