Fruitful collaboration between the DLR aerodynamicists and our Helmholtz AI consultant team for Aeronautics, Space, and Transportation.
One of the simplest and oldest ways to visualize fluid flow in experiments is the application of tufts.Tufts are small pieces of rope/wire that align themselves with the local flow direction. In regions of separated flow, the tufts typically display a very unsteady behavior. At DLR, Institute of Aerodynamicsand Flow Technology tufts or flow cones are used in both flight test and wind tunnel experiments (fixed wing and helicopter) to determine the local flow on a surface to improve the understanding of the flow and provide validation data for numerical models. The tufts/flow cones are filmed by a camera (either video or single shot). Afterwards the images are processed in order to obtain the local flow direction in terms of angle, shape, temporal behavior, etc. of the tuft. In order to evaluate the temporal behavior of a tuft it is paramount to compare the status in one image to the status of the same tuft in another image.
In this paper, the DLR researchers developed an AI system to automatize fluid flow visualizations from in-flight images with tufts. In contrast to manual analysis by humans, the developed AI system enables automation, adding the benefit of flexibility, reproducibility, lack of human bias, and scalability to vast number of images and tufts per image. When compared to traditional image processing techniques, the developed system is able to cope with the issues that arise outside of a controlled environment, such as large changes in lighting conditions, perspective and background scenes.
The project was funded by Helmholtz AI HAICU voucher: “Automated Tuft Recognition For Flow Visualization”. The consultant was from the DLR Institute of Robotics and Mechatronics located at Oberpfaffenhofen, while the researchers were from the DLR Institute of Aerodynamics and Flow Technology, located at Braunschweig. Despite the physical distance and difference in research areas, the team fruitfully collaborated, thanks to the platform provided by the HAICU.
Some of the key aspects of the paper:
Development of an AI system to learn fluid flow visualizations from in-flight images with tufts via semantic segmentation. As a result, real-world demonstrations of the proposed concept are provided, for the first time to our knowledge.
Uncertainty-driven techniques to perform the semantic segmentation without requiring any annotations of segmentation masks. The system only needs the labelling of few bounding boxes, and only a single image where all the semantic class labels are specified.
Application to data-sets from manned helicopter flights as well as a UAVs flying in the stratosphere (higher than 20km altitude). In particular, the manned helicopter data was collected by the researchers at the DLR Institute of Aerodynamics and Flow Technology.