Modelování složitých dynamických soustav s využitím neuronových sítí

Abstract

This master's thesis focuses on the application of neural networks for modeling the dynamics of nonlinear systems. The primary objective is to compare the performance of machine learning methods with traditional physical modeling through two distinct case studies: a relatively simple magnetic levitation system and the complex multi-signal dynamics of a railway vehicle’s secondary suspension.

Two modeling concepts were selected for investigation: the Nonlinear AutoRegressive network with eXogenous inputs (NARX) and the State-Space Neural Network (SSNN). The study examines their robustness, closed-loop stability, and generalization capability across both systems. The results demonstrate that the NARX-based approach is highly effective even when using a simple architecture, achieving a fit with reference data exceeding 90\% and 75\% depending on the task. In contrast, the SSNN state-space approach showed inferior performance compared to the NARX model and required significantly higher computational resources during the training phase.