A multi-task learning model with feature separation constraint for parameter prediction of gas–liquid two-phase flow

Abstract Precise measurement of gas void fraction is crucial for the control and optimization of various industrial processes. However, due to the intricate dynamics inherent in two-phase flow, measuring gas void fraction remains a challenging task. In this work, we propose a novel deep learning model called Feature Separation Constraint Network (FSCN). First, we conduct gas–liquid two-phase flow experiments in vertical upward pipelines, where temporal and spatial signals about the fluid are collected using a four-sector distributed conductivity sensor. Second, a spatio-temporal convolutional neural network is employed to extract features, and flow pattern classification serves as an auxiliary task to provide global features for gas void fraction prediction. Subsequently, an expert network structure with a feature separation constraint is designed to enhance prediction performance. Finally, comparative analyses with competitive baselines demonstrate FSCN’s superior capability in gas void fraction estimation. A series of ablation experiments confirm the efficacy of each component.