Control-Oriented Aerodynamic Modeling with Automatic White-Box Structure Based on Dendrite Net

The establishment of an aerodynamic surrogate model is the basis of the optimization and control design for aerial vehicles. The fitting expression construction based on data relation relies on empirical studies to estimate polynomial structures, which is fast but troublesome in structure selection. The complexity of polynomial structure selection can be avoided by using large amounts of data to train the neural network with a large number of neurons, but the black box model of neural network is not conducive for optimization iteration due to its poor interpretability and large amounts of computation. In this study, a control-oriented aerodynamic modeling method with automatic white-box structure based on Dendrite Net is proposed to construct the explicit expressions with strong interpretability that does not depend on empirical studies. First, the Dendrite Net model is built to demonstrate that its prediction accuracy and training speed are better than the typical BP network. Second, by using the white-box property of Dendrite Net, the polynomial expression can be automatically constructed to predict aerodynamic performance based on the coefficient spectrum of the Dendrite Net. The experimental verification shows that the polynomial established in this paper is superior to the related method of artificial structural analysis, in terms of the accuracy of polynomial and the simplicity of expression construction. The proposed method has the advantages of fast computation, strong interpretability, accurate expression of results, and independent of empirical studies, which not only has the advantages of polynomial fitting and neural network fitting, but also overcomes their disadvantages.