A Convolutional Neural Network for Detecting Faults in Power Distribution Networks along a Railway: A Case Study Using YOLO
João Pedro Augusto Costa, Omar Andrés Carmona Cortes
Abstract
This work presents a Convolutional Neural Network (CNN) called YOLO for detecting failures in components of power lines along a railway. The task is a significant challenge because the CNN has to recognize the object and then classify it in real-time. Moreover, some extra difficulties are presented in the task, such as similarity in terms of color, the intersection of components, the component size, and climate conditions. The failure scenarios have been simulated in a laboratory containing all the structures found in real-world power lines along railways. The laboratory allowed us to build the image dataset containing $$708$$ images with annotations that have been used for training the neural network. Three versions of the Yolo V3 were compared against the state-of-the-art convolutional neural network called Tiny Yolo. Results have shown that Yolo V3 version 2 adequately detects the objects and faults, reaching a precision of $$98\% $$, a recall of $$95\% $$, and a MAP of $$96.58\% $$.