Neural Ordinary Differential Equations for Data-Driven Regulation of Aerospace System

초록

This study presents a data-driven regulator for a rotary-wing Unmanned Aerial Vehicle (UAV) utilizing neural Ordinary Differential Equations (ODEs). Gradients essential for optimizing the loss function are computed through automatic differentiation, employing both forward and reverse modes with symbolic differentiation rules to enhance computational efficiency. The ODE model comprises two fully connected hidden layers. The adaptive moment estimation method is employed for updating the neural network. The UAV is equipped with a position and attitude autopilot system. A single scenario is utilized to train the neural network within the ODE framework, aiming to regulate the UAV's state. © 2024 15th Asia-Pacific International Symposium on Aerospace Technology, APISAT 2024. All rights reserved.

키워드