Deep learning based liquid level extraction from video observations of gas–liquid flows. (December 2022)

• Record Type: Journal Article
• Authors: Olbrich, Marc; Riazy, Leili; Kretz, Tobias; Leonard, Terri; van Putten, Dennis S.; Bär, Markus; Oberleithner, Kilian; Schmelter, Sonja
• Is Part Of: International journal of multiphase flow. Volume 157(2022)
• Publication Date: 2022-12
• Subjects: Multiphase flow -- Gas–liquid interface -- Deep learning -- Image processing -- Convolutional neural network -- Videometry; Multiphase flow -- Periodicals; Écoulement polyphasique -- Périodiques; Multiphase flow; Periodicals; 620.1064
• DOI: 10.1016/j.ijmultiphaseflow.2022.104247
• Languages: English
• ISSNs: 0301-9322

Abstract

Abstract: The slug flow pattern is one of the most common gas–liquid flow patterns in multiphase transportation pipelines, particularly in the oil and gas industry. This flow pattern can cause severe problems for industrial processes. Hence, a detailed description of the spatial distribution of the different phases in the pipe is needed for automated process control and calibration of predictive models. In this paper, a deep-learning based image processing technique is presented that extracts the gas–liquid interface from video observations of multiphase flows in horizontal pipes. The supervised deep learning model consists of a convolutional neural network, which was trained and tested with video data from slug flow experiments. The consistency of the hand-labelled data and the predictions of the trained model have been evaluated in an inter-observer reliability test. The model was further tested with other data sets, which also included recordings of a different flow pattern. It is shown that the presented method provides accurate and reliable predictions of the gas–liquid interface for slug flow as well as for other separate flow patterns. Moreover, it is demonstrated how flow characteristics can be obtained from the results of the deep-learning based image processing technique. Highlights: The method extracts the gas–liquid interface from video data of multiphase flows. The used deep learning model predicts accurate segmentation maps of liquid and gas. Reliability of the method is shown by a detailed error analysis. Versatility of the model is demonstrated for independent and unseen data sets.