Estimativa de vazões em escoamentos tubulares através da análise de imagens térmicas por redes neurais

Abstract

Tradicional flow measurement methods are invasive and costly to operate. This paper proposes a non-invasive method for flow rate monitoring using thermal imaging and convolutional neural networks (CNN). The approach integrates principles of heat transfer, thermography, and artificial intelligence, aiming to demonstrate that different flow conditions generate characteristic thermal patterns on the outer surface of pipelines. A controlled experimental system was developed, enabling the capture of thermal images under four distinct flow conditions. The collected images were then used to train CNN models via the transfer learning technique. The model that exhibited the best performance was DenseNet169, featuring a classifier layer with 512 neurons, which achieved 75% accuracy and 0.7550 F1-Score. This result demonstrates that thermal images contain sufficient and discernible information to correlate temperature patterns with flow conditions,thereby validating the proposed method.