Enhanced Identification of the Parameters of a Jeffcott Rotor from Run-up Transients using Physics-informed Neural Networks with Sinusoidal Activations

Conference proceedings

en

Accurate identification of rotor-dynamic parameters is essential for the health monitoring of rotating machinery. Classical methods such as FRF fitting and Kalman filtering often require long stationary datasets and perform poorly under noisy transient conditions, such as during run-up tests. Physics-informed neural networks (PINNs) have emerged as a data-efficient alternative; however, standard tanh/ReLU architectures are susceptible to spectral bias, which limits their ability to capture rapidly varying responses. In this study, we employ a PINN with sinusoidal representation networks (SIREN) to estimate the stiffness, damping and eccentricity of a Jeffcott rotor undergoing constant angular acceleration directly from non- stationary run-up data. Sinusoidal activations mitigate spectral bias and improve the representation of the chirp-like dynamics near resonance. Using synthetic displacement signals with controlled noise, we demonstrate that the SIREN-based PINN outperforms a tanh-activated baseline in terms of both accuracy and training efficiency. Beyond full-signal runs, feasibility studies are discussed using only short run-up windows, showing that reliable parameter identification is possible from sub-second transients. The results suggest that using sinusoidal representations in PINNs is a good way to estimate parameters from data where the system is only temporarily operating. This is in addition to steady-state approaches and could lead to online health monitoring systems that work in practice.

PINN, parameter estimation, sinusoidal activation, rotor dynamics, time-domain analysis

2025-09-25