Dispositivo de borda baseado em rede neural convolucional para inspeção multiespectral de painéis fotovoltaicos

• Ano de defesa: 2024
• Autor(a) principal: Di Renzo, André Biffe
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Brasil; Programa de Pós-Graduação em Engenharia Elétrica e Informática Industrial; UTFPR
• Palavras-chave em Português: Painéis solares; Localização de falhas (Engenharia); Redes neurais (Computação); Sistemas embarcados (Computadores); Algorítmos computacionais; Processamento de imagens; Visão por computador; Inteligência artificial; Solar panels; Fault location (Engineering); Neural networks (Computer science); Embedded computer systems; Computer algorithms; Image processing; Computer vision; Artificial intelligence; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA; Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/35695

The use of photovoltaic panels for sustainable electricity generation has been growing worldwide. As a result, large solar plants must be monitored to find defects, quickly avoiding extended shutdowns in energy generation. There are several techniques for monitoring solar plants. Some of them are visual inspection in the visible spectrum, infrared thermography, and the acquisition of fluorescence images after excitation of the panel by ultraviolet light. All of these techniques use cameras to capture images of the panel for later analysis. Even with só many techniques, each of them has its shortcomings. In this work, a multispectral camera was developed to acquire and classify images of photovoltaic panels in real-time, using an edge device to connect all the hardware and process the images. In addition, a light concentrator was developed só that ultraviolet light inspections can be carried out efficiently. This research project also reviewed techniques that can be used to improve or clarify the aforementioned techniques, aiding in the inspection of solar panels. Regarding computational algorithms, two methodologies were tested for inspecting photovoltaic panels. One uses direct image classification (without object detection) and classifies the acquired image with a convolutional neural network. The other technique uses a complete algorithm that detects and classifies solar panels using convolutional neural networks. Both methodologies presented results similar to or better than those found in the literature, with accuracy values above 90% for the direct classification methodology and accuracy above 86% for the methodology with object detection. The presented solution can be an essential tool for the maintenance team to perform preventive maintenance, allowing early identification of problems and planning the maintenance schedule of the solar plant.