Assessing similarity requirements for effective transfer learning across a network of rigid frame bridges

In the field of structural health monitoring, the integration between vibration-based systems and artificial intelligence algorithms is becoming particularly attractive for real-time assessment of bridge infrastructure. Given the need for sufficiently large datasets to train a robust classifier and the lack of labelled bridge response data, especially in damage conditions, one of the main challenges lies in leveraging data-driven monitoring techniques, based on measured structural response components, to enable effective management of entire bridge networks. To avoid separate training for each bridge, transfer learning strategies could therefore represent a valuable solution to enhance the available dataset by collecting information from similar structures and leveraging the acquired knowledge to make inferences across the network. However, there is still a need to check if particular similar structural characteristics could improve knowledge transfer for a certain damage class, since transfer performance is strongly correlated with the degree of similarity between two structures. Focussing on domain adaptation and working with properly-selected features, this paper proposes a methodology to (i) assess bridge similarity via the use of specific indexes and (ii) study transfer effectiveness under varying structural similarity and damage typologies. For validation purposes, an extensive simulation campaign is carried out to build a large network of numerical bridge models, characterised by multiple configurations of three-span rigid frame bridges and subjected to two damage scenarios. The results show the possibility to group similarly-behaving structures together and identify damage-dependent bridge properties that are required for a successful transfer.