Control of nonlinear mechatronics systems by using universal learning networks

Nonlinear elements such as friction, dead zone, backlash are mixed in most mechatronics systems, and these are factors that make control accuracy of systems decrease and cause an oscillation. It is difficult to deal with the problems of modeling and control of such systems using common sigmoidal neural networks because there exist nonsmooth nonlinearities in the systems, and there is no easy way to incorporate knowledge and experiences accumulated from past study. In this paper, a design method for nonlinear mechatronics control systems is proposed, in which both of the control object and its controller are represented by using a Universal Learning Network, and the network parameters are trained using a random search algorithm called RasID. In this approach, knowledge and experience about models and controllers are easily incorporated into the network including the nonlinear elements and their compensation elements expressed by non-differentiable functions. Some simulations of a nonlinear crane control system with dead-zone characteristic were carried out. The effectiveness of the proposed design method is illustrated via simulations.