Prediction of Lift and Drag Coefficients for Aircrafts Based on CNN-ATT

The accurate establishment of the aerodynamic model of an aircraft is the basis for subsequent research on various aspects of the aircraft. Deep learning, with its powerful ability to fit nonlinear problems, has gradually become the focus of research on various difficult problems. In this paper, we propose a method for predicting the lift and drag coefficient of an aircraft based on a convolutional neural network. Our dataset is different from most wing aerodynamic parameter prediction datasets. Our dataset is unique in that it incorporates Mach and angle of attack information into images of the overall aircraft, with the lift coefficient and drag coefficient as the outputs. Our neural network is distinct from others in that it utilizes the final fully connected layer as the output layer for direct prediction of lift and drag coefficients in the regression task. Furthermore, we have incorporated an attention mechanism into our model to enhance its generalization ability and prevent it from being affected by noise or irrelevant information. We refer to this network as convolutional neural network with Attention mechanism(CNN-ATT). Our experiments demonstrate that our method outper-forms traditional approaches using vectors composed of aircraft structure and other information as inputs, as well as methods employing BP and LSTM networks, in terms of both information representation capability and prediction accuracy. Specifically, we have identified algorithmic and technical improvements that contribute to the superior performance of our model. We have also considered other factors that may affect model performance.