NARX Models of Two-Phase Microchannels Flow in Comparison

Abstract

In this work, two structures of data-driven models have been optimized and compared for the identification and tracking of fast two-phase flows in microchannels. Two-phase flow, consisted of an interlaced sequence of two fluids, as water and air, traveling in a microchannel is defined slug flow and it can be generated by their interaction at a junction. An extensive experimental campaign was performed to collect data and the processes was optically monitored. Two structures of Nonlinear AutoRegressive with eXogenous (NARX) input models, by using Neural Networks (NN) and Wavelet Networks (WN), were compared for modeling the slug flow passage. Two types of patterns were chosen to train and test the networks: single-flow pattern, one per experiment, and multi-flow patterns containing more experimental conditions. The test on single flow patterns highlights the robustness of the models in tracking the slug flow passage and the test on multiple flows patterns confirms the possibility to have one model for different experimental conditions. To underline the potential of these models, some indices were considered to evaluate their performance. The proposed models can represent an important step towards the development of predictive control for real-time System-on-Chip applications.