关键词: 多源数据融合, 涡格法, 飞机概念设计, 径向基神经网络, Lambda翼无人机

Abstract: The conceptual design phase of aircraft requires extensive analysis and comparison of the aerodynamic performance for numerous configuration candidates. High-fidelity methods like Computational Fluid Dynamics (CFD) are computationally prohibitive. Rapid analysis tools like the Vortex Lattice Method (VLM) often fail to meet accuracy requirements. Single-Fidelity Data-Driven surrogate models, which directly map geometric design variables to aerodynamic characteristics, suffer from prediction accuracy that is heavily dependent on the size and selection of training samples. To address the trade-off between computational accuracy, efficiency, and sample size, this study proposes a rapid aerodynamic prediction method that fuses multi-fidelity data from VLM and CFD simulations. Based on a Radial Basis Function Neural Network (RBFNN), the framework employs VLM to obtain low-fidelity aerodynamic parameters efficiently. These parameters, together with the original design variables, Mach number, and wetted surface area, form the enhanced input vector to improve the model's understanding of aerodynamic. The network is trained to predict the discrepancy between the multi-fidelity data. This approach significantly reduces the required number of high-fidelity training samples and improves computational efficiency. The method is validated using a Lambda-Wing UAV dataset. The results show that with 25 training samples, the UAV’s high-fidelity aerodynamic parameters can be accurately predicted. The prediction accuracy for all parameters outperforms that of single-source data-driven surrogate models, and is comparable to the co-Kriging method—with certain parameters even demonstrating superior prediction performance. Further validation against three typical engineering scenarios confirms the method's general applicability. The relative prediction error for aerodynamic coefficients remains within 3%, while the computational time is reduced by two orders of magnitude. The fusion method is applied to the optimization of the Lambda-wing UAV scheme. The cruise lift-to-drag ratio of the optimal scheme is increased by 11.6%, and the predicted relative error was less than 1.2%. This effectively enhanced the design efficiency and holds certain prospects for engineering application.

Key words: multi-fidelity data-fusion, VLM, aircraft conceptual design, RBFNN, Lambda-Wing UAV