A new deep neural network framework with multivariate time series for two-phase flow pattern identification

Abstract

Uncovering flow dynamic behavior of different flow patterns is an important foundation of multiphase flow research. But the traditional classifier is still adopted in the flow pattern identification based on statistical features of experimental measurements, and the utility of data is not sufficient in previous works. Therefore, a novel deep neural network framework is proposed to leverage abundant details of signals. The data was input into the new model after two innovative slicing operations, which combines BiLSTM and CNN to extract the deep characteristic information of different flow patterns. In addition, attention mechanism and residual connection are introduced to improve the network performance. Meanwhile, the dynamic experiment of vertical gas-water two-phase flow is carried out, four-channel conductance signals under five typical flow patterns, namely bubble flow (BF), slug flow (SF), bubble-slug transitional flow (BSF), churn flow (CF) and slug-churn transitional flow (SCF), are collected to feed the network. Finally, in order to verify the effectiveness of the proposed model, some comparative experiments are designed and implemented. The results demonstrate that our proposed model outputs more precise flow pattern identification, which opens up a new way for investigating industrial multiphase flow.