Detecção automática de equipamentos de pátio em uma subestação de distribuição utilizando redes neurais convolucionais profundas

• Ano de defesa: 2024
• Autor(a) principal: Gomes, Andreas Anael Pereira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Brasil; Programa de Pós-Graduação em Sistemas de Energia; UTFPR
• Palavras-chave em Português: Subestações elétricas; Energia elétrica - Distribuição; Visão por computador; Termografia; Redes neurais (Computação); Falhas de energia elétrica - Detecção; Electric substations; Electric power distribution; Computer vision; Thermography; Neural networks (Computer science); Electric power failures - Detection; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA::SISTEMAS ELETRICOS DE POTENCIA; Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/34343

Multiple factors can aggravate energetic losses in transmission lines, such as the deterioration of contacts or corrosion. Due to that, there is a high demand in the industry for predictive technical maintenance to detect and correct these faults as fast as possible, to guarantee the reliability of energy distribution. Thermography is a non-invasive inspection technique that is quite popular in the industry, since it does not require the equipment to be turned off to be done, with advances in recent years for its AI automatization. Since each equipment may require specific segmentation techniques due to their individual max admissible voltage values, automated thermographic inspection is normally limited to either specific equipment or the detection of generic temperature raises that usually are false positives for faults. This work proposes a combination of neural network architectures for object detection in a power distribution substation. By identifying individual electrical equipment in optical images captured by thermographic cameras and cropping the area of interest in those objects, it would possible to apply a more adequate segmentation of hot spots in the corresponding thermographic image. The object detection models will be compared through different metrics that compare regions of interest annotated by humans on images from a test set that was not used for training with the outputs of the detection models.