Estimating unknown dynamics of a quadrotor aerial vehicle artificial neural networks

Atapattu, Sachithra H. (2020) Estimating unknown dynamics of a quadrotor aerial vehicle artificial neural networks. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (4MB)

Abstract

This thesis presents the use of artificial neural networks (ANN) to accurately estimate the unmodelled dynamics of a quadrotor aerial vehicle. Due to the complexity of aerodynamic models, ground effects, actuator non-linearities, and disturbances that occur during operation, usually a complete dynamic modelling of a quadcopter is difficult to achieve. A quadcopter’s flight stability mainly depends on its control system and its navigation system. Therefore, the knowledge of the platform’s unknown dynamic model will allow enhanced design of the autonomous system. This thesis focuses on estimating the unmodelled dynamics using machine learning (ML) techniques to achieve better design of control and navigation systems of quadcopters. In this work, the dynamic behaviour of a quadcopter is expressed as a combination of a nominal model and an unknown model. The nominal model is derived from classical rigid-body dynamic equations, while ML approaches are used to estimate the unknown model. Among the available ML techniques, two commonly used approaches; Gaussian process regression (GPR) and ANN are evaluated in this thesis. Training and testing of the methods were done offline, using a dataset gathered by manually flying an AscTec Hummingbird quadcopter in an OptiTrack motion capture environment. For the comparison of the selected ML techniques, the disturbance force was learnt initially. Since ANN showed better results than GPR in initial comparison, ANN was chosen and further evaluated with different architectures. The disturbances in force and torque were learnt separately using two independent ANNs. These estimated residual dynamics were added to the nominal dynamic model to obtain the compound dynamics. The results demonstrate reduction in model error in comparison to the nominal model. To demonstrate the use of learnt residual dynamics in improved controller design, a trajectory tracking controller was implemented in a simulated environment. This simulation was designed to track a desired trajectory in the presence of added disturbances estimated using the ANN model. The results are compared with the trajectory tracking results obtained with only the nominal model. As expected, the controller with disturbance correction using ANN model showed improved performance.

Actions (login required)

View Item