Monday  |  Salon 12  |  02:00 PM–02:20 PM

#17096–Transfer Learning for Structural Damage Classification: Transfer the Knowledge from Cyber to Physical Systems

Structural Health Monitoring (SHM) plays a crucial role in ensuring the integrity and reliability of structures and infrastructure systems. Machine Learning (ML) techniques have shown immense potential in automating SHM processes. However, a significant challenge arises when dealing with infrastructure SHM due to the scarcity of available data from damaged states for ML training. To address this issue, researchers often rely on Finite Element Models (FEMs) to generate simulated damaged data. Nevertheless, the utilization of uncalibrated FEMs introduces inherent uncertainties that hinder the effectiveness of supervised ML models for SHM. In this study, we investigate the capability of Transfer Learning (TL) to mitigate modeling uncertainties in supervised damage state classification. Our approach involves employing a Deep Neural Network (DNN) architecture comprising convolutional neural networks (CNNs). We construct a FEM of a laboratory structure and validate it against measured structural responses under known excitations. Subsequently, FEM models with a set of variables, relative to the validated model, are created . These models serve as source domains for generating diverse datasets, and TL techniques are employed to determine the optimal level of disparity between the source domains' parameters and the target domain for effective TL implementation. By leveraging TL using feature extraction (FE) and joint training (JT), our proposed methodology aims to enhance the prediction accuracy of damage states in experimental setups. The findings of this study contribute to advancing the field of SHM by addressing the challenges associated with limited real-world damaged data and modeling uncertainties.

Burak Duran   University of New Hampshire
Yashar Eftekhar Azam   University of New Hampshire