Physics-guided generative adversarial network for probabilistic structural system identification

Yu Yang

doi:10.1016/j.eswa.2023.122339

2024-04-01

PRE

Yu Yang

Yongming Liu *

DOI：10.1016/j.eswa.2023.122339

原文链接

原文求助

摘要

En 中文

Structural system identification (ID) is an important tool for many structural and infrastructure applications, such as structural health monitoring and structural model updating. A novel physics-guided generative adversarial network (PG-GAN) is proposed in this study for probabilistic structural parameter ID. The PG-GAN leverages the powerful distribution learning ability of GANs while also enabling physical awareness by incorporating additional generators for physical parameters and developing a physics-based loss function that links the data generator with the parameter generators through existing physics knowledge to guide the training of generative models. Two experiments are performed to validate the effectiveness of the PG-GAN for structural parameter ID and demonstrate its application in structural damage detection. The PG-GAN is shown to successfully capture the uncertainty inherent in structural parameter estimation and accurately locate and quantify structural damages in terms of probabilistic distributions of the damage extent. The main contributions of this study are threefold: (1) compared with traditional methods for structural parameter ID, the proposed PG-GAN provides an intelligent framework for probabilistic identification of structural parameters without the need for any assumption on the data distribution and does not require an initial guess of parameters; (2) unlike traditional data-driven GANs, the proposed PG-GAN provides a linkage between GAN and physics knowledge to enables physical awareness in generative model training; and (3) the proposed PG-GAN can serve as an intelligent diagnostic system to better inform the decision-making on structural health management as it can quantify the uncertainty in the structural damage extent.

Keyword:

Generative adversarial network

System identification

Probabilistic inference

Uncertainty quantification

Physics-guided machine learning