Flow pattern recognition for horizontal gas-liquid two-phase flow with a hybrid deep-learning model

In gas-liquid two-phase flow, it is crucial to accurately recognize the flow pattern which serves as the foundation for the study of flow characteristics. Considering the complexity associated with two-phase flow, it is still a great challenge to correctly identify flow patterns. This research proposes an innovative deep learning-based method to categorize flow patterns encountered in gas-liquid flow. A conductance sensor is adopted to acquire the flow information in a horizontal pipeline. The impact of noise on the measured voltage signal is reduced by wavelet threshold processing and sparse coding. For data augmentation, overlapping sampling is conducted. With extracted time-domain features as the input, a hybrid deep-learning model which integrates convolutional neural network and bidirectional long short-term memory is established and trained to identify different flow patterns. Performance comparison between the hybrid model and various deep learning techniques is made. The findings demonstrate that the proposed method's flow pattern identification accuracy reaches 99.85 % indicating that dynamic behavior of the gas-liquid two-phase flow can be well revealed.