Exploring the Potential of Transfer Learning Applications for Structural Damage Classification

Abstract

Structural Health Monitoring (SHM) is essential for ensuring the safety and maintaining the functionality of structures and infrastructure systems, and Machine Learning (ML) techniques have shown great potential in rendering SHM as an automated process. However, unlike many other application areas of ML, when dealing with infrastructure SHM, there is little to no data available from damaged states to be used for ML training. To cope with this issue, a typical Finite Element Model (FEM) of a representative bridge beam structure is used for generating a damaged dataset. Secondly, creating a generalized network that can successfully work and perform a classification task for a variable set of structures seems to be one of the biggest challenges in the field of structural damage detection. In this work, the capability of Transfer Learning (TL) via Feature Extraction (FE) and Joint Training (JT) in generalizing the network is explored. Modeling uncertainties for supervised damage classification can be mitigated by utilizing a Deep Neural Network (DNN) comprised of Fully Connected (FC) and Convolutional (CONV) layers. In this regard, FEMs of three simply-supported beam structures with varying lengths were constructed, and acceleration time-history data was obtained from the interior nodes under the applied load history at the midpoint. The damage states were represented by the change in flexural stiffness within a range from 10% to 90%. Subsequently, the potential knowledge transfer was successfully implemented from the source domain to the target domains via TL and JT. Then, the proposed two-dimensional (2D) CNN network was tested with a target dataset that was not included in the training from another beam with a different length. The results indicate that FEMs could generate numerous source domains to achieve a generalized model even with including uncertainties.