Adaptive spatio-temporal feature extraction and analysis for horizontal gas-water two-phase flow state prediction

Abstract

Gas-water two-phase flow widely exists in chemical industrial processes, which contains multiple flow states such as bubble flow, slug flow, wave flow, etc. Transition states between different flow states in the practical flow process present complex dynamics and spatio-temporal evolution characteristics. Accurate flow state prediction is of great significance for efficient utilization of energy, reasonable control of chemical reaction and safe operation of processes. However, with the temporal correlations and complex spatial manifold in the dynamic flow process, it is difficult for conventional methods to precisely characterize and predict the flow states. To predict the gas–water two-phase flow state especially the transition state, this work proposes a flow state prediction strategy that extracts the spatio-temporal features of the process from multimodal measurement signals. The locality preserving slow and steady feature analysis (LPS2FA) is first proposed to extract state features by concurrent consideration of the high-order temporal dynamics and manifold structure to characterize the flow states. Moreover, incremental LPS2FA (IncLPS2FA) is developed to learn the transition characteristics online, where the tuning parameter is adaptively adjusted by comparing the distance information in “time neighborhood” and “space neighborhood”. Then, canonical correlation analysis (CCA) is utilized to maximize the correlation between the past and future state features to establish the prediction model. Additionally, a novel indicator called long-term crest factor similarity (LTCFS) is defined to intuitively indicate the transition states by quantifying the long-term waveform variation and equilibrium relationship of the state features. The effectiveness and superiority are verified by the experiment of gas–water two-phase flow.