Two-Phase Flow Regimes Characterization via Pressure Heat Map Recognition

Abstract

Precise recognition of flow regimes in industrial two-phase flow system is reliant on visual aids. However, opacity of pipe medium poses a significant challenge and presents a pressing need for reliable identification of flow regimes. This study aims to characterize stratified, slug, elongated bubble and dispersed bubble flow by evaluating the performance of four prominent deep learning architectures. Numerical investigation was conducted for pressure signal data collection for individual flow regimes by varying inlet gas and liquid superficial velocities. Frequency with time domain heatmaps were generated from pressure signals and were pre-processed. Data sets were trained, validated and tested using EfficientNetB80, MobileNetV2, Xception, and DenseNet deep learning architectures. Gradient-weighted Class Activation Mapping (Grad-CAM) were created to visualize the targeted features in each architecture. Corresponding performances were evaluated to identify the best deep learning architecture for flow regime characterization from pressure signals heatmaps. MobileNetV2 with accuracy of 92.5% outperformed EfficientNetB80 and DenseNet each with accuracy of 88.46% and Xception having accuracy of 84.62%. Higher rate of confusion in distinguishing between elongated bubble and dispersed bubble flow lead to lower accuracies in the applied lagging models. MobileNetV2 proves to be an effective CNN architecture for identifying two-phase flow regimes from pressure signal heatmaps in industrial systems.