| description abstract | Air pockets trapped in water pipelines are a common phenomenon and can lead to different air-water two-phase flow patterns: stratified, blowback, plug, and bubbly flows. The two former flows contain a large amount of air and should be carefully monitored for pipeline safety, while the two latter flows have relatively low air fractions and can be regarded as normal operating states of pipelines. Hence, flow pattern identification is key to diagnosing the operating state of pipelines. In this paper, a new data analysis method based on complex network theory is proposed to identify the features of flow patterns using pressure signals. The pressure signals of different flow patterns, collected from an experimental facility, were used to characterize the nodes and edges (i.e., connections) in the complex network. The closely linked nodes with dense edges could be aggregated to form a cluster (i.e., community). An unsupervised machine learning technique is then used for community clustering in the network. The results show that the complex network constructed from pressure signals can be divided into several communities, representing different phases (i.e., air, water, or mixed phases) of the air-water flows. Therefore, the flow patterns can be identified in terms of the cluster features and topological features, which are represented by indicators including modularity, graph density, average path length, and transitivity. The impacts of two structural parameters of the complex network, i.e., window size and sliding step, are analyzed. Sliding step is shown to have a more significant impact on the flow pattern identification than window size. This study shows that the complex network approach is effective for flow pattern identification in air-water two-phase flows and could be potentially used for identification of pipeline operational states. | |
