Sliding mode online learning for flight control applications in unmanned aerial systems

Implementing adaptive flight control strategies into unmanned aerial systems (UAS) contains a high potential to improve the degree of automation. This is especially the case regarding automatic operation under difficult atmospheric conditions or even system failures. A neural control strategy enables the UAS to improve its flight characteristics and to respond to unknown, non-linear flight conditions. Here, a learning flight control system for a fixed-wing UAS is realised using a systematic two-stage approach by firstly implementing a sustainable offline-trained basic knowledge and subsequently improving these characteristics during flight. Within the automated offline-step large groups of neural networks are trained with the required behaviour, which is derived from measured data. This phase shows that the necessary learning task can be achieved by multi-layered feedforward-networks. The training success of all networks is then evaluated with statistical methods and networks are selected for online application. The online learning step is realised with a control architecture comprising a neural network controller and a neural observer which predicts the system's dynamics and delivers the training signal for the contoller network. An important element of the control strategy is to determine a consistent error signal for the online training of the neural controller. This is done by backpropagation of a measured error through the inverse dynamics of the observer network. Since the inverse dynamics have to be very precise in order to train the controller adequately, a stable sliding mode control (SMC) algorithm for network training is introduced. This online adptive algorithm significantly improves the observer's charcteristics and with it the system's performance.