Identification of Macro-Moving Stage in Precision Positioning System Based on Neural Networks

A special neural network with non-smooth activation function is proposed in this paper to approximate the non-smooth nonlinearities such as dead zone coupled with hysteresis in macro-moving stage of precision positioning systems.In order to train the nonsmooth neural network,the generalized gradient is introduced into the Levenberg-Marquardt algorithm to model the behavior of the macro -stage.The expanded input space method is used to transform the multi-value mapping of hysteresis into the one-to-one mapping.For modeling a complex nonlinear system,the necessity of expanded input space is proven and its corresponding expanded way is proposed. The experimental results show that the satisfactory modeling performance is achieved.