Neural Networks for Multiobjective Adaptive Structural Control

Abstract

The concept of using neural networks to predict future values of constrained performance variables as part of an adaptive structural controller is proposed. Neural networks are first trained using simulated responses and then they are used in simulations with input disturbances having different frequency contents. Because of this, as well as other factors, one or more prespecified constraints may be violated. The neural networks are designed to permit predictions of these unsafe vibration levels, stresses, actuator saturations, etc., before they can occur. To prevent their occurrence, the control laws may be adaptively tuned, maximizing safety and reliability. To illustrate the proposed methodology, numerical examples are shown using linear quadratic Gaussian control of a simple example three-story building model subjected to earthquake excitation, with an active brace used for control. Elman neural networks are chosen, filtered white noise is used as the input disturbance for training, and actual earthquake records are used as input disturbances for testing the neural networks and finally the adaptive controller.