Hysteresis modeling of piezoelectric micro-positioning stage based on convolutional neural network

Abstract

The inherent hysteresis nonlinearity of piezoelectric actuator degrades the positioning accuracy of the micro-positioning stage. Prandtl–Ishlinskii model is widely used for piezoelectric hysteresis modeling, yet it is a rate-independent model with weak generalization ability. To overcome this problem, we proposed a convolutional neural network model based on the Prandtl–Ishlinskii model, which consists of a rate-dependent Prandtl–Ishlinskii model layer and convolutional network layer. The rate-dependent Prandtl–Ishlinskii model layer extends the traditional Prandtl–Ishlinskii model to describe the rate-dependent hysteresis behavior. The convolutional network layer with deep learning ability extracts the deep features of the input signal to improve the generalization ability of the hysteresis model. The experiment results indicate that the standard error of the proposed hysteresis model to predict displacement at unmodeled frequencies has been reduced by 18.74%–36.75% in comparison with the Prandtl–Ishlinskii model, which verifies that the proposed hysteresis model has not only higher accuracy but also stronger generalization ability.

Comprehensive Design Method of a High-Frequency-Response Fast Tool Servo System Based on a Full-Frequency Error Control Algorithm

“…Hu et al proposed a convolutional neural network model based on the Prandtl-Ishlinskii model. The standard error of the proposed hysteresis model in predicting the displacement at unmodelled frequencies was reduced by 18.74-36.75% [17]. Currently, the servo control algorithms for fast servo tools and piezoelectric actuators can effectively address the hysteresis phenomenon of a piezoelectric actuator.…”

Design and control of an x-y-θz parallel piezoelectric nanopositioning stage with a large moving platform

“…Aljanaideh et al (2013) proposed a ratedependent Prandtl-Ishlinskii model, which is constructed as a superposition of weighted rate-dependent play operators. Hu et al (2021) The feedforward controller based on the hysteresis inverse model can accelerate the response speed and reduce the hysteresis nonlinearity of the system to a certain extent. However, it cannot eliminate the positioning and tracking errors of the system.…”

“…Aljanaideh et al (2013) proposed a rate-dependent Prandtl–Ishlinskii model, which is constructed as a superposition of weighted rate-dependent play operators. Hu et al (2021) proposed a convolutional neural network model based on the Prandtl–Ishlinskii model, consisting of a rate-dependent Prandtl–Ishlinskii model layer and a convolutional net-work layer. Zhou et al (2021) introduced the dynamic rate of change of the input signal into the static weight of the classical PI model, forming a rate-dependent hysteresis model.…”