A Software Technique for Oil-Water Two-Phase Flow Measurement: CapsNet with Multi-task Learning

Abstract

Flow parameters measurement is beneficial for understanding oil-water two-phase flow. Due to the changeable flow structures of oil-water two-phase flow, the prediction of superficial velocity of oil-water two-phase flow in large diameter pipes is still a challenging problem. In this paper, a novel soft measurement technique based on Capsule Network (CapsNet) is developed to predict the superficial velocity. Firstly, a vertical upward oil-water two-phase flow experiment in a 125 mm ID pipe was conducted, and response signals at different flow conditions were obtained by a vertical multi-electrode array (VMEA) conductance sensor. Then, in order to increase the number of samples without losing information, a new data pre-processing (1D-to-2D) technique is used. Finally, a novel multi-task learning network based on CapsNet is designed to predict the flow pattern and superficial velocity of each phase. To verify the advancedness of the method, we compared the proposed network with its variations and other competitive networks. The results suggest the proposed network achieves the best performance for prediction of flow pattern and superficial velocity. The proposed method presents great potential for handling high-dimensional, time-varying and nonlinear problems in multiphase flow.