Analytical and experimental study of nonlinear structural control using neural networks

A comprehensive analytical and experimental study of actively controlled nonlinear structures using the learning capabilities of the neural networks is presented. The method utilizes the neural networks learning capability of the control tasks and referred to as neuro-control method. The neurocontrollers are developed and trained by the use of independent neural network models called the emulator neural networks. The emulator neural networks provide a training path for the neurocontrollers as well as identify non-parametric models of the controlled system. This study is developed and conducted in two consecutive phases. The first phase comprised the analytical study of nonlinear structural control. In this phase, two neurocontrollers with their associated emulator neural networks are developed, trained, assessed and tested in numerical simulation of a three story steel frame model. First neurocontroller is trained when the response of the structure remained within the linearly elastic range, and has been called the linearly-trained neurocontroller ( LTC). The second neurocontroller is trained from the nonlinear response of the structure, and has been called the nonlinearly-trained neurocontroller (NTC). The emulators effectiveness in model reprehension is presented and discussed. Similarly, both neurocontrollers, effectiveness and robustness are evaluated and presented. The second phase of this study includes a comprehensive experimental verification of the structural neuro-control method. These experiments are carried out on the earthquake simulator at the University of Illinois at Urbana-Champaign. Two linearly-trained neurocontrollers, referred to as U3A and UA, with different architecture, feedback and sampling rates are developed, experimented and evaluated. The training procedure for the neurocontrollers were achieved by the aid of multiple emulator neural networks in parallel-series fashion. These emulator neural networks vary in their prediction capabilities, sampling rates and time delays. Together, they constitute the neurocontroller training source. The stability and robustness of the neurocontrollers are demonstrated analytically and experimentally. Finally, a linear quadratic Gaussian optimal regulator ( LQG) is designed from an experimentally identified model. Then, the optimal controller is assessed and compared to the neurocontroller results. Additionally, the stability and robustness of the controller is illustrated and outlined. It is shown experimentally that the neurocontrollers performance is superior to the optimal controller performance.