Objective Flow Pattern Identification and Classification in Inclined Two-Phase Flows Using Machine Learning Methods

David H Kang, Purdue University

Abstract

Two-phase modeling and simulation capabilities are strongly dependent on the accuracy of flow regime identification methods. Flow regimes have traditionally been determined through visual observation, resulting in subjective classifications that are susceptible to inconsistencies and disagreements between researchers. Since the majority of two-phase flow studies have been concentrated around vertical and horizontal pipe orientations, flow patterns in inclined pipes are not well-understood. Moreover, they may not be adequately described by conventional flow regimes which were conceptualized for vertical and horizontal flows. Recent work has explored applying machine learning methods to vertical and horizontal flow regime identification to help remedy the subjectivity of classification. Such methods have not, however, been successfully applied to inclined flow orientations. In this study, two novel unsupervised machine learning methods are proposed: a modular configuration of multiple machine learning algorithms that is adaptable to different pipe orientations, and a second universal approach consisting of several layered algorithms which is capable of performing flow regime classification for data spanning multiple orientations. To support this endeavor, an experimental database is established using a dual-ring impedance meter. The signals obtained by the impedance meter are capable of conveying distinct features of the various flow patterns observed in vertical, horizontal, and inclined pipes. Inputs to the unsupervised learning algorithms consist of statistical measures computed from these signals. A novel conceptualization for flow pattern classification is developed, which maps three statistical parameters from the data to red, green, and blue primary color intensities. By combining the three components, a flow pattern map can be developed wherein similar colors are produced by flow conditions with like statistics, transforming the way flow regimes are represented on a flow regime map. The resulting dynamic RGB flow pattern map provides a physical representation of gradual changes in flow patterns as they transition from one regime to another. By replacing the static transition boundaries with physically informed, dynamic gradients between flow patterns, transitional flow patterns may be described with far greater accuracy. This study demonstrates the effectiveness of the proposed method in generating objective flow regime maps, providing a basis for further research on the characterization of two-phase flow patterns in inclined pipes. The three proposed methods are compared and evaluated against flow regime maps found in literature.

Degree

M.Sc.

Advisors

Kim, Purdue University.

Subject Area

Artificial intelligence|Electrical engineering

COinS