Deep Convolutional Identifier for Dynamic Modeling and Adaptive Control of Unmanned Helicopter.
Yu Kang,Shaofeng Chen,Xuefeng Wang,Yang Cao
DOI: https://doi.org/10.1109/TNNLS.2018.2844173
IF: 14.255
2019-01-01
IEEE Transactions on Neural Networks and Learning Systems
Abstract:Helicopters are complex high-order and time-varying nonlinear systems, strongly coupling with aerodynamic forces, engine dynamics, and other phenomena. Therefore, it is a great challenge to investigate system identification for dynamic modeling and adaptive control for helicopters. In this paper, we address the system identification problem as dynamic regression and propose to represent the uncertainties and the hidden states in the system dynamic model with a deep convolutional neural network. Particularly, the parameters of the network are directly learned from the real flight data of aerobatic helicopter. Since the deep convolutional model has a good performance for describing the dynamic behavior of the hidden states and uncertainties in the flight process, the proposed identifier manifests strong robustness and high accuracy, even for untrained aerobatic maneuvers. The effectiveness of the proposed method is verified by various experiments with the real-world flight data from the Stanford Autonomous Helicopter Project. Consequently, an adaptive flight control scheme including a deep convolutional identifier and a backstepping-based controller is presented. The stability of the flight control scheme is rigorously proved by the Lyapunov theory. It reveals that the tracking errors for both the position and attitude of unmanned helicopter asymptotic converge to a small neighborhood of the origin.