An Attention-Augmented Long Short-Term Memory Network for PINN-based Structural Health Monitoring
Résumé
Structural Health Monitoring (SHM) plays a vital role in maintaining the safety and functionality of structural and mechanical elements in different engineering sectors. Conventional model-based SHM techniques often struggle with the intricacies of complex systems due to oversimplified models. In contrast, data-driven, model-free methods provide speed and scalability but often fail to incorporate system physics, which leads to challenges in generalizing. Physics-informed neural networks (PINNs) have emerged as promising solutions by merging data-driven models with physical insights. Nevertheless, the majority of PINN applications in SHM concentrate on estimating dynamic system states, while inverse estimation remains less explored.
Most existing inverse methods demand complete state measurements, which are rarely achievable in practical scenarios. This research presents a novel PINN framework-a physics informed attention-augmented Long Short-Term Memory network (Pi-Attn-LSTM)-designed for estimating parameters in structural systems without the need for comprehensive state data, thereby enhancing its feasibility. The model utilizes a temporal attention mechanism in an encoder-decoder setup to adaptively focus on critical features in sequential data, enhancing the accuracy of parameter estimation in dynamic environments. Validation is executed using both numerical simulations of multi-degree-of-freedom systems, including an IASC-ASCE-Benchmark problem, and experimental tests on data from a scaled aluminium frame. Moreover, comparative studies between the proposed Pi-Attn-LSTM and a typical Pi-LSTM model demonstrate the superior robustness and accuracy of Pi-Attn-LSTM. The extensive validation study shows that Pi-Attn-LSTM not only speeds up convergence but also reliably identifies localized structural degradation, highlighting its potential as a viable tool for SHM in real-world scenarios.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |