An intelligent data-driven flow regime recognition method for horizontal air-water two-phase flow

Flow regime recognition is very important in the two-phase flow measurement. However, considering that two-phase flow is much more complicated than single-phase flow, flow regime cannot be accurately identified by mechanism model. In this work, a novel intelligent data-driven method based on image encoding and transformer is proposed to recognize typical flow regimes encountered in the horizontal air-water flow. Dynamic experiment is carried out using a ring-shaped conductance sensor to collect voltage signals of bubble flow, bubble-slug flow, slug flow, slug-stratified flow and stratified flow. To highlight the characteristic differences between different flow regimes, the measured signals are encoded into two-dimensional images. To classify the encoded images of the five flow regimes, transformer models are then established. With the encoded images as the input of the model, flow regime identification is implemented by training of the model. The results demonstrate that the characteristics of different flow regimes can be better reflected in the encoded image with Gramian angular field. Meanwhile, the recognition accuracy of Swin Transformer is advantageous to that of Vision Transformer in the classification of the encoded images of the five flow regimes. Comparing with other identification methods, the method which combines Gramian angular field with Swin Transformer shows the best performance in the recognition of the flow regimes. The total accuracy reaches as high as 99.1 % This study offers an alternative for accurate flow regime recognition in two-phase flow measurement.