Loading...
Publication
Aplicação de Modelos de Machine Learning para Classificação de Mecanismos de Falha e Deteção de Fissuras em Ligas Metálicas Aeronáuticas
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 4.21 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
O presente trabalho explora a aplicação de dois modelos baseadas em machine learning, nomeadamente redes neuronais convolucionais, para avaliação de falhas em ligas metálicas de componentes aeronáuticos da Força Aérea Portuguesa. O primeiro foi projetado para a classificação de mecanismos de falha utilizando imagens obtidas através do microscópio eletrónico de varrimento, enquanto o segundo se foca na deteção de fissuras de fadiga. Para a tarefa de classificação foi utilizada uma arquitetura VGG16, ajustada com uma classification head e treinada utilizando técnicas de transfer learning. Para a deteção foi implementado um modelo baseado em Faster R-CNN com ResNet 101 e Feature Pyramid Network, pré-treinado e configurado para identificar e localizar fissuras de fadiga de vários tamanhos. Ambos os modelos foram treinados e validados em conjuntos de dados anotados manualmente e com recurso a técnicas de data augmentation. O modelo de classificação atingiu uma exatidão de aproximadamente 93% e precisão de 92%, enquanto o modelo de deteção apresentou uma exatidão de cerca de 89% e precisão de 90%. Estes resultados destacam a capacidade das redes neuronais convolucionais em automatizar tarefas complexas, reduzindo o tempo de análise de grandes quantidades de imagens de uma só vez.
The present work explores the application of two machine learning based models, namely convolutional neural networks, for the evaluation of failures in metallic alloys of aeronautical components of the Portuguese Air Force. The first model was designed for the classification of failure mechanisms using images obtained through a scanning electron microscope, while the second focuses on the detection of fatigue cracks. For the classification task, a VGG16 architecture was used, adjusted with a customized classification head and trained using transfer learning techniques. For the detection task, a model based on Faster R-CNN with ResNet 101 and Feature Pyramid Network was implemented, pre-trained and configured to identify and locate fatigue cracks of various sizes. Both models were trained and validated on manually annotated datasets and data augmentation techniques were applied. The classification model achieved an accuracy of approximately 93% and a precision of 92%, while the detection model presented an accuracy of around 89% and a precision of 90%. These results highlight the ability of convolutional neural networks to automate complex tasks, reducing the time required to analyze large amounts of imagens at once.
The present work explores the application of two machine learning based models, namely convolutional neural networks, for the evaluation of failures in metallic alloys of aeronautical components of the Portuguese Air Force. The first model was designed for the classification of failure mechanisms using images obtained through a scanning electron microscope, while the second focuses on the detection of fatigue cracks. For the classification task, a VGG16 architecture was used, adjusted with a customized classification head and trained using transfer learning techniques. For the detection task, a model based on Faster R-CNN with ResNet 101 and Feature Pyramid Network was implemented, pre-trained and configured to identify and locate fatigue cracks of various sizes. Both models were trained and validated on manually annotated datasets and data augmentation techniques were applied. The classification model achieved an accuracy of approximately 93% and a precision of 92%, while the detection model presented an accuracy of around 89% and a precision of 90%. These results highlight the ability of convolutional neural networks to automate complex tasks, reducing the time required to analyze large amounts of imagens at once.
Description
Keywords
Classificação Deteção Fissuras de Fadiga Fractografia Machine Learning Mecanismos de Falha Redes Neuronais
Convolucionais Transfer Learning