A Transfer Learning Approach for Cross -Structural Health Monitoring of Steel and Concrete Systems

Structural Health Monitoring (SHM) is vital for maintaining the integrity of critical infrastructure, yet the variability of structural materials —particularly between steel and concrete —poses a significant challenge for generalizing data -driven models. This study presents a transfer learning -based approach that enables cross -material damage detection using limited labeled data from the target domain. A deep convolutional neural network (CNN), initially trained on the SDNET۲۰۱۸ dataset for surface crack detection in reinforced concrete elements, was adapted to monitor fatigue and impact damage in steel components using the Case Western Reserve University (CWRU) vibration dataset. Feature transfer and domain adaptation were performed through layer freezing and fine -tuning techniques to mitigate domain shift. The proposed model achieved ۹۴.۱% accuracy and an F۱ -score of ۰.۹۱ in the target domain; significantly outperforming baseline models trained from scratch. These results confirm that transfer learning can effectively reduce data dependency and improve damage recognition across structurally dissimilar systems. The framework offers a scalable and cost -effective solution for SHM in mixed -material infrastructure networks.

Keywords:

Transfer learning , Structural health monitoring (SHM) , Steel structures , Concrete structures , Convolutional neural networks (CNN) , Cross -domain learning , Damage detection , Vibration data , Crack detection , Deep learning