Finite-Memory-Structured Online Training Algorithm for System Identification of Unmanned Aerial Vehicles With Neural Networks
Hyun Ho Kang, Dong Kyu Lee, Choon Ki Ahn
Abstract
In this article, we propose a novel finite-memory-structured online training algorithm (FiMos-TA) for neural networks to identify and predict the unknown functions and states of an unmanned aerial vehicle (UAV). The proposed FiMos-TA is designed based on a system reconstructed by accumulating the states from the UAV dynamics. The system is redefined by replacing the unknown nonlinear functions of the UAV with neural networks, and a random walk modeling is adopted to design a training algorithm. The proposed FiMos-TA with a finite memory structure updates the weights of the neural network by accumulating the refined measurements of a UAV on the receding horizon. The training law of the proposed FiMos-TA is obtained by introducing the Frobenius norm and confirms a robust performance against modeling uncertainties and identification errors. The robustness and accuracy of the proposed FiMos-TA are verified through experiments.