Abstract:

This paper presents the design and implementation of a machine learning-based framework aimed at enhancing the numerical modeling of nonlinear flight dynamics in unmanned aerial vehicles (UAVs). Traditional numerical solvers, though reliable, face challenges in real-time computation and adaptability to nonlinear aerodynamic variations. The proposed hybrid framework integrates data-driven learning components with classical numerical analysis to improve both computational efficiency and modeling precision. Using hybrid datasets from simulations and experimental flights, Long Short-Term Memory (LSTM) and Transformer architectures were developed as numerical surrogates for nonlinear state estimation. The results demonstrate that the proposed machine learning-enhanced framework outperforms the conventional fourth-order Runge–Kutta (RK4) method in terms of accuracy, adaptability, and computation time, achieving up to a tenfold improvement in inference speed. Furthermore, the model shows robustness under disturbances and physical interpretability through feature attribution analyses. This study represents a crucial intermediate phase in developing a comprehensive machine learning–driven numerical environment for nonlinear UAV flight dynamics analysis and control.