A transfer learning strategy for SHM of roadway bridges via domain adversarial neural networks

<p>In the context of structural health monitoring of infrastructures, the combination of vibration-based systems with artificial intelligence algorithms is becoming particularly appealing. Nonetheless, one of the major challenges lies in effectively managing a large network of bridges and scheduling monitoring activities appropriately. Each bridge typically needs separate training if considering traditional machine learning models, which is often impractical when dealing with numerous structures. Moreover, the lack of labelled data hampers the implementation of supervised learning classifiers. To address these challenges, transfer learning offers the possibility to make inferences on a partially labelled/unlabelled target domain by transferring information from a labelled source domain. In this paper, a methodology based on domain adversarial neural networks is adopted to create a domain-invariant space where the classifier can generalize effectively across a network of bridges. This study examines the transfer between two post-tensioned concrete bridges, the Z24 bridge and the S101 bridge, focusing on the case in which the Z24 bridge serves as the source domain to successfully perform damage classification in the S101 bridge, where only healthy labels are assumed available.</p>