Characterization of transient differential pressure signal features and flow pattern identification in horizontal two-phase flow through a constriction with machine learning models

Flow pattern identification and phase flow rate measurement of two-phase gas–liquid flows are of fundamental importance for process monitoring and control in several industrial applications. Differential pressure ($\Delta p$) based flow sensors are robust and reliable devices, with no moving parts, therefore very suitable for extreme environment applications such as deep offshore. These well-known sensors typically relate the mean differential pressure across a throttle device to the mixture flow rate. However, for the determination of phase flow rates, information about phase fraction is necessary. Beyond the average differential pressure, a wealth of information can be extracted from the transient $\Delta p$ signal that can be useful for flow pattern identification and phase flow rate determination, without the use of additional sensors. In this paper, we present a thorough analysis of those features that have been extracted from the PDF, PSD, and DWT representations of the differential pressure signal. These features are then used for flow pattern determination based on deep neural networks, support vector machine, and K-nearest neighbor classifiers. The data are extracted for the flow of water–air mixture ranging from 0.03 to 1.28 m/s liquid superficial velocity $j_l$ and 0.03 to 20 m/s of gas superficial velocity $j_g$ in a horizontal configuration of a 25.4 mm internal diameter, therefore covering most flow patterns encountered in gas–liquid flows. Two orifice plates with diameters of 12.7 mm and 18.8 mm were used as throttle devices. Through data correlation analysis, a feature selection was performed following a geometry independence criterion. Therefore, the selected features are expected to be representative of the underlying flow characteristics of the upstream flow, irrespective of the orifice geometry. Results show that the prioritization of these selected parameters as inputs for the classifiers results in a more generalizable model.