Speaker: Simon Giard-Leroux, M. Sc. student in Computer Science, Department of Computer Science, Faculty of Science, Université de Sherbrooke
Abstract: In the electrical power industry, arc flash studies are performed by engineers to evaluate the incident energy to which a person working on equipment would be exposed if an accident caused a short circuit. This energy can vary according to several parameters, one of the main ones being the duration of the arc, which depends on how quickly the protective equipment supplying the equipment where the fault occurs is cut off. It is therefore important to correctly identify the protective equipment in an electrical network, including the types of fuses, which can be difficult to identify with the naked eye from photos of electrical installations. However, they can be identified by their physical characteristics, such as their color or shape. This observation task must currently be done manually, so a more automated solution would be advantageous.
In parallel, the field of object detection using deep learning has experienced a remarkable growth in the last few years in order to localize and identify the type of different objects in images. By applying a supervised learning strategy, it is possible to train and optimize an object detection model based on neural networks that can identify fuses in new images never seen before. In this dissertation, we will discuss the use of deep learning-based object detection techniques to automate the identification of electrical fuses, including the Faster R-CNN, RetinaNet and DETR models. We will pose the problem of fuse identification in the context of arcing studies, describe the principles behind the object detection techniques used in deep learning, and propose a methodology to optimize a final detection model that can be used in industry with a high identification performance, allowing to significantly accelerate the work of electrical engineers in this task.
A paper detailing our methodology to achieve a final AP50 performance of 91.06% has been submitted for publication to the journal IEEE Transactions on Industrial Informatics and is currently under review. This result demonstrates that fuses can be adequately predicted in new electrical survey photos. The developed code, the dataset of more than 12,000 fuses and a user interface to use the final model in an industrial context are shared openly with the scientific community.
Jury member, presidentPierre-Marc Jodoin, Professor, Department of Computer Science, Faculty of Science, Université de Sherbrooke
Jury member, research director: Martin Vallières, Professor, Department of Computer Science, Faculty of Science, Université de Sherbrooke
Jury member, research co-director: François Bouffard, Professor, Department of Electrical and Computer Engineering, McGill University
Jury member, external evaluator: Christian Gagné, Professor, Department of Electrical and Computer Engineering, Université Laval
All interested persons are cordially invited.
Teams link: https://bit.ly/3zSvNew