Neural networks and fuzzy logic for structural control

A new method for the active control of structures is proposed in this study. This method is based on the use of learning capability and adaptivity of neural networks and the high degree of flexibility and adjustability acquired through utilization of fuzzy logic. This method is classified as a "learning control method" with emphasis on the important role of the learning capabilities of the controller. However it can be classified as a smart or an intelligent control method too. This method has been called the "neuro-fuzzy control method" and its corresponding controller, the "neuro-fuzzy controller". Neuro-fuzzy controllers can theoretically cope with any nonlinearity, delay and imperfection in the controlled structure. Hence they can be considered as general controllers for structural purposes.In this method, a neural network called the "emulator neural network" is trained to learn to predict the response of the structure from the history of response and control signals. It learns about all the sources of nonlinearity and time delay, actuators capacity and any imperfections in the whole control system, implicitly. Then it is used in a preliminary control of the structure and the training of another neural network called the "neuro-controller". Neuro-controller has all the required knowledge of controlling the structure. At last a supplementary fuzzy controller is constructed to improve on the performance of the "neuro-controller". These two controllers which work in series, constitute the "neuro-fuzzy controller".In this study, the neuro-fuzzy control method is explained and its capabilities are numerically assessed through its application to the digital control of a three storey steel frame structure, subjected to different simulated earthquake excitations. Also for the sake of comparison, the predictive optimal control method is used in the control of the same structure, subjected to the same excitations. Then the results of the neuro-fuzzy control and the predictive optimal control methods are compared to each other. It is shown that the neuro-fuzzy controller is able to provide better results than the predictive optimal controller.Also, it is proposed to use as the criteria for the evaluation of capabilities of any control method, the three characteristics of adaptivity, prediction capability and simplicity of that method. It is discussed and demonstrated that the neuro-fuzzy control method satisfies these criteria better than the other proposed methods.Neural network related issues have played important roles in the progress of this study. These issues such as improvements on the learning speed of the muiti-layer feed-forward neural networks are discussed in this article too.U of I OnlyETDs are only available to UIUC Users without author permissio