Physics-guided deep neural network for structural damage identification. (15th September 2022)

• Record Type: Journal Article
• Authors: Huang, Zhou; Yin, Xinfeng; Liu, Yang
• Is Part Of: Ocean engineering. Volume 260(2022)
• Publication Date: 2022-09-15
• Subjects: Damage identification -- Vibration response -- Finite element model -- Supervised learning method -- Physics-guided neural network; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2022.112073
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: The physics-driven method via finite element model and the data-driven methods via supervised learning is commonly used in the analysis of structural damage identification. The finite element model can be susceptible to environmental noise and modeling errors, which leads to model identification results deviating from the real damage. The supervised learning method faces the challenge of insufficient labeled damage data, which limits the learned model in identifying damage scenarios that are not contained in the training data. In this study, a novel physics-guided data-driven method is proposed to incorporate physical perception into the data-driven method for identifying structural damage. Specifically, the physics from the finite element model (e.g., structural dynamic characteristics) is incorporated into the neural network model to guide the damage feature learning from scarce measured data. In addition, a loss function with physical significance is developed to evaluate the discrepancy between the finite element model and the measured data, which guides the neural network to learn the damage features of the real structure. The learned neural network model has the potential to improve the damage detection results. The proposed methodology is validated by a numerical case study of a suspension bridge model and an experimental study of a frame structure. Highlights: The proposed method addresses the challenge of the lack of damage labels. The cross physics-data domain loss function is core of the PGDNN framework. Compared with 38.6% of CNN, the damage localization accuracy of PGDNN reaches 72.8%. The proposed method is verified by the numerical and experimental cases.