Integrating model-based and data-driven simulator for health management of a reusable rocket engine

Abstract

There is considerable need to generate large amounts of training data that include various operating conditions for fault detection and diagnosis with machine learning in a reusable rocket engine. A system-level simulation model has been developed in which reduced-order models are employed to simulate the global behavior of a reusable rocket engine. Although some components of the engine are not modeled at the system level due to their complexity, they are included among the items inspected during fault detection and diagnosis. This study has developed a regression model for simulating the behavior of such components based on the results of static-firing tests on a reusable rocket engine developed in Japan. The regression model used measurements that can be modeled by the system-level simulation and treated the ones that cannot be modeled as response variables. To identify the operating conditions, the explanatory variables are divided using the Gaussian mixture model in advance, and the Ridge regression models are then trained from the clustered explanatory variables on each cluster. This method reasonably predicts the response variables, even if the static-firing testing includes varying operating conditions such as the combustion phase with varied throttling and the chill-down phase.