Wang Zihao1 ,Li Gun1 Liu Dawei2 Chen Dehua1, 2 , Zhang Shujun1
1.University of Electronic Science and Technology of China, School of Aerospace Engineering, Chengdu 611731
2.China Aerodynamics Research and Development Center, High-Speed Aerodynamics Institute, Mianyang 621000
doi: 10.7638/kqdlxxb-2023.0146
The flutter of supercritical wings greatly impacts the safety and stability of transport aircraft. Accurately and efficiently determining the flutter boundary has always been a research hotspot. In current flutter research, constructing unsteady reduced-order models is an important analytical method. With the development of artificial intelligence, the use of neural networks as reduced-order models has been proposed by more and more scholars. However, existing neural network-based reduced-order models require a large amount of computational aerodynamic data as training sets, while using RANS equations to compute unsteady aerodynamic data requires a significant amount of time. How to obtain a more accurate flutter boundary with a limited amount of computational data is a question worth studying.
This paper constructs a flutter boundary prediction framework for supercritical wings based on Long Short-Term Memory (LSTM) neural networks: Using RANS equations to compute a small amount of steady aerodynamic data, the increase in angle of attack at a given Mach number is abstracted as a change in LSTM time series, predicting the aerodynamic coefficients at other angles of attack under this Mach number, and then obtaining the initial angle of attack for flutter based on the variation of the aerodynamic coefficient curve, thereby determining the flutter boundary.
For the CHN-T1 model, based on its computational data, a predictive model for aerodynamic coefficients and a model for determining the initial angle of attack for flutter based on LSTM were designed to accurately predict the trend of aerodynamic coefficients at a given Mach number, and achieve rapid determination of the initial angle of attack for flutter; by integrating the initial angle of attack data, the flutter boundary for the CHN-T1 model was determined, and the accuracy of the results was verified with wind tunnel test data.
1) The LSTM model has good predictive ability for the trend of aerodynamic coefficient changes, with a root mean square error maintained within 2%; at the same time, it performs excellently in determining the initial angle of attack for flutter, with the error of the flutter boundary kept within 2%.
Figure 1 LSTM Unit Structure Diagram
2) Verification results with different training-test ratios indicate that the proposed method has generalization capability, can overcome a certain degree of interference, and establishes a robust flutter boundary prediction model. At the same time, different training-test ratios can affect the prediction and determination accuracy of the model, so convergence analysis needs to be conducted before modeling to ensure the generalization capability of the model.
Figure 2 Comparison of Flutter Boundaries for CHN-T1 Model Obtained by Different Methods
The current flutter boundary prediction methods use training and testing data sourced from high-precision data solved by the N-S method, while wind tunnel test data is mainly used to validate the accuracy of the flutter boundary. There has been no multi-source intelligent fusion research of experimental data and computational data. Future work will further improve the model’s generalization capability and prediction accuracy, and it is also planned to apply this model to aerodynamic data prediction for various wings to further expand its universality.
(The sources of the above-related figures can be found in the original text)
Editing and typesetting: Li Lu
Click the link below to view historical articles
Research Paper | Study on Supercritical Airfoil Design of Integrated Body Layout
Research Paper | Aerodynamic Model CHN-T2 Design for Wide-Body Aircraft
Column | Design and Database Application of CAE-AVM Model Cruise Configuration
Column | Model-Free Adaptive Control of Shock-Induced Flutter of Airfoils
Column | Numerical Simulation of CHN-T1 Model Based on Nested Grids
▼▼▼Click below “Read Original” to further view the details of this article, and feel free to cite it.
