Systematic Feature and Method Selection for Two-Phase Flow Classification Using Multi-Wavelength Near-Infrared Sensors

Accurate classification of two-phase gas-liquid flows is critical for the optimization and safety of industrial processes. This work presents the Multi-Feature Multi-Domain (MFMD) methodology, a systematic framework for flow pattern classification in horizontal air-water flows. Five representative patterns-bubble, plug, slug, stratified, and wave-were identified using collimated, non-intrusive nearinfrared (NIR) optical sensors. The approach relies on the fusion of data from three discrete laser wavelengths (850 nm, 980 nm, and 1310 nm), selected for their distinct absorption characteristics in water and sensitivity to interfacial, hybrid, and volumetric features. Features were extracted from the time (Probability Density Function, PDF), frequency (Power Spectral Density, PSD), and time-frequency (Discrete Wavelet Transform, DWT) domains. We compared an interpretable feature selection strategy, Recursive Feature Elimination (RFE), with two transformation methods: supervised Linear Discriminant Analysis (LDA) and unsupervised Principal Component Analysis (PCA). Classification was performed using Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest (RF), and Artificial Neural Networks (ANN). Results show that combining sensors at the suggested wavelengths is crucial for high performance. The RFE approach achieved 100% accuracy under the evaluated conditions when merging features from all data sources. Moreover, efficient models were obtained by fusing signals across wavelengths regardless of the reduction method. This work culminates in the Optimal Feature and Method Selection (OFMS) guide, which synthesizes the findings into a practical tool for selecting effective strategies while balancing maximum accuracy and model parsimony.