Data-driven approach for modeling and controlling thin-film piezoelectric actuator

The precise and robust control of piezoelectric actuator is a critical concern in microelectromechanical systems (MEMs), particularly in applications that necessitate the use of high voltage and frequency. Lead Zirconate Titanate (PZT) is a commonly utilized material in such applications due to its large piezoelectric coefficient. However, the material nonlinearities, such as hysteresis and creep, that are inherent in PZT pose significant challenges for control efficiency.

In response to these challenges, various control approaches that incorporate feedforward and/or feedback control have been proposed. These approaches utilize analytical models, such as the Preisach, Prandlt-Ishlinskii and Duhem models, to phenomenologically describe the hysteresis phenomenon. The advantage of this approach lies in the ability to employ an inverse hysteresis operator in control design. Nonetheless, analytical method is limited to the quasi-static regime and does not capture the material’s inherent asymmetrical hysteresis loop. To address this limitation, more complex models, such as neural networks, have been integrated into the control loop as the plant or compensator due to its versality of representing complicated nonlinear function.

The purpose of this study is to investigate the use of machine learning in the modeling and control of thin-film piezoelectric actuators. A deep neural network, consisting of recurrent and convolutional layers, will be trained using measurement data to predict the displacement response of the actuator to unipolar and bipolar voltage inputs with varying amplitudes and frequencies. In addition, another neural network will be used as a hysteresis compensator, trained to predict voltage control inputs based on the displacement history. To improve the efficiency of the control loop, an additional neural network will be employed as a feedback controller to account for uncertainties and external disturbances. The ultimate goal of this work is to develop a fully data-driven, hysteresis-free piezoelectric microactuator.