Analysis of Nonlinear Learning Methods in the Adaptive Control Context of an Unmanned Helicopter

Kurzfassung

One research focus in DLR’s (German Aerospace Centre) Institute of Flight Systems is generation of artificial neural network benchmark for system identification uses. This thesis presents intermediate results which were obtained within forementioned research focus. The research will continue after the submission of this thesis. In this work, available literature were reviewed and reference criteria were produced, to be used for selection of suitable machine learning algorithms. With help of the reference criteria, multilayer feed forward neural networks (FFNN) and echo state networks (ESN) were selected for system identification use. wo different configurations of FFNN were applied to modelling of an unmanned helicopter dynamics. Lastly, approaches for capturing model uncertainties with FFNN were explained and it was shown that an existing flight dynamics model could be improved with a neural network correcting it. Finding optimal neural network structures for these different identification tasks, and providing performance comparison between FFNN and ESN were left as future work.