Ai-based evaluation of hybrid rocket combustion tests

Abstract

High-speed camera recordings of rocket combustion tests were analyzed using deep learning. In hybrid rocket engines thrust is generated via the chemical reaction of a solid fuel and a fluid oxidizer, as such they can combine some of the advantages of solid and liquid rockets. One important metric in the design of hybrid rocket engines is the regression rate, which is the rate with which the surface of the solid fuel recedes during combustion. The regression rate depends on many factors, such as the makeup of the fuel, the flowrate of the oxidizer and the pressure in the combustion chamber. In order to investigate these dependencies, several combustion tests were performed at the German Aerospace Center and recorded with a high-speed camera. The shape of the remaining fuel is detected in each frame and the regression rate is computed as the rate of change of the height of the fuel. Detecting the shape of the fuel using classical computer vision algorithms is impractical because the optical access to the combustion chamber may be partially obstructed by soot, reflections or flames forming at the side of the fuel. Therefore, a neural network with U-net architecture is trained to detect the shape of the fuel. This training requires manual labelling, which can only be done for a small fraction of the frames of a given high-speed recording. Hence, several forms of data augmentation are employed to avoid overfitting and improve robustness with respect to the prominent sources of noise in the data.