Realization of 2 DOF Control System Using Multilayered Neural Networks

Recently, many kinds of neural controllers have been studied as a new paradigm of control methods. However, most of these are aimed at internalizing inverse models of plants in neural networks. Therefore, in those methods, neural networks have been obliged to learn nonproper systems. As a result, neural networks cannot perform correctly as compensators. On the other hand, in conventional control fields, new control methods based on the two-degree-of-freedom (2 DOF) system have been developed and their effectiveness has been reported in many papers. However, this methodology for constructing control systems is based on the fact that controlled plants are linear systems. In consequence, it is not clear whether these methods are also effective for nonlinear systems. In this study, a neural controller based on the structure of the 2 DOF control system is investigated. The main aim of this study is to make the neural controller compensate for plants with time delay or friction, ensuring stability of the system by means of conventional feedback control (PI-control). For the neural controller, the number of neurons in the input layer can be determined by the transfer functions of the model and the plant. Moreover, the structure of the neural controller has recurrence. The weights of this recurrent network can be obtained by using a genetic algorithm. To confirm the feasibility of our proposed method, we carry out simulations using a plant with a time delay and friction as a practical example.