Feature Selection and Virtual Sensing Based Mixture Ratio Estimation for Liquid Propellant Rocket Engine Control Systems

In the field of modern aerospace research, the control of rocket propulsion engines is steadily evolving towards intelligent closed-loop control strategies such as model-predictive and H∞ controllers. These approaches reduce component wear, enhance performance, and ensure the system stays within a safe operating range. Altogether this allows us to extend the lifespan of rocket engines and promote their reusability. However, the implementation of smart controllers in such complex and dynamic environments is challenging and requires accurate knowledge of critical engine parameters. Quantities that cannot be directly measured must be inferred from alternative sources, e.g., via process models or virtual sensing.

The propellant mixture ratio in the combustion chamber, which is crucial for engine optimization and performance, may serve as a prime example of such a quantity. It is essential to accurately measure propellants' mass flow rate to determine this ratio, typically using flow meters. However, weight and space constraints for this instrumentation make it unsuitable for in-flight use and limit its application to test beds. In this manuscript, we focus on a virtual sensing approach that exploits machine learning models trained with steady-state simulation data. We demonstrate how the mixture ratio can be estimated using a virtual sensing approach based on other easily available signals. To facilitate the adoption of our approach, we prioritize potential features via a machine learning procedure that assigns a relative importance score to each signal, supporting the domain expert’s final decision process. Utilizing the identified signals, we develop a regression model to precisely estimate the propellant mixture ratio. We train the model using both noise-free signals generated from simulation as well as signals with added noise. That approach better reflects the measurements encountered in actual in-flight scenarios.

The data-driven feature ranking allows the identification of input signals that positively impact the virtual sensing model's performance. As a result, the final machine learning model can accurately estimate the mixture ratio, demonstrating high precision in both noise-free simulation environments and noisy conditions, emulating real in-flight scenarios. Future research will be conducted to evaluate the model’s performance on actual rocket engine measurements, collected on testbenches, and to extend the estimation to transient conditions.