Zur Anwendung neuronaler Netzwerke in adaptiven Flugregelungssystemen

The present work makes a contribution to the systematic integration of artificial neural networks into adaptive control systems. Besides the basic principles of mapping non-linear dynamic processes using neural networks, a structured methodology for the network design is illustrated, which allows a statistical evaluation of the learning success. In addition, to allow for an improved transparency of neural networks used in control systems a paradigm change regarding their learning methods is undertaken. This means that a neural network is treated as a dynamic system whereby the learning within such a network becomes a controlled process. Therefore, the application of the concept of sliding mode control to learning techniques of neural networks is possible, which thereby can be augmented with an intrinsic stability condition. Through the dynamic calculation of the learning rate this learning approach is able to constrain the network’s states into the globally asymptotically stable sliding mode. The practical relevance of this learning method is illustrated utilising two dissimilar control strategies: the so-called indirect adaptive neuro-control, which does not need any mathematical model of the controlled system, and the nonlinear dynamic inversion, for which a high degree of knowledge of the system dynamics is essential. The results of both control methods underline the profound as well as robust capabilities of neural networks to map complex non-linear dynamic processes. Furthermore, particular advantages arise from the combination of dynamic inversion and neural networks trained with the sliding mode learning method. This is the case because constraining the network’s states to the sliding mode offers an enhanced stabilising effect upon the error dynamics of the closed-loop control system. Using the example of a damaged unmanned aircraft under perturbation it is shown, that the sliding mode learning technique due to the dynamic calculation of the learning rate offers a higher speed of convergence than comparable established methods. The results especially emphasise the advantages of the interdisciplinary combination of control theory and machine learning techniques for the design of complex control systems.