Unmanned aerial vehicles (UAVs) have the potential for use in a number of high-precision applications, including industrial inspection and precision agriculture. Quadrotor UAVs are one of the most commonly used UAV platforms, as they are agile, lightweight and relatively inexpensive. A significant limitation of current quadrotor UAV flight control systems is their inability to adapt to changing flight conditions, as well as their inflexibility to changing airframe properties. Conventional methods utilise linearised or simplified models of flight dynamics which have limited capacity to respond to changing system parameters, and use control methods that are limited to low accuracy applications or to specific flight conditions. The development of a generalised, flexible flight control system for quadrotor UAVs would allow for a greater range of applications across a wide range of airframes, from specialist to civilian UAVs. This work investigates the development of such a flight control system.

Firstly, a nonlinear dynamics model of a quadrotor UAV was developed. This dynamics model used a quaternion orientation representation; the majority of existing models use an Euler angle representation, which have known singularities, whereas using a quaternion representation allows for singularity-free orientation representation as well as more efficient computation. The dynamics model was developed in a minimal sense, with the core dynamics effects derived from first principles. The model was then parametrised into 13 parameters, using a lumped parameter approach as the initial model was structurally non-identifiable. Unmodelled non-ideal dynamics effects were included by augmenting each dynamic axis with a parameter representing the total dynamic impact of unmodelled effects on that axis.

The dynamics model developed includes aerodynamic drag effects modelled using the Rayleigh drag equation. There is little consensus in the literature on how best to model aerodynamic drag for a multirotor UAV, so an experimental study of the drag behaviour of a quadrotor airframe was conducted. A quadrotor airframe was mounted at a range of pitch and yaw angles in a wind tunnel, and the drag forces on the airframe were measured across a range of wind speeds. The results of this study showed that a Rayleigh drag model accurately models the UAV’s aerodynamic drag behaviour.

An experimental flight platform was then developed to allow for physical data collection. This platform consisted of a standard cross configuration quadrotor airfarme, with Hall effect sensors added to measure the angular velocities of the propellers. The Hall effect sensors and inertial measurement unit used were calibrated against known reference values, using test fixtures designed for each sensor, and their signal-to-noise ratios (SNRs) were derived, with all sensors having SNRs greater than 25 dB. Ground truth flight data was obtained using a motion capture system, and required the implementation of synchronisation between the onboard sensor clock and the external motion capture clock. A modified form of the IEEE 802.11 clock synchronisation protocol was implemented, with an infrared LED used to obtain a 20 Hz synchronisation pulse with the motion capture system.

Next, identification of the model’s parameters was investigated. Parameter identification methods for UAV systems in the literature are typically applied prior to flight operations, whereas this thesis considers the use of such methods during flight operations to allow for online tuning and adaptation to changing flight conditions. Four common white-box parameter identification methods were implemented and were applied to both synthetic data and experimental flight data in an open-loop control configuration. Error metrics for the performance of each method were defined, including a definition of pose error that captured both the overall desired behaviour of the airframe, as well as including the quaternion orientation representation. Application of the parameter identification methods to the synthetic data showed that Levenberg-Marquardt gradient descent had the best performance, including when synthetic noise was added to the simulated data, with this method yielding pose errors 75% smaller than the other methods at SNRs down to 20 dB. Levenberg-Marquardt gradient descent was also the most performant method on the experimental flight data, with median quaternion distance errors below 0.3 rad, but maximum distance errors comparable to the experimental flight envelope.

Finally, a nonlinear model predictive control system that incorporated both the nonlinear dynamics model and online parameter identification was developed. A novel cost function based upon the pose error metric was used to prioritise trajectory tracking, with comparison with a standard quadratic cost function showing that the pose error cost function gave better tracking performance and was better able to track target quaternion orientations. The controller was tested on synthetic flight trajectories, including airframe parameter changes and external wind disturbances, and was shown to maintain accurate trajectory tracking when subjected to these conditions, although the controller’s high computation cost prevented its use in an experimental setting.