High-performance data analytics of hybrid rocket fuel combustion data using different machine learning approaches

Kurzfassung

Hybrid rocket fuels pose several advantages over conventional solid or liquid bi-propellant solutions in terms of safety, cost and thrust controllability. For paraffin-based fuels, the droplet entrainment process plays a major role in the combustion kinetics. This phenomenon is well reported in the literature, but dedicated quantitative analyses of optical measurements are still lacking. In this study, k-means++ clustering with different numbers of clusters and spectral clustering were employed to high-speed video data of four different combustion experiments with varying fuel and oxidizer configurations. The goal was to identify short-term turbulences and irregularities in the burning kinetics, which could further resolve the process of droplet entrainment. Our results show that k-means++ is able to identify main flow phases, but cannot resolve short-term structures, even with a large number of clusters k = 20. Spectral clustering, which is based on graph theory, requires the computation of an adjacency matrix, which becomes computationally expensive for large data sets. We implemented a highly parallel version of the algorithm, which allowed computation of the pairwise similarities on 30 000 video images in approximately 1 h (150 processes). The similarity matrix gives a qualitative assessment of the combustion kinetics, and several short-term irregularities could be identified. Full spectral clustering of the data yielded quantitative partitioning of the individual frames into long-term main components and short-term fluctuations. Results indicate that a fuel combination of paraffin with the addition of 5% polymer yields the most homogeneous combustion kinetics, and that the oxidizer flow has a substantial influence. The distributed implementation of the algorithm allows future investigations to be conducted on many experiments, giving more insight into the mechanisms of the combustion process.