Detecção e classificação de níveis de cavitação em sistemas industriais: uma abordagem de clusterização e comparação de diferentes arquiteturas de redes neurais convolucionais

Abstract

This work proposes an approach for the detection and classification of cavitation levels in industrial systems using clustering techniques and a comparison of different convolutional neural network (CNN) architectures. Cavitation, a phenomenon that causes mechanical damage in pumping systems, was analyzed through images captured in an experimental setup. The methodology involved constructing an image dataset, where an acquisition system was configured to capture cavitation bubble formation in a transparent acrylic tube. Lighting was controlled to ensure image quality, and a pressure transducer was used to monitor suction pressure levels. Architectures such as VGG16, VGG19, ResNet50, and ResNet101 were explored, with the application of filters like Prewitt and Sobel to enhance feature detection. The clustering process used GMM to group the data, and dimensionality reduction was performed using PCA. The results showed that the VGG19 architecture achieved 88% accuracy in supervised classification, while ResNet101 obtained 91%. Additionally, the VGG architectures consumed less RAM compared to the ResNet architectures, especially with large volumes of data. Performance analysis also revealed that the application of filters improved cluster quality, facilitating the separation of cavitation levels. On diversified test data, with variations in lighting and captured on different days, the model with VGG19 achieved an average precision of 79.97%. This study highlights the effectiveness of computer vision and machine learning techniques in predictive cavitation monitoring, contributing to efficient maintenance and cost reduction in industrial environments.