Application Of Deep Learning In Feedforward Control Of Piezoelectric Ceramic Actuator

The synchrotron radiation beamline is a nonlinear system with a complex structure and a wide variety of equipment.The traditional manual beam adjustment mode requires experienced engineers to adjust the positions of the motion axes on the beamline equipment within their range one by one until meeting the experimental requirements,the whole process is time-consuming and labor-intensive.To improve the operating efficiency of the beamline,the researchers based on the differential evolution algorithm and based on the principle of “survival of the fittest,survival of the fittest”,the researchers realized the beamline intelligent adjustment system,which shortened the beam adjustment time from the original hours to within 30 minutes.The system requires good repeatability of the motion axis of the equipment.When the pitch axis of the monochromator is added,the algorithm does not converge due to the inherent hysteresis nonlinearity and creep characteristics of the piezoelectric ceramic.Therefore,the popularization and application of the intelligent beam adjustment system are limited.Piezoelectric ceramics have the characteristics of fast response,high resolution,small size,large driving force,low heat generation,and low noise.They are widely used in high-precision equipment.To improve the control accuracy and motion repeatability of piezoelectric ceramics,mathematical modeling methods are usually used for compensation.The deep neural network emerging in recent years not only has strong nonlinear fitting ability,feature extraction ability,and fault tolerance ability,but also outperforms previous mathematical modeling and control algorithms in the face of massive,high-dimensional data.The Shanghai Synchrotron Radiation Facility,which operates nearly 5,000 hours a year,has massive data resources,and the high availability of data brings great opportunities for the application of deep neural networks in the field of synchrotron radiation.This paper takes the high-precision and the high-repeatability motion of the piezoelectric ceramic actuator as the control goal,and studies the structure of the deep neural network,the design method of the model,the establishment of the data set,the training process of the model and the experimental verification.The main work and achievements are as follows:1)A multi-input-single-output multilayer perceptron is designed for piezo-electric ceramic feedforward compensation.The input voltage,output displacement,and target displacement of the previous position are used as the input of the neural network.The output voltage of the controller is used as the output of the neural network.After100,000 epochs of training,the model with the smallest loss value is obtained.The model is used to test the actual displacement when the input signal is triangular wave signal,a sinusoidal signal,a damped sinusoidal signal and an amplified sinusoidal signal under 10 Hz operating frequency.When the operating frequency is set to 10 Hz,20Hz,50 Hz and 100 Hz in turn,the root mean square errors of the actual displacement and the target displacement of linear motion are 0.0098μm,0.00983μm,0.01007μm and 0.01063μm,respectively.The test results show that the deep neural network has good feedforward compensation ability and good frequency generalization ability.2)In order to improve the training speed and explore the feasibility of online learning,a recurrent neural network including a GRU layer and a multi-layer perceptron is designed for piezoelectric ceramic feedforward compensation.When the input signal is triangle wave,sine signal,amplitude sine signal,amplitude sine signal under 10 Hz operating frequency,the maximum displacement error is-0.465μm.Due to the complex structure of the recurrent neural network,which makes it difficult to converge during training and to obtain a better model,the actual control effect is not as good as the multilayer perceptron in 1).3)On the basis of 2),an optimization method based on the least squares method is proposed to fit the parameters of the piezoelectric ceramic main hysteresis loop to obtain the linear fitting relationship between the input voltage and the output displacement.The control of the network was fine-tuned so that the maximum error dropped from-0.465μm to-0.273μm.This paper is the first domestic application of a deep neural network to the field of synchrotron radiation beamline station control.The deep neural network modeling,data set establishment,training process research and other achievements for piezo-electric ceramic nonlinear systems are used to improve the performance of a single motor.The motion accuracy and repeatability of beamline equipment and the realiza-tion of intelligent control of the entire beamline station have certain reference value.