Deep Learning for Identifying Hysteresis Models of Piezoceramic Actuators in the Linear Frame

Abstract

Piezoceramic actuators have been already applied in precision positioning in terms of simple and compact structure, free from noise, and high theoretical positioning resolution. However, the hysteresis characteristic limits the further improvement of positioning accuracy. Nowadays neural networks (NNs) have revolutionized progress in the identification and global linearization tasks, which makes it potential to employ deep learning in identifying hysteresis model of piezoceramic actuators in the linear frame. This paper aims at achieving the identification of the hysteresis model by means of NNs. Based on the Preisach model, datasets are obtained in Matlab. Identification of nonlinear coordinates is accomplished by the auto-encoder afterwards. As a result, weights of the various network branches are computed. Comparing the hysteresis model displacement output with the NNs reconstruction, it is found that the curves are basically consistent. The experimental results confirm the correctness of this method, which is of great significance for analysis and control of nonlinear systems.