Abstract
Many studies have been undertaken in order to apply both the flexibility and learning ability of neural networks to control systems. With regard to servo control, an adaptive type controller and an learning type controller are proposed, and verified by applications to a force control system. An adaptive type identifier has also been studied. The convergence speed of this type of identifier is fast because learning is performed within one trial. In contrast, although the convergence speed of the learning type identifier is slow because its learning requires after several trials, it offers the possibilities of greater robustness when the plant dynamics are unknown and there is external noise. Jordan has stuided several types of identifiers using neural networks. However, learning type identifiers for discrete systems have not been studied yet. They should be studied because of the advantages to the offer of software servo control.
Therefore, this paper proposes a practical design method for both a learning type direct transfer identifier and an inverse transfer function identifier for discrete systems. The direct transfer function identifier constructs the forward dynamics of the object plant in the neural network. The inverse transfer function identifier constructs the inverse dynamics of the object plant in the neural network. Analytical approach describes the local stability condition of the proposed identifiers when both the object plant and the neural networks are linear. Simulation results for a second order plant confirm both the characteristics, and the nonlinear saturated function (sigmoid function) effect.
