Neural adapted controller learned on-line in real-time

Abstract An adapted controller consisting of a PD structure aided by a neural network is considered. It is dedicated to an unstable magnetic levitation system (MLS) to maintain a levitated sphere in the real-time. The main motivation is to create a self-tuning intelligent device that can easily change the controller parameter estimates based on adaptive update rules. How to formulate these rules in order to not disturb or ruin the system stability is the most serious question. This issue is discussed and verified by real-time experiments carried out at our apparatus that enables to vary its parameters. The neural network learned on-line guarantees the robustness of the control despite the system parameter are changed and disturbances are introduced. One can expect a better stabilizing performance as well.