基于深度自动编码器神经网络的飞行器翼型参数降维与优化设计

The traditional parametric description methods of aircraft airfoil have lead to low optimization efficiency and heavy calculation workload because of large amount of variables in optimization design. A neural network model based on deep autoencoder(DAE) is proposed to solve the dimensionality reduction of description parameters in airfoil optimization design. The physical meaning of each parameter output by the model is analyzed,the dimensionality reduction effect of the model on airfoil description parameters is compared with that of the proper orthogonal decomposition (POD) method. Under the given design objectives and constraints, an optimization design framework based on surrogate model and genetic algorithm is used for RAE2822 airfoil optimization in the transonic flow. The airfoil optimization design effects of the proposed model,Class Shape Function Transformation(CST) method and POD method are compared,which proves that the proposed method with neural network based on DAE has higher optimization efficiency,and it performs obviously better than both CST and POD methods in drag reduction design of RAE2822 in transonic flow.