Treffer: Physics-aware neural network integrated with stochastic subspace identification for quantification of physical changes

Structural Health Monitoring (SHM) is essential for ensuring the reliability of structural and mechanicalcomponents across various engineering domains. Traditional model-based SHM techniques oftenstruggle with complex systems and the limited availability of accurate physical models. On the contrary,data-driven, model-independent approaches, while simple and fast, frequently lack a comprehensive understandingof system physics and suffer from generalization issues. In recent years, Physics-InformedNeural Networks (PINNs) have emerged as a promising alternative that leverages both data and physicalmodels. Despite their success in state estimation for structural systems, limited research has focusedon inverse applications. A significant challenge in using PINNs for parameter identification lies in theircomputational demand. To address this, this study proposes a novel integration of Stochastic SystemIdentification (SSI) and Physics-Informed Neural Networks (PINN) for joint input-state-parameter identificationin structural systems (SSI-Pi-LSTM). SSI employs statistical analysis and subspace identificationtechniques to reliably estimate state-space matrices and dominant modal parameters from structuralresponse data. These parameters can then be incorporated into the PINN framework, reducing estimationtime and improving accuracy. This combined approach aims to bridge the gap between efficiency andprecision in structural parameter identification.