Learning algorithms for neural networks and development of neural-network-based active vibration absorbers

This thesis deals with the development of learning algorithms for recurrent and multilayer neural networks and application of neural networks to the control of vibration in rotordynamic systems. These learning algorithms are based on the concept of terminal attractors and the attractive condition used in the sliding mode control theory. Terminal attractors, which are based on the violation of the Lipschitz condition, represent singular solutions of dynamical systems. The fact that the system can reach singular solutions (or the desired solutions) in a finite time is utilized to enhance the learning rates of neural networks. The derivations of these new learning algorithms are formulated for both recurrent and multilayer neural networks. An inverse kinematic problem associated with a two-link robot manipulator is chosen as an example to verify the usefulness of new learning algorithms. Simulation results for both neural networks are presented.; A neural-network-based active vibration absorber has been developed to optimally suppress rotor vibrations caused by rotor unbalance. The unique feature of this new vibration absorber is its ability to optimally control vibration at different rotor speeds. Numerical examples dealing with a single-degree-of-freedom spring-mass system and a multi-degree-of-freedom rigid rotor supported by magnetic bearings are presented to verify the advantages of this novel neural-network-based active vibration absorber.