Comparison of Deep Learning Architectures for Nonlinear System Identification of a Hysteretic Piezoelectric Precise Positioner

Victor Henrique A. Ribeiro; Pedro H. L. S. P. Domingues; Gilberto Reynoso-Meza; Micky Rakotondrabe; Leandro dos Santos Coelho; Helon V. H. Ayala

doi:10.20906/sbai.v1i1.2658

Victor Henrique A. Ribeiro Industrial and Systems Engineering Graduate Program, Pontifícia Universidade Católica do Paraná, Curitiba
Pedro H. L. S. P. Domingues Department of Mechanical Engineering, Pontifícia Universidade Católica do Rio de Janeiro, Rio de Janeiro
Gilberto Reynoso-Meza Industrial and Systems Engineering Graduate Program, Pontifícia Universidade Católica do Paraná, Curitiba
Micky Rakotondrabe Laboratoire Génie de Production, National School of Engineering in Tarbes (ENIT-INPT), University of Toulouse, Tarbes, France
Leandro dos Santos Coelho Industrial and Systems Engineering Graduate Program, Pontifícia Universidade Católica do Paraná, Curitiba
Helon V. H. Ayala Department of Mechanical Engineering, Pontifícia Universidade Católica do Rio de Janeiro, Rio de Janeiro

DOI: https://doi.org/10.20906/sbai.v1i1.2658

Keywords: System identification, Hyperparameter tuning, Bayesian optimization, Deep Learning, Convolutional neural network, Long short term memory

Abstract

The characterization of hysteretic components poses a difficult nonlinear system identification problem. Several studies have addressed this by employing artificial neural networks, where deep learning (DL) has recently gained attention in system identification tasks. However, there is a lack of studies comparing different deep neural network (DNN) architectures. Therefore, this work proposes the comparison of three DNN architectures, including feedforward neural networks (FFNN), long short term memory (LSTM), and convolutional neural networks (CNN), for the characterization of a piezoelectric positioning system (positioner) typified by hysteresis. Moreover, Bayesian optimization is employed for hyperparameter tuning in all DNN architectures. Results show that all DL architectures achieved desirable values for the coefficient of determination (R2) and root mean squared error (RMSE). However, LSTM obtains the best overall results, outperforming both the FFNN and CNN, being a more appropriate black-box architecture for identifying frequency-dependent hysteresis loop shapes.