Neural networks for data-driven modeling of an AMB suspended rotor system : an implementation study
Shishkov, Aleksandr (2025)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025051612480
https://urn.fi/URN:NBN:fi:amk-2025051612480
Tiivistelmä
The thesis work presented a data-driven system identification approach for an Active Magnetic Bearing (AMB) suspended rotor system using feedforward neural networks (FNNs).
Building on the methodology of Nevaranta et al. (2023), the thesis report focused on a practical, step-by-step documentation of neural networks implementation to approximate the nonlinear dynamics of AMB systems under both single-input single-output (SISO) and multiple-input multiple-output (MIMO) excitation conditions.
Datasets were generated using a Simulink model of the AMB system, and detailed implementation procedures with relevant source codes are provided. The resulting FNN models demonstrated strong predictive performance on both simulated and experimental datasets, accurately capturing system dynamics with minimal prediction error.
This thesis bridges the gap between theoretical research and applied engineering practice, offering a validated framework for the deployment of neural network-based identification models in AMB applications and setting the foundation for future extensions towards control functionalities and dynamic modeling under variable rotational speeds.
Building on the methodology of Nevaranta et al. (2023), the thesis report focused on a practical, step-by-step documentation of neural networks implementation to approximate the nonlinear dynamics of AMB systems under both single-input single-output (SISO) and multiple-input multiple-output (MIMO) excitation conditions.
Datasets were generated using a Simulink model of the AMB system, and detailed implementation procedures with relevant source codes are provided. The resulting FNN models demonstrated strong predictive performance on both simulated and experimental datasets, accurately capturing system dynamics with minimal prediction error.
This thesis bridges the gap between theoretical research and applied engineering practice, offering a validated framework for the deployment of neural network-based identification models in AMB applications and setting the foundation for future extensions towards control functionalities and dynamic modeling under variable rotational speeds.