Learning control of unmanned aerial vehicles using artificial intelligence-based methods

Abstract

In recent years, many research activities have focused on the developments for unmanned aerial vehicles (UAVs) due to their usefulness in providing cost-effective solutions to dangerous, dirty and dull tasks. In many applications, it is crucial for UAVs to be able to fly autonomously in uncertain environments under variable operating conditions. In such circumstances, an intelligent capability of the flight controller is a must rather than a choice. Model-free controllers propose alternative solutions to the model-based controllers without requiring a precise system's model which is often either unavailable or time-consuming to obtain. One branch of model-free methods is composed by fuzzy logic controllers (FLCs) due to their capability of delivering excellent control in the presence of uncertainties. However, one weakness of FLCs is that their parameters have to be tuned to deal efficiently with uncertainties. On the other hand, neural networks are computing models which progressively improve their performance by learning from training examples. Hence, artificial neural networks (ANNs) and deep neural networks (DNNs) propose learning approaches to enhance control strategies. Nevertheless, the main disadvantage of neural networks is that their inner workings are difficult to interpret. The limitations of fuzzy logic and neural networks were a driving force behind the creation of hybrid systems where the combination of DNN and FLC can overcome the drawbacks of each individual method.

This thesis focuses on the aforementioned artificial intelligence-based control methods that enable UAVs to accurately track 3D trajectories. The investigation starts from the simplest static type-1 FLC, through interval type-2 FLC, to the most efficient novel fuzzy mapping-based controllers. In this thesis, it was demonstrated that the analytical representation of the fuzzy mapping facilitates the tuning of the parameters in FLCs. Next, the controllers based on ANNs and DNNs with learning capabilities were investigated. In this thesis, it was verified experimentally that the proposed approaches can improve real-time control performance. Finally, a novel deep fuzzy neural network framework which profoundly fuses DNN and FLC for online training was proposed and validated under a variety of operating conditions.