Aircraft dynamics modeling at high angle of attack incorporating residual transformer autoencoder and physical mechanisms

Abstract

Aircraft dynamics modeling is an important part of advanced control law design and flight safety. To address the challenge of longitudinal dynamics modeling using a small amount of flight test data under high-angle-of-attack (AOA) conditions, we propose an effective approach that integrates machine learning with physical mechanisms. First, a low-fidelity aircraft dynamics model based on physical analysis is established. Second, a Residual Transformer (ResTrans) autoencoder is designed to extract temporal and spatial features from flight motion history under high-AOA conditions. These features are then used to compensate for the modeling errors of the low-fidelity model through a deep neural network (DNN)-based fusion module, resulting in a high-fidelity aircraft dynamics model. Moreover, a physics-informed closed-loop multi-step dynamics evolution (PI-CMDE) paradigm is developed for constructing loss functions, ensuring stable and efficient parameter optimization of the high-fidelity model. Finally, a simulation model of a scaled F-16 aircraft is used to generate a small set of high-AOA flight test data for training and testing the high-fidelity model. Experimental results demonstrate that, compared to three representative aircraft dynamics modeling baseline methods, the proposed approach achieves higher modeling accuracy and better generalization performance, highlighting its advanced capabilities.