Flow Pattern Identification Of Boiling Vapor-liquid Two Phase Flow In Mini-channels Based On Neural Networks

It is found in the studies of boiling heat transfer that the vapor-liquid two-phase flow situation on the interface-flow pattern, can greatly affect not only the two-phase flow and heat transfer characteristics of the vapor-liquid, but also the accurate measurement of flow parameters and the determination of the operating characteristics of two-phase system. So the analysis of the pattern recognition of boiling heat transfer is an important part for the vapor-liquid two-phase flow.Firstly, water is taken as the working fluid. The width of the channel is set to be 2mm,1.5mm,1mm and 0.5mm for the boiling heat transfer experiments. The channel is covered with plexi-glass, which is nature to be transparent to achieve the visualization of the experiment. In the study, mass flow rate, temperature, pressure at the entrance of the test section, and fluctuation signal of the differential pressure are measured, besides, thermo-couples are set along the direction of flow to measure the temperature value of different positions.Then the signal obtained use wavelet denoising methods of matlab for processing. The wavelet mode selects wavelet 2-D, the mother wavelet selects haar, the threshold mode selects the unscaled white noise and horizontal details coefs, threshold function selects the soft threshold. After treatment, the signal can greatly assist the experimenter to determine the flow pattern, and reduce subjective errors. After the signal analysis and processing, it can be learned in this experiment that there are three types of flow pattern:single-phase flow, slug flow and limited slug flow.Finally, a dimensionless number is gained by computing the experiments temperature, fluid velocity and differential pressure and other datum. The dimensionless number is taken as the input vector with the temperature measured by thermo-couples; three corresponding flow:single-phase flow (0 01), slug flow (011) and limited slug flow (111) are taken as the output vectors. In this paper, four neural networks:BP, RBF, SOM, and Elman neural networks are established. Part of the datum is used for the establishment and tra i n i ng of neural networks, other part of the datum i s used for neural network validation. The results show that the recognition of flow pattern of BP and Elman neural network are better than another two networks, with the rate of 90% or more, while the correct rate of SOM is less than 50%. And the recognition rate of RBF neural network is just between them. According to the classification of the good results of the two neural networks, the user interface is established, so the users can output the data directly.This paper provides a new method of identifying flow patterns. Comparing with other identification methods, the subjective recognition errors are decreased and the recognition rate of flow can be improved.