关键词: 上升管, 气液两相流, 流型识别, 卷积神经网络, 实验验证

Abstract:

A flow pattern recognition method that combines pressure difference data recursive graph and convolutional neural network is proposed to address the difficulty of identifying gas-liquid two-phase flow patterns in the gathering and lifting pipeline system. Two-phase flow experiments were conducted using air and water as media in the gathering and lifting pipeline system, and four typical flow patterns were observed: bubbly flow, annular flow, slug flow, and severe slug flow. Based on the phase space reconstruction method, recursive graphs of pressure difference signals for each flow pattern were constructed, and it was found that the recursive graphs of the four flow patterns all exhibited different characteristics. Due to the relatively regular period of annular flow, there are more main diagonals in the recursive plot, and the state trajectory has higher repeatability, showing a more regular appearance; For bubbly flow, due to the random distribution of bubbles, the pressure difference signal exhibits irregularity in the recursive graph, with isolated points occupying a large proportion and fewer diagonal structures, presenting a certain degree of randomness; For slug flow, in the state of intermittent motion between gas and liquid phases, short diagonal lines and block like small structures coexist in the recursive graph. Three different classical convolutional neural networks (LeNet, AlexNet, ResNet) were constructed to extract the diagonal and grid like features of the recursive graph of flow regime pressure difference data. The model performance was evaluated using the AUC (Area Under Curve) of Receiver Operating Characteristic Curve (ROC). The results showed that LeNet achieved the highest accuracy, sensitivity, and specificity in identifying various flow patterns.

Key words: riser, gas-liquid flow, flow pattern identification, convolutional neural network, experimental validation