ADAPTIVE INVERSE CONTROL OF SYSTEMS WITH ACTUATOR NONSMOOTH NONLINEARITIES

摘要

Neural Networks (NN) have been used extensively in feedback control systems. Most of the applications rely on the universal approximation property of NNs. However, in most real industrial control systems there are piecewise continuous (PC) functions for which approximation results in the literature are sparse. Examples include friction, deadzones, backlash, and so on. In this paper a modified NN structure is given for approximation of PC functions of the sort that appears in friction and other motion control actuator nonlinearities. The NN consists of units having jump approximation basis functions. This modified NN and a NN based on classical sigmoidal activation function can approximate any PC function with discontinuities at a finite number of known points. Industrial motion device actuator nonlinearities fall in this category of functions, therefore, the modified NN structure is ideal for application in robotics and other industrial systems. The performance of the NN structure is demonstrated with two example control systems, one with unknown, unsymmetrical deadzone and the other one with backlash nonlinearities (ball and plate).