Research Of Modeling Fluid Dynamic Data Based On Deep Neural Networks

The research of Aerodynamic focuses on various power generated by the movement of objects in the air.As an important branch of Fluid Mechanics,it directly affects the progress and development of aerospace industries in various countries.In order to deal with the problem of excessive time consumption when acquiring aerodynamics data by traditional research methods,aerodynamic data modeling has become a hot research point in the field of aerodynamics.Currently,most methods of aerodynamic data modeling rely on aerodynamic mechanisms and are limited by the high complexity of aero-physical models.It is difficult for such methods to fundamentally reduce time costs.The method without physical model does not rely on the aerodynamic mechanism,and directly establishes the aerodynamic data mapping relationship to ensure a certain accuracy while greatly reducing the calculation cost.However,such existing research is mainly based on traditional machine learning techniques,and the accuracy of the obtained aerodynamic data is inadequate.Therefore,this thesis combines deep learning technologies such as cluster networks,multi-task learning,and mixture-of-expert networks,and proposes effective modeling methods for aerodynamic data.Comprehensive validations are executed on various aerodynamic datasets for verifying the effectiveness of proposed models.The main works of this thesis are summarized as follows:1.Regarding to the basic scenario,we follow the idea of divide-and-conquer,and design a cluster network based aerodynamic model.The model includes multiple sub-networks,and each sub-network automatically learns to implement different functions via training data,thus completing the effective modeling of aerodynamic data.Aiming at the problem that the entire model may be dominated by individual sub-networks during model training,a new loss function is proposed to improve the model training effect.At the same time,the training data is grouped in a pioneering way to help divide the functions of each sub-network,and finally obtain some better local models.A systematic comparison experiment with existing aerodynamic data modeling methods proves that proposed methods significantly improve model accuracy and calculation speed.2.For the relatively complex scenario which holds less training data,the multi-task learning based aerodynamic data modeling methods are proposed.Models based on multi-task learning have excellent model accuracy when tasks are highly relevant and stable.In order to further improve the performance of the model,the cluster network is used to improve the learning effect of multiple tasks sharing features.Thus,we propose an aerodynamic model combining multi-task learning and the cluster network.3.In order to solve the problem of model performance degradation by uneven task correlations,models based on the multi-gate mixture-of-expert network are designed to implement effective aerodynamic data modeling.This model adds a dedicated gating network for each task to ensure that different tasks utilize the expert network in the optimal way they need.However,due to the differences in the complexity of different tasks in aerodynamic data,the gate network has a negative impact on simple tasks,and retains a positive impact on complex tasks.Therefore,a residual gated network based on the idea of residual network is proposed to improve the multi-gate mixture-of-expert model.Comprehensive experiments show that the new model has excellent performance when facing simple and complex aerodynamic data modeling tasks.