Automated flow pattern recognition for liquid-liquid flow in horizontal pipes using machine-learning algorithms and weighted majority voting

The simultaneous liquid-liquid flow usually manifests various flow configurations due to a diverse range of fluid properties, flow-controlling processes, and equipment. This study investigates the performance of machine learning (ML) algorithms to classify nine oil-water flow patterns (FPs) in the horizontal pipe using liquid and pipe geometric properties. The MLs include Support Vector Machine, Ensemble learning, Random Forest, Multilayer Perceptron Neural Network, k-Nearest Neighbor, and weighted Majority Voting (wMV). Eleven hundred experimental data points for nine FPs are extracted from the literature. The data are balanced using the synthetic minority over-sampling technique during the MLs training phase. The MLs' performance is evaluated using accuracy, sensitivity, specificity, precision, F1-score, and Matthews Correlation Coefficient. The results show that the wMV can achieve 93.03% accuracy for the oil-water FPs. Seven out of nine FPs are classified with more than 93% accuracies. A Friedman's test and Wilcoxon Sign-Rank post hoc analysis with Bonferroni correction show that the FPs accuracy using wMV is significantly higher than using the MLs individually (p<0.05). This study demonstrated the capability of MLs in automatically classifying the oil-water FPs using only the fluids' and pipe's properties, and is crucial for designing an efficient production system in the petroleum industry.