Abstract

Hysteresis is a nonlinear phenomenon that significantly complicates analysis and diagnosis. Reduced-order models that account for hysteresis, such as the Bouc-Wen model, are often calibrated using heuristic methods and require substantial domain expertise. This paper proposes a parameter estimation approach for computational models using physics-informed neural networks. A neural differential equation enables the network to capture hysteretic behavior, facilitating the derivation of critical parameters directly from vibration data while efficiently delivering highly accurate and reliable hysteresis predictions. An experimental setup involving a structure with a bolted joint calibrates a Bouc-Wen model, considering various preload and tightening torque conditions to demonstrate applicability. The results show that the identified model can accurately capture the hysteresis loop and the underlying dynamics present in the experimental data.

Keywords: hysteresis; model calibration; parameter estimation; neural differential equation.