| Recently, fuzzy logic systems and neural networks have emerged as effective tools for modeling and controlling complex or uncertain processes. In light of their similarities and differences, upsurge of interest centers on merging fuzzy systems and neural networks into integrated systems. The aim of this dissertation is to establish a generalized framework integrating both fuzzy systems and neural networks. Models pertaining to this framework possess fuzzy thinking, reasoning and the learning ability of neural networks. Several computationally efficient algorithms are developed. In particular, SRAT (selective rule activation technique) and HNFS (hybrid neuro-fuzzy system) are proposed to tackle the so-called curse of dimensionality encountered in conventional neuro-fuzzy models, which make it difficult to implement existing neuro-fuzzy schemes. The SRAT filters can constrain the size of the neuro-fuzzy model based on the predefined closeness function. Since the number of rules is reduced, learning time and computational expense are significantly reduced. Equipped with an optimal search algorithm, the SRAT can also detect dependencies of the input variables automatically. The HNFS is a fuzzy inference system with each rule consequent containing a modified feedforward neural network, which is developed to simplify the computation. Along with this hybrid neuro-fuzzy architecture, the sliding mode learning method is incorporated into the identification mechanism. Simulation results show that the resulting system can identify complex and/or uncertain systems accurately. Furthermore, this hybrid architecture bridges over the well developed linear system theory and the new neuro-fuzzy models. Hence, the linear control techniques can be applied to design robust control laws. Analytical expression and computational efficiency are the major advantages of the new control laws. Demonstration of the effectiveness and the robustness of the integrated identifier-controller scheme has been achieved by balancing an inverted pendulum and controlling a multilink robotic system in the presence of substantial parameter variation and disturbances. Because of the identification accuracy and computation efficiency, the developed integrated neuro-fuzzy system can be realized with current VLSI technology and has promising applications in different areas such as robotics, flight vehicles, and other industrial processes. |
