Fuzzy-Neural Controller and Real-Time Implementation of A Ball Balancing Beam

Nonlinear dynamic ball balancing beam has been successfully controlled by applying conventional methods, neural networks, and fuzzy logic respectively. Conventional methods necessitate strong mathematical and control background to derive equations. Neural networks learn to balance a ball, but the ball never settles down due to the fact that \discrete resolution of the boxes representation" was used. Fuzzy logic has continuous representation; however, it takes a lot of e orts to incorporate human knowledge into rules. In order to have a continuous representation learning system with less rules and mathematics, a system with blend of neural networks and fuzzy logic is proposed. Fuzzy logic membership functions are utilized to fuzzify input parameters; neural network interpolates the fuzzy rule set; after defuzzi cation, the output is used to train a smaller size of neural network; the weights of the later neural network can be adjusted to ne tune the controller. This controller balances balls with one third of the required 27 rules. With learning capability, it approaches its goal more frequently in general. In this paper, the design of the fuzzy-neural controller is discussed, the hardware setup is shown, and the performance is evaluated.