Nonlinear System Identification for Quadrotors with Neural Ordinary Differential Equations
Mingqian Wang,Jianshan Zhou,Xuting Duan,Dezong Zhao,Pinlong Cai,Junda Zhai,Xinyu Liu,Chenghao Ren
DOI: https://doi.org/10.1109/ICUS58632.2023.10318403
2023-01-01
Abstract:Quadrotors have been widely used in various civilian fields, such as traffic monitoring, agriculture and crop monitoring, and delivery services, among others. Quadrotors exhibit highly complex flight dynamics, characterized by inherent instability, strong nonlinearity, and significant coupling. Representing their flight dynamics necessitates the use of multiple input multiple output system models. Disturbances affecting one axis can readily propagate to other degrees of freedom, potentially resulting in performance degradation or even destabilization. Up to now, modeling the dynamics system forms the foundation for studying quadrotor control. Traditionally, dynamics system models have been derived from first principles and physical insights. However, the dynamics governing these quadrotors are known to be highly nonlinear and time-varying, making it challenging to determine the control parameters using conventional models. To address this, system identification, a data-driven method that utilizes flight experiment data, demonstrates good performance in modeling the dynamics system of quadrotors. This data-driven approach extracts informative features from the data to develop system models. In this paper, we propose a system identification method using neural ordinary differential equations. The results indicate that our method achieves high accuracy and performs well in nonlinear system identification.