Entwicklung einer Triebwerksregelung für ein LOX/LNG Triebwerk mit elektrischen Pumpen

Dies ist die aktuellste Version dieses Eintrags.

Kurzfassung

In this master thesis, a hybrid control system combining concepts from classical control theory and model-free reinforcement learning is developed and evaluated. Advancements in the space industry, driven by the so-called new space boom, are introducing new challenges for modern launch vehicles. In particular, the propulsion subsystem of future rockets must operate for longer durations, offer higher reliability, and integrate into increasingly complex hierarchical control structures. These new requirements demand advanced engine control systems capable of solving complex multivariate control problems such as deep throttling and life-extending control strategies. The German Aerospace Center at the Institute of Space Propulsion in Lampoldshausen is researching and testing novel control approaches to solve some of these problems by the means of intelligent control. In particular, the usage of reinforcement learning-based control is being investigated due to its capability of solving complex multivariate control problems for nonlinear systems such as rocket engines. While the applicability of such control structures of has been previously proven, the black box character of the resulting control laws remains a challenge for the adoption of reinforcement learning-based control. Building upon previous research performed at the Institute of Space Propulsion, this thesis develops a hybrid control structure that integrates a reinforcement learning-based control law with two static proportional-integral (PI) controllers. The question is posed how such a system compares to a classical control theory-based approach when solving a trajectory tracking problem. In addition, it is investigated how the robustness properties of such a hybrid control system relate to the properties of the static PI control system. In total, three control systems are developed. The investigated hybrid control system, a pure reinforcement learning-based control system and a static PI control system. The pure reinforcement learning-based control system and the PI control system serve as baselines to compare the hybrid control system against. The control systems are investigated in both their ability to solve a trajectory tracking problem and their robustness against model uncertainties. The investigations are performed in simulation using a numerical model of an example cryogenic methane (LNG) and liquid oxygen electric pump cycle rocket engine, developed by The Exploration Company. The simulation is implemented in Ecosim, an industry standard simulation tool used in the design process of rocket propulsion system. The results demonstrate that the hybrid control system learns a form of automatic gain scheduling. Using this, the control system presents superior performance over both the pure reinforcement learning-based control system and a static PI control system. The control actions remain interpretable, alleviating the problem of opaque black box control laws learned by a reinforcement learning system. The developed control presents a promising approach to combine the ability of reinforcement learning-based control systems in solving complex multivariate control problems for nonlinear system with the interpretability and analysability of a classical control system to increase the confidence in the control systems performance.

Verfügbare Versionen dieses Eintrags

• Entwicklung einer Triebwerksregelung für ein LOX/LNG Triebwerk mit elektrischen Pumpen. (deposited 24 Jun 2025 12:53) [Gegenwärtig angezeigt]