Icing prediction method for arbitrary airfoil using deep neural networks

The icing of the aircraft will affect the aerodynamic performance and is the main factor affecting flight safety. Accurate prediction of ice shape can provide strong support for anti-icing work and is of great significance for ensuring flight safety. The traditional numerical icing research methods can hardly meet the requirements of ice accumulation evaluation under multiple icing conditions. Neural network methods provide a robust way for the ice prediction task. The current machine-learning prediction model for airfoil icing can only predict the icing shape of a specific airfoil or a class of airfoils and does not have the universality of icing prediction for a general airfoil. To solve this problem, a deep neural network-based icing prediction method for a general airfoil is proposed, which is suitable for low-speed incompressible flow. The method uses the airfoil pressure coefficient to abstract the characteristics of airfoils, combines the parameters of the flow field and the cloud field as the input, and uses the Fourier series fitting coefficient of the two-dimensional ice curve as the output. By this means, a prediction model using a deep neural network is established, and the icing prediction task of a general airfoil is preliminarily realized. The experimental results of various examples show that the proposed method has a good ice shape prediction effect for a single airfoil or a general airfoil, and the relative error of the main characteristic parameters of the ice shape prediction is not more than 15%.