Classification of two phase flows using linear discriminant analysis and expectation maximization clustering of video footage

Abstract

Two-phase adiabatic flow of refrigerant R134a was studied and an automated flow regime detection algorithm was developed. For a saturation temperature of 15 °C, 5 s video data was captured for mass fluxes ranging from 200 kg/m 2 s to 500 kg/m 2 s in smooth horizontal tubes. The vapor quality varied between 0 and 1, the I.D. was 8 mm. Slug flow, intermittent flow and annular flow were discerned. A flow regime is assigned to each video stream as a whole, instead of on a frame by frame basis. In this way, each data point contains both high resolution temporal and spatial information. The dimensionality of the data is reduced by a combination of manual parameter selection and linear discriminant analysis. The classification of the reduced data is then done by unsupervised clustering with the expectation maximization algorithm. Based on our visual classification, all of the slug and annular flows are correctly recognized, intermittent flows are identified with 95% accuracy. The method can be used for automatic online flow regime identification.