Scalable Anomaly Detection in High-Speed Combustion Imaging

Abstract

High-speed imaging plays a crucial role in aerospace combustion research, providing detailed insights into transient phenomena such as turbulence-flame interactions and instability mechanisms critical for propulsion systems. However, the vast volume of high-speed video data exceeds single-node processing capacities, demanding scalable, parallel analysis solutions. To address this challenge, we leverage the Helmholtz Analytics Toolkit (Heat), an open-source Python library for distributed array computing, to enable scalable anomaly detection in high-speed combustion video data. Specifically, we developed a parallelized implementation of the Local Outlier Factor (LOF), a well-established density-based algorithm for detecting local (i.e., small-scale) anomalies, to support massively parallel execution across multiple GPUs. This allows the identification of critical instabilities and localized extinction events that are otherwise difficult to detect due to the immense data volume and complexity. The parallelized implementation of the LOF is applied to high-speed video data from hybrid rocket combustion experiments performed at the German Aerospace Center (DLR), where 120 000 images were processed. Our results illustrate that by leveraging Heat’s parallelized framework, memory and runtime constraints---which in the standard implementation of the LOF would require about 10 billion image comparisons---can be significantly mitigated. This demonstrates the potential of HPC-driven anomaly detection to advance aerospace research. This research was supported by the European Space Agency through the Open Space Innovation Platform https://ideas.esa.int as a Early Technology Development Agreement and carried out under the Discovery Program ESA Early Technology Development (Research Agreement No. 4000144045/24/NL/GLC/ov). Disclaimer: The view expressed in this publication can in no way be taken to reflect the official opinion of the European Space Agency.