Obtaining a Digital Twin of Systems via Approximation with DNN and Kautz Functions

This paper proposes a hybrid architecture for obtaining digital twins of dynamic systems under conditions of parametric uncertainties and unmodeled dynamics through approximation using deep neural networks (DNNs) and orthonormal Kautz functions. In the classical framework of digital twin operation based on supervised machine learning, orthonormal Kautz functions are used to approximate systems with real and complex poles, thereby extending the applicability of the approach. A DNN architecture has been developed for extracting the decomposition coefficients, ensuring high accuracy even in the presence of noise and parameter variations. The proposed method has been tested and validated using both simulation and experimental data. The data were obtained from a real electrohydraulic system via a measurement and control setup. Graphical results are presented, confirming the high accuracy and practical applicability of Kautz functions in the digital twin structure.