Sistemas passivos, como circulação natural, têm sido cada vez mais utilizados para refrigeração de reatores nucleares. A capacidade de um fluido de transferir calor está fortemente relacionada com seu padrão de escoamento, especialmente quando em escoamento bifásico. Estes padrões vêm sendo utilizados em experimentos e modelos de predição de parâmetros que medem esta capacidade. Uma das técnicas não invasivas que vêm sendo utilizadas é a automatização da determinação do padrão de escoamento por meio de imagens. Este trabalho aplicou Redes Neurais Convolucionais para a classificação de imagens de diferentes padrões de escoamento bifásico relacionados à instabilidade chugging da circulação natural. Estas redes, que têm sido o estado-da-arte em classificação de imagens, não se baseiam em características pré-escolhidas, permitindo investigação de novas características para essa tarefa. São comparadas arquiteturas destas redes com diferentes graus de complexidade. Atualmente, a aplicação destas redes ao problema de escoamento bifásico é uma área pouco explorada. No subconjunto de teste, foi obtido um F1-Score médio ponderado de 0,99 e uma acurácia de 99,5%. Os resultados do trabalho mostram que as redes neurais convolucionais apresentam um bom desempenho preditivo e que detêm recursos ainda não explorados para fins de classificação de padrões em imagens de escoamento bifásico.

Passive systems using natural circulation have been applied to new designs of nuclear power plants. Heat transfer capacity of a fluid is strongly correlated with its flow pattern, and these patterns have been used as basis to experiments and prediction models to obtain heat transfer related parameters. One of the most recently investigated non-invasive techniques to estimate, and predict those parameters, has been the imaging of two-phase flow patterns under natural circulation. These images are usually classified based on automated algorithms using artificial intelligence or machine learning techniques which were usually based on predefined feature extraction. This work has used Convolutional Neural Networks (CNNs) to classify images of two-phase flow patterns of natural circulation instabilities. This neural network, which has been the state-of-the-art for image classification, is not based on previous chosen image features, enabling more accurate classification results, and allowing investigation of novel features to obtain the best classification. This work has investigated different structures and configurations of these neural networks and has verified their application to the two-phase flow problem. The CNNs were applied to experimentally acquired images of a Natural Circulation Circuit where cyclical instabilities (called chugging) were generated. The database was composed of 1152 images which were divided, mainly, in three different classes relative to each of the three stages of the chugging cycle. An accuracy of 99.5% with a weighted F1-score mean of 0.99 was obtained for the test subset. These results have testified a good predictive capacity of these neural networks. During this work development, many implementation details were verified and described, including the use of Class Activation Mapping technique which allowed better comprehension of classification mechanisms used by CNNs. Many future works are recommended based on this initial CNN application for pattern recognition tasks on two-phase flow images.