Research On Deep Learning Method For Joint Modeling Of Aerodynamic Layout And Flight Status Features

Data-driven intelligent aerodynamics is an emerging interdisciplinary research.The aerodynamic performances of an aircraft are simultaneously influenced by flight status parameters and aerodynamic layout parameters.How to accurately obtain the aerody-namic performances of the aircraft using deep learning methods based on these two types of parameters is an important issue.When the aerodynamic layout parameters are fixed,numerical simulation,wind tunnel testing,and flight testing are commonly used to obtain dataset to train aerodynamic models.Using numerical simulation methods to calculate numerical solutions of nonlinear partial differential equations(PDEs)is time-consuming and hardware-consuming.Wind tunnel testing can be easily affected by tunnel walls and supports.Flight testing is difficult to cover all flight status.Consequently,aerodynamic data samples typically exhibit limitations and discrepancies in data modes,which presents two main challenges for deep learning research in aerodynamic modeling:1.Limited Data Issue:aerodynamic data samples are often scarce,but the range of them may be large,resulting in sparse aerodynamic samples.This further increases the difficulty of the aerodynamic model convergence;2.Data Mode Discrepancy Issue:due to significant discrepancies in aerodynamic performances corresponding to different flight status,exist-ing methods are easily influenced by other data modes when learning specific limit data modes.This leads to large prediction errors.In addition,when the aerodynamic status parameters are fixed,and the aerodynamic performances of the aircraft need to be deter-mined based on aerodynamic layout parameters,layout parameters are typically organized in two ways:one is parametrization features such as wingspan,chord length,sweep angle,or polynomial coefficients;the other is Euclidean coordinates.Thus,the third challenge is extracting geometric-features of complex shapes.When aircraft’s shape undergoes sub-tle variations,its aerodynamic performances may undergo significant changes.However,both of these Euclidean distance/vector-based description methods are inadequate for per-ceiving subtle changes in the aircraft’s surface structure,resulting in inaccurate predictions of the aircraft’s aerodynamic performances.Regarding the above challenges,the specific research content and innovations of this dissertation are as follows:1.To address the convergence challenges of deep learning models posed by the lim-ited and sparse data samples,this dissertation proposes an optimal approximation theorem for a Generative Adversarial Network(GAN)discriminator.The theorem demonstrates that when GAN is used to handle nonlinear,limited and sparse data samples,the Ra-dial Basis Function Neural Network(RBFNN)is the optimal form of the discriminator.Based on this theorem,this dissertation further proposes two novel generative models:Radial Basis Function-based GAN(RBF-GAN)and Radial Basis Function Cluster-based GAN(RBFC-GAN).These models are used to generate a large amount of performances with flight/flow status parameters,to expand the dataset and meet the demands of sub-sequent modeling.Experimental results show that,compared to GAN,the Mean Square Error(MSE)of data generated by RBF-GAN and RBFC-GAN is reduced by 65.33%and77.37%,respectively,and the stability is improved by 34.62%and 72.31%,respectively.2.In response to the challenge of existing modeling methods failing to reflect data mode discrepancy,this dissertation proposes a Multi-task Learning(MTL)model for data mode discrepancy adaptive modeling.MTL categorizes flight/flow status data modes within the dataset into multiple subtasks,and then employ a Cluster Net to learn the nonlin-ear mapping contained in each subtask in a distributed way.Experimental results demon-strate that this model can reduce mutual interference between different data modes during the learning process.It results in a 1.85%decrease in MSE for predicted velocity in non-steady flow fields and an 35.11%decrease in MSE for predicted pressure in steady flow fields.3.In response to the challenge of Euclidean space-based description methods fail-ing to capture complex aircraft layout features,this dissertation places the aircraft shape into a Riemannian space.This dissertation proves a theorem that the approximation of aircraft shape using Bézier curves/surfaces can construct a segmented smooth topological manifolds.Based on this theorem,this dissertation further extract manifold features with clear mathematical meanings as layout features of aircraft.Experimental results show that,compared to coordinate data,the reconstructed airfoils with manifold features as inputs are smoother,and the MSE of reconstructed airfoils decreased by 41.30%.4.Based on the theorem proposed in 3,and the discrepant data learning mechanism proposed in 2,a method called Riemannian manifold feature and Discrepant Data-based Fusion Method(MDF)is further proposed to conduct joint learning research focusing on flight status features and aerodynamic layout features.Experimental results show that,compared with state-of-the-art method,MDF can reduce the predicted MSE of the coeffi-cient of drag(C_D)in UIUC airfoil dataset by 54.56%.For DLR-F11 aircraft,the predicted MSE for the wing surface coefficient of pressure(C_P)decreased by 55.54%.The significance of this dissertation is as follows:1.Integrating aerodynamic modeling methods with artificial intelligence explores new methodologies in intelligent aerodynamics studies.2.Through studies of data generation for flight status parameters and adaptive mod-eling for data mode discrepancy,this study enables fast and accurate prediction of aero-dynamic performances concerning limited flight status data mode,which significantly reduces the risk in flight testing and shortens the iterative design process for aircrafts.3.Conducting manifold feature extraction in Riemannian space for aircraft with complex shapes enables the fast acquisition of manifold features with clear mathemati-cal meanings,which can accurately describe the geometric structures of aircrafts.This presents a feasible approach for aircraft reverse design and aerodynamic layout optimiza-tion.4.The developed joint modeling approach for aerodynamic layout features and flight status features obtains higher-precision predictions of aircraft aerodynamic performances.This meets various demands in aircraft design,numerical flight simulation,combat sim-ulations,etc.It also supports accurate assessment of next-generation aircraft’s flight per-formances and fast improvements/modifications of aircrafts in critical defense research areas.