A carregar...
Publicação
SISTEMA DE VISÃO TERMOGRÁFICO BASEADO EM APRENDIZAGEM PROFUNDA PARA DETEÇÃO DE DEFEITOS INTERNOS EM MATERIAIS COMPÓSITOS DE MATRIZ POLIMÉRICA
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 3 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
A presente dissertação contemplou o desenvolvimento de um sistema de visão
termográfico baseado em aprendizagem profunda para a deteção de defeitos internos em
materiais compósitos de matriz polimérica reforçados com fibras de vidro, carbono ou kevlar.
Para o seu desenvolvimento, foram utilizados provetes produzidos por manufatura
aditiva, especificamente pelo processo Fused Deposition Modeling (FDM). A metodologia
englobou o uso da termografia ativa como ferramenta para adquirir imagens termográficas e
posterior rotulações, seguido da utilização de um modelo Region Convolutional Neural
Network (R-CNN). O treino e teste desse modelo foram realizados por meio de Transfer
Learning (TF), utilizando uma ResNet 50. Algoritmos como Selective Search (SS) e
Intersection Over Union (IoU) foram fundamentais para a sua implementação.
A implementação de métodos de undersampling e oversampling para o balanceamento
de dados, juntamente com a variação do Learning Rate (LR), foram fatores a ter em conta para
a obtenção de resultados nas diferentes métricas (precisão, recall, F1-score e exatidão) na fase
de teste da CNN. Para os testes de inspeção offline da CNN utilizando imagens foram obtidos
resultados superiores a 92% para todas as métricas.
Por fim, foram realizados testes de inspeção inline utilizando um vídeo completo do
ensaio de inspeção (termografia), alcançando uma exatidão na deteção de defeitos de 84,4%.
This dissertation involved the development of a thermographic vision system based on deep learning to detect internal defects in polymer matrix composite materials reinforced with glass, carbon or kevlar fibres. For its development, specimens produced by additive manufacturing were used, specifically by the Fused Deposition Modelling (FDM) process. The methodology included the use of active thermography as a tool for acquiring thermographic images and subsequent labelling, followed by the use of a Region Convolutional Neural Network (R-CNN) model. This model was trained and tested by means of Transfer Learning (TF), using a ResNet 50. Algorithms such as Selective Search (SS) and Intersection Over Union (IoU) were fundamental to its implementation. The implementation of undersampling and oversampling methods for data balancing, together with the Learning Rate (LR) variation, were factors to be taken into account in order to obtain results in the different metrics (precision, recall, F1-score and accuracy) in the CNN testing phase. For the CNN offline inspection tests using images, results of over 92% were obtained for all the metrics. Finally, inline inspection tests were carried out using a complete video of the inspection test (thermography), achieving a defect detection accuracy of 84.4%.
This dissertation involved the development of a thermographic vision system based on deep learning to detect internal defects in polymer matrix composite materials reinforced with glass, carbon or kevlar fibres. For its development, specimens produced by additive manufacturing were used, specifically by the Fused Deposition Modelling (FDM) process. The methodology included the use of active thermography as a tool for acquiring thermographic images and subsequent labelling, followed by the use of a Region Convolutional Neural Network (R-CNN) model. This model was trained and tested by means of Transfer Learning (TF), using a ResNet 50. Algorithms such as Selective Search (SS) and Intersection Over Union (IoU) were fundamental to its implementation. The implementation of undersampling and oversampling methods for data balancing, together with the Learning Rate (LR) variation, were factors to be taken into account in order to obtain results in the different metrics (precision, recall, F1-score and accuracy) in the CNN testing phase. For the CNN offline inspection tests using images, results of over 92% were obtained for all the metrics. Finally, inline inspection tests were carried out using a complete video of the inspection test (thermography), achieving a defect detection accuracy of 84.4%.
Descrição
Palavras-chave
deteção termografia Region Convolutional Neural Network (R-CNN) Transfer Learning (TF) Selective Search (SS)