Artificial neural network approach for experimental model identification and control of structures

Most active control studies have focused on the simulation approaches to understand and predict the behavior of structures under earthquake loading. However, issues such as controller and sensor interface and structural complexities can not be accounted for during simulation studies. The paper provides an experimental study of active control of structures using neural networks and examines the behavior of structure and the generalization to new inputs under dynamic loading. A series of experiments were conducted in which a designed 3-story steel scaled model, mounted on a shake table, was tested. While previous studies were limited to only three loadings, the purpose of the additional work contained in this paper is to apply six loadings with varying bandwidths to examine the generalization behavior of the trained neural network. In addition, in this research fiber optic sensors as well as accelerometer, were mounted on the centers of the top and the base frames in the longitudinal direction of the structure. This paper discusses the details of the scaled model structure, and the various loadings that were used to examine the generalization of the trained neural network architecture to new inputs. The results indicate that the neural network controller, developed based on the dynamics of the structure, can significantly reduce the vibration of the structure, and is able to generalize to variable new inputs.