Time series identification, classification, and display methods for periodic flow patterns

Abstract

In the oxygen turbopump of a LOX/kerosene rocket engine, low-temperature gas-liquid flow and phase change processes occur, representing a typical case of direct contact condensation (DCC). DCC exhibits periodic flow characteristics under certain operating conditions, which are crucial for the normal operation of the rocket engine. This study aims to propose a temporal identification method for flow patterns with periodic characteristics and introduce a new method for displaying flow pattern diagrams to enhance the accuracy of flow pattern identification and description. Through experimental observations, the flow conditions are classified into three main flow patterns: crawling, chugging, and condensation oscillation. This study used You Only Look Once (YOLO) to extract single frame image features and Long Short-Term Memory (LSTM) to extract time series features. After evaluating the confidence of the model, the average accuracy of YOLO and LSTM were 99.33 % and 88.33 %, respectively. Furthermore, the flow pattern diagram proposed in this study effectively integrates temporal information of flow pattern variations with the proportions of different flow patterns. Based on the new flow pattern diagram, the variation rules of flow patterns under low mixing pressure and low mass flow rate were discussed. The trends of flow pattern proportion changes at non-operating points were effectively predicted, and the results align closely with traditional flow pattern diagrams. These findings provide an analytical basis for the mixing process of oxygen-rich gas and liquid oxygen in the inter-pump tubing of LOX/kerosene engines and can support future research.