Study On Transfer Learning Modeling And Interactive Visual Method Of Aerodynamic Data

Performance prediction is an important part of aircraft optimization design.The existing performance prediction methods from low to high accuracy include rapid engineering method,high precision numerical simulation,wind tunnel test and flight test.However,with the improvement of prediction accuracy,the cost of data acquisition is also increasing.Aerodynamic proxy model is an efficient means to obtain aerodynamic data.However,the reliability and prediction accuracy of the aerodynamic proxy model need a large number of high-precision data sets as support,and the acquisition cycle of high-precision data is long and the cost is high.To obtain sufficient high-precision data will inevitably consume a lot of computing time and resources.The variable confidence model can use a large number of low precision data sets for auxiliary modeling,thus reducing the model’s dependence on high precision data sets.According to the research,transfer learning is often used to solve the problem of multi-data source and multi-task.How to use transfer learning technology to build variable reliability aerodynamic model is the topic of this thesis.In this thesis,transfer learning modeling of aerodynamic data is studied,and a variable reliability aerodynamic modeling method based on transfer learning is proposed,and good prediction results are obtained.At the same time,the algorithm is encapsulated on the basis of this study,and an interactive and visual platform for aerodynamic modeling is constructed,which is convenient for researchers to apply the method in this thesis and manage aerodynamic data and models.Specific research contents are as follows:Firstly,aerodynamic modeling,variable-reliability model and transfer learning are summarized,and the application of transfer learning in variable-reliability model is established.In addition,according to the transfer learning classification perspective of learning methods,existing transfer learning methods for nonlinear fitting problems are selected,and experiments are designed to compare the results of three transfer learning methods,Tr Adaboost.R2,CDM and pre-training fine-tuning,on multi-source aerodynamic modeling,and screen suitable transfer learning algorithms.Secondly,based on the pre-training fine-tuning algorithm,the influence of different neural networks on transfer learning is compared horizontally.According to the above experimental results,a variable confidence aerodynamic modeling method based on transfer learning is proposed.The method combined the aerodynamic data fusion theory and transfer learning method,and designed the regression network structure based on Long Short-Term Memory neural network(LSTM).A parameter tuning mechanism combining pre-training fine-tuning and network layer freezing is used for migration training to obtain a highly reliable aerodynamic model.In order to verify the effectiveness of the algorithm,the XFLR calculation data(low precision)and wind tunnel test data(high precision)of NACA2414 airfoil are taken as research objects,and a large number of XFLR data are used to build a low-fidelity model as a pre-training model.The transfer learning modeling experiment of wind tunnel data with data volume from 1/2 to 1/10 is designed on the pre-training model.The Additive Scaling Function Based Multi-fidelity Surrogate Model(AS-MFS)is compared with the unmigrated LSTM model and Additive Scaling function based Multi-Fidelity Surrogate Model(AS-MFS).The experimental results show that the proposed method achieves higher prediction accuracy under all data quantities,and the prediction accuracy of drag and lift-drag ratio increases by7.22% and 8.85% on average compared with that before migration.Compared with AS-FMS,the average increase is 8.66% and 4.36%.Finally,based on the previous research work,a dynamic modeling interactive visual system is constructed.The system has the functions of data management,model construction,model prediction and visualization of prediction results.The system is tested in this thesis.The test results show that the system is easy to operate and can meet the requirements of interactive and visual aerodynamic modeling.