Structural health monitoring of bridges and viaducts is a crucial field to ensure infrastructure safety and durability. However, one of the main challenges is the scarcity of available data for newly monitored structures, which makes it difficult to train reliable machine learning models. Additionally, assessing the health condition of a structure at the beginning of the monitoring period is challenging, as at least six months to one year of data collection is typically required. This reduces the efficiency and responsiveness of monitoring campaigns. This thesis aims to explore the use of Transfer Learning and Domain Adaptation techniques to address this issue. Specifically, the objective is to develop methods that leverage data from an already monitored bridge to improve the analysis of another similar bridge with limited available data. To achieve this goal, latent space projection techniques will be used to align data distributions from different structures. Some relevant methods include Transfer Component Analysis (TCA) and Joint Distribution Adaptation (JDA). These approaches help mitigate domain differences and improve the model’s ability to generalize to new structures, allowing for faster health condition assessment with fewer data required.The research will be conducted using real-world data collected from monitored bridges, providing an excellent opportunity for practical application.
Contacts: Viviana Giorgi, Lorenzo Bernardini