Optimized spectral feature selection with lightweight CNN for precise volume fraction estimation in three-phase flows

The ability to accurately calculate volumetric fractions for three-phase flow: gas, water and oil, is critical for industrial applications in the oil and gas and chemical engineering industry. This paper develops a methodology that incorporates the acquisition of signals using X-rays, spectral features and machine learning. Spectral features representing mean energy, variance, full width at half maximum (FWHM), entropy, kurtosis, and cepstral coefficients, were obtained from 108 Monte Carlo N-Particle (MCNP) energy simulation sets through annular, homogeneous and stratified flow regimes and 36 combinations of volume fractions ranging from 10 % to 80 %. The Beetle Antennae Search (BAS) algorithm optimizes the spectral features using measures that reduce the number of features to six features to improve overall prediction efficiency, these six features are the final input for a lightweight convolutional neural network (CNN). The chosen samples from the MCNP simulations (86 samples for training input and 22 for testing input, referred to as "the dataset") were used for training and testing the model in order to predict the three volume fractions. The predictions resulted in low root mean square error (RMSE) values, documented as 0.15 for gas, 0.13 for water and 0.14 for oil with R² values all exceeding 0.94; residuals were assessed to determine the model's stability and showed that the residuals were clustered around 0 for both training and testing sets. This approach offers a real-time and computationally-efficient technique for use in industry, but the reliance on simulated data and number of existing samples means there is still room for improvement in the future. The study demonstrates a scalable method for multiphase flow analysis, which could be verified with real experimental data in future studies to improve on the overall detection method reliability.