A neural network implementation of real-time fuzzy predictive control

Fuzzy predictive controllers have been applied to several applications with good control performance. However, this methodology often leads to nonconvex optimization problems, which are difficult to solve for fast processes, i.e. processes with small sampling times. This paper proposes a new methodology to apply a fuzzy predictive controller in real-time by using a neural network architecture, which receives data from the process and computes the control actions. Thus, the neural network is learned off-line, and its final structure guarantees that control actions are computed very rapidly. An internal model control structure is used to cope with model-plant mismatches and disturbances. The proposed methodology is tested in a realistic simulation of an experimental robot manipulator, where force and position are both controlled. The proposed scheme reveals very good control performance.