Multiphase flow measurement is essential in industries like oil and energy, where accurate
data can enhance production quality and reduce costs, particularly during processes like
oil extraction. Traditional sensors, such as differential pressure and gamma-ray sensors,
struggle with the complexity and variability of these flows, often leading to significant
errors.
This research introduces a novel machine learning-based approach combined with physical
sensors to improve the accuracy of multiphase flow measurements. It focuses on
multimodal electrical tomography sensor arrays, a non-intrusive method, and applies
advanced machine learning techniques, including convolutional neural networks (CNNs),
long short-term memory (LSTM) networks, and Transformer models. These models
capture spatial and temporal features of the flow and were rigorously evaluated through
experiments at the University of Leeds Multiphase Flow Laboratory, where data was
gathered from 185 different flow conditions.
Two models were developed: a CNN-LSTM model and an end-to-end Multiphase Flow
Estimation model (MFENet). MFENet, based on a generative framework and multi-head
attention, achieved outstanding prediction accuracy, with relative mean errors of 2.4% for
gas flow rates and 1% for liquid flow rates, demonstrating superior performance compared
to conventional methods.
Furthermore, the study explored dimensionality reduction techniques, reducing the input
size of tomographic images to just 21 mean concentration values. This approach
maintained high predictive accuracy, reducing the gas flow rate prediction error to 2.1%.
Ablation experiments further refined temporal and spatial input projections, enhancing
computational efficiency and overall performance.
This research demonstrates the potential of combining machine learning with physical
sensors to create a robust, real-time framework for accurate multiphase flow measure
ments, offering significant improvements in both accuracy and computational efficiency
over traditional methods.
