Processamento de imagens para a extração de parâmetros de escoamentos bifásicos gás-líquido utilizando redes neurais convolucionais

Abstract

Multiphase flows are widely encountered in several industrial sectors, particularly in the nuclear, chemical, and oil and gas industries. These flows are composed of two or more phases, which may assume different geometric configurations depending on the physical properties of the phases and the operating conditions. In oil and gas production, the flows are predominantly composed of water, oil, and natural gas. In the Brazilian context, the pre-salt reservoirs account for a large share of the country’s oil and gas production, corresponding to approximately 80% of the total, according to data released by the National Agency of Petroleum, Natural Gas and Biofuels (ANP). The exploitation of these reservoirs occurs mainly in deepwater environments, characterized by high pressures and high CO₂ contents, features that directly affect flow behavior and require new strategies for the design, operation, and monitoring of production systems. The acquisition of experimental information on this phenomenon is essential for establishing engineering correlations and for developing models that describe its physical behavior. This work presents a methodology based on image processing techniques for the extraction of characteristic parameters of gas–liquid twophase slug flow. Experimental data are obtained using a test facility with controlled temperature and pressure, employing model fluids (mineral oil and sulfur hexafluoride) to emulate field conditions. Images acquired at high acquisition rates are segmented using a U-Net convolutional neural network, enabling the distinction between liquid and gas regions, with an average Dice coefficient higher than 0.95. Elongated bubbles are identified and reconstructed through the aggregation of consecutive frames, based on the bubble nose velocity. The parameters that characterize the studied flow are estimated and compared with classical models from the literature, manual measurements, and techniques based on capacitive sensors, showing deviations on the order of ±10% relative to the adopted references. The proposed methodology demonstrates the capability to automatically and robustly estimate the main characteristic parameters of gas–liquid two-phase slug flow.