Neural network output feedback training for optimal vibration suppression

Neural network control of flexible structures demonstrate better settling time and energy dissipation than linear design methods. Optimal tuning of passive vibration absorbers for reduced order control is examined using linear and nonlinear cases. Quasi-Newton optimization performed well in the passive control case, but not in an active nonlinear example. In both cases unsupervised least mean squares or backpropagation techniques were unstable. Lessons on unsupervised training for dynamic system control are based on training data. Output feedback parameters are “tuned” for best disturbance rejection using total energy as a cost function. The passive case trained reaction mass actuator coefficients which used position and velocity feedback for stiffness and damping respectively. Laboratory data for the linear single neuron case validated the model fidelity. The active case used air jet thrusters as a bang-bang type control for a 20 bay truss. This work illustrates the usefulness of neural networks for linear and nonlinear control, in this case for flexible structures