A parallel framework of physics-informed neural networks for model identification of linear and nonlinear systems
Résumé
Recent advances in sensor and communication technologies have enabled widespread monitoring of complex structures using both model-based and data-driven methods for Structural Health Monitoring (SHM).
While model-assisted approaches struggle with the complexity of real-world systems, purely data-driven methods offer simplicity and speed but lack physical interpretability. Physics-Informed Neural Networks (PINNs) offer a promising middle ground by integrating data with governing physical laws. However, most existing PINN applications focus on solving forward problems, with limited attention to inverse parameter estimation relevant to SHM.
This study presents a parallel PINN framework for efficient state and parameter estimation in complex structural systems. To address the computational burden of high-dimensional systems, the global dynamics are reformulated into a set of coupled lower-dimensional subproblems, one for each degree of freedom, without any physical partitioning of the domain. This algebraic decoupling retains structural continuity through shared global matrices, enabling localized learning while preserving global consistency. The approach is validated on both linear and nonlinear systems using harmonic and real earthquake excitations, demonstrating robustness under high noise, sparse data, and varying system complexity. An experimental study on a scaled shear frame further supports its practical relevance for real-world SHM.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |