DGLR-Publikationsdatenbank - Detailansicht
M. Winter, C. Breitsamter
In the present work, a Monte-Carlo-based aerodynamic reduced-order modeling process is developed to estimate statistical errors caused by the random training data segmentation. The reduced-order models (ROMs) considered here are constructed by means of a linear or nonlinear system identification. Therefore, training, validation, and test datasets provided by a computational fluid dynamics (CFD) solver are exploited. However, system identification tasks always involve parameter optimization and function fitting problems that are sensitive to the choice of the initial parameters or the training data composition, respectively. Consequently, an unfavorable random starting point may lead to a poor ROM performance. A remedy to overcome those model uncertainties is the application of a Monte-Carlo training and application strategy. To assess the effectiveness of the proposed ROM framework, the procedure is demonstrated by modeling the unsteady transonic aerodynamics of the pitching and plunging NLR 7301 airfoil. Various ROM techniques are trained and applied within the Monte-Carlo framework to show their simulation capabilities compared to the respective full-order CFD solution. The focus is particularly laid on the evaluation of the ROM solutions fluctuation due to different random initializations. It is shown that some ROM approaches exhibit a very good agreement combined with a low sensitivity to the training data partitioning, which is highly beneficial for a reliable and accurate application.
Deutscher Luft- und Raumfahrtkongress 2017, München
Deutsche Gesellschaft für Luft- und Raumfahrt - Lilienthal-Oberth e.V., Bonn, 2017
21,0 x 29,7 cm, 9 Seiten
Stichworte zum Inhalt:
Reduced-Order Modeling, Unsteady Aerodynamics