Improved performance of data-driven simulator of liquid rocket engine under varying operating conditions

Abstract

Integrating a model-based and data-driven approach to generate sufficient number of training data that include time-varying operating conditions is needed to develop a prognostics and health management technique using machine learning for a reusable rocket engine. A data-driven approach was used to predict the engine behavior, which cannot be determined using a model-based approach. This paper proposes a clustered regression model to predict various operating conditions for a reusable rocket engine. The training data were divided in advance according to the operating conditions using the Gaussian mixture model, and multiple regression models were individually trained on each cluster. By switching the regression models according to the operating conditions, the time-series data comprising various operating conditions, such as the combustion and chill-down phases and transition between phases, were accurately predicted when compared with the prediction results of a single regression model. The prediction results of the present hybrid approach agree reasonably well with the static firing test results of a reusable rocket engine if the explanatory variables are adequately included. By integrating model-based and data-driven approaches, adequate training data can be generated for prognostics and health management of the reusable liquid rocket engine.