Zero-shot Learning Model with Dynamic and Static Feature Fusion for Industrial Gas-liquid Two-phase Flow State Identification

Gas-liquid two-phase flow is prevalent in various industrial processes. Due to safety constraints and high costs, there exists a lack of training samples for flow states in real industry, resulting in traditional data-driven methods no longer being feasible. To address this challenge, a zero-shot learning (ZSL) model for flow state identification is proposed from the perspective of attribute description and attribute migration. In feature extraction, considering the simultaneous existence of both dynamic and static characteristics in gas-liquid two-phase flow, a feature fusion method is introduced. This method integrates dynamic and static features, where dynamic features are extracted through slow feature analysis from a large set of dynamic variables, while static features use a small set of static variables directly. The feature-to-attribute prediction is accomplished by mining the association mapping between the state attributes and the extracted feature information, thereby enabling flow state identification in case of missing training samples. The effectiveness of the proposed method is verified through comparative experiments, and a better average recognition accuracy is achieved for three types of zero-shot cases than only using dynamic or static feature extraction methods.