Time and frequency domain physics informed neural networks for parameter estimation : comparative study of time-domain and frequency-domain approaches

Abstract

This thesis investigates parameter estimation of second-order dynamical systems using Physics-Informed Neural Networks (PINNs) in both the time and the frequency domain. A linear system with known model structure is considered, where the damping and stiffness coefficients are identified while the mass is assumed known. Simulation experiments with different levels of measurement noise are conducted to compare estimation accuracy, response reconstruction, and robustness. The results show that both methods can effectively recover the unknown parameters and reconstruct the system response. The frequency-domain method performs better for damping estimation, while the time-domain method gives better stiffness estimation. Overall, the study shows that time-domain and frequency-domain information are complementary and that PINNs provide an effective approach for parameter estimation.