Intelligent Learning Based Modeling Identification And Control Of Hysteretic Systems With Application To Piezo Nano-Motion Stage

The piezoelectric-driven nano-manipulation stage is the core component of nanoscale detection equipment represented by scanning probe microscopy and atomic force microscopy.The positioning and control accuracy of the nano-manipulation stage determines the quality of nano-detection.However,the existence of complicated hysteresis nonlinearity in piezoelectric ceramic actuators severely impairs the control accuracy of nano-stages,causes system oscillations,and even results in instability,which poses a significant challenge for modeling and precision control of piezoelectric nano-stages.This dissertation takes the piezoelectric nano-manipulation stage as the research object and deeply investigates the theory and method of hysteresis nonlinearity modeling,parameter identification,and precise compensation control based on intelligent learning algorithms,aiming to eliminate the influence of hysteresis on system’s performance,to realize the nano-precision motion control of the piezoelectric stage,and to extend its application in ultra-precision detection and manufacturing equipment.The main research contents of this dissertation are as follows.A modified particle swarm optimization-based Duhem hysteresis model identification method is proposed to address slow convergence and the local optima problems of conventional particle swarm optimization(PSO).A randomness operator is introduced in the optimization process,which acts separately on each search space dimension to improve PSO’s convergence and parameter identification accuracy.Experimental investigations are conducted on the constructed piezo-actuated nano stage and compared with the conventional particle swarm optimization method and genetic algorithm(GA).The experimental results demonstrate significant improvements by the proposed method in hysteresis model identification accuracy.To explore the application of intelligent learning algorithms in dynamic modeling,a novel gray box neural network-based piezoelectric hysteresis modeling and identification method is proposed.A system-level quasi-static differential hysteresis model is used to construct the gray box neural network.To describe the rate-dependent hysteresis behavior,a dynamic model based on nonlinear autoregressive moving average with exogenous input(NARMAX)is introduced into the quasi-static model,in which specific neural network weight functions are designed to reflect the model parameters.The parameters of the gray box hysteresis model are determined by neural network training.To deal with the multi-valued property of hysteresis,a generalized input gradient mapping is proposed to convert the multi-valued mapping of hysteresis into a one-to-one mapping.Simulation and experimental results on the piezo-actuated nanopositioning stage show that the proposed gray box neural network method’s accuracy is significantly higher with less training time than existing black-box and PSO-based parameter identification methods.A high-speed tracking control method for the piezo-actuated nano stage is proposed based on inverse hysteresis model feedforward compensation and grey box neural network.A novel gray box inverse hysteresis model is constructed as the feedforward hysteresis compensator,and the parameter identification problem of the inverse hysteresis model is transformed into the gray box neural network training problem.After that,the model predictive controller is introduced as a high dynamic feedback controller,and an integral compensator is employed to eliminate the steady state error caused by the external disturbances.Experimental results indicate that the proposed feedforward hysteresis compensation strategy improves tracking performances of the nano stage,and achieves the advantages of low modeling error and fast training speed compared with the existing methods in the literature.To overcome the problem that the inverse hysteresis-based control methods are susceptible to external disturbances,an enhanced intelligent predictive control strategy is proposed,which does not require an inverse hysteresis model.First,a gray box neural network hysteresis model with the generalized input gradient is dynamically linearized.Then the dynamically linearized hysteresis model is integrated with the predictive control scheme to produce an enhanced predictive control scheme with squeezed search space.The enhanced model predictive control scheme can reduce the computational burden and improve the high-speed precision tracking performance.The convergence of the proposed predictive control method is analyzed by driving a convergence analysis.Theoretical analysis and experimental results on the piezo-actuated manipulation stage indicate that the proposed method significantly improves convergence speed and high-frequency trajectory tracking performance compared with the conventional model predictive control method.