A hybrid deep learning model towards flow pattern identification of gas-liquid two-phase flows in horizontal pipe

Horizontal gas-liquid two-phase flow is common in oilfield gathering and transportation systems, and the flow pattern significantly impacts pipeline operation. To ensure the safety and stability of pipeline operations, it is essential to develop more precise identification techniques for gas-liquid two-phase flow patterns. Traditional deep learning methods have single features, are difficult to capture the dynamic features of the flow pattern, and have poor ability to deal with disturbing information. This paper introduces a novel conductance sensor, with an optimized measurement design to achieve higher accuracy. Gas-liquid two-phase flow experiments were conducted on a horizontal pipe experimental platform with an inner diameter of 50 mm using this sensor, and flow parameters were collected under different flow patterns. The TFN-STFT-CBAM-Anomaly Transformer model was proposed, utilizing the raw data collected by the conductance sensor as input for identifying gas-liquid two-phase flow patterns, thereby enhancing identification efficiency. This model integrates the time-frequency transformation (TFN) network with the Convolutional Block Attention Module (CBAM) attention mechanism and synergizes with the Anomaly Transformer module. The model is effectively able to extract multi-level and multi-scale features to understand the dynamic evolution of the flow pattern in a more comprehensive way, with higher expressive capability and better robustness in coping with anomalous perturbations in practical applications. Additionally, ablation experiments were conducted to validate the effectiveness of the model's components. The superiority of the proposed model was further validated through comparison with the latest methods. The model achieved an accuracy of 97.36 % for flow pattern identification using raw data with different flow patterns.