Tracking Control Using Recurrent-Neural-Network-Based Inversion Model: A Case Study on a Piezo Actuator

Abstract

This article is concerned with designing a broadband, high accuracy, and computationally efficient real-time controller for piezo actuators (PEAs). The essential component proposed is a recurrent-neural-network (RNN) based inversion model (RNNinv) used to compensate for the PEA nonlinearities. However, the obtained RNNinv may not efficiently model the low-frequency and/or time-varying dynamics of the system due to the limited length of the RNN training set. To address this issue, a linear model embedded with an error term is used to model the low-frequency dynamics in case it is not precisely modeled by RNNinv, and a predictive controller based on this linear model is then designed for precise output tracking. To validate the proposed control framework, the closed-loop stability condition is derived, and the RNNinv stability in unforced mode is investigated. To further improve the accuracy, a mechanism is proposed to separate the controlling dynamics to achieve higher accuracy for applications that cover broad and/or high-frequency ranges. The proposed approach was implemented on a commercial PEA and its performance was demonstrated through comparison with other controllers.