An application of domain adaptation for population-based structural health monitoring

In the field of civil infrastructure, Structural Health Monitoring generally suffers from a scarcity of labelled damage-state data. To solve this issue, this work adopts a Transfer Learning approach for leveraging information from a source structure, characterised by a rich class of damage labels, to improve inferences on a target structure with limited knowledge. The goal is to train a machine learning algorithm on a bridge undergoing damage and to afterwards transfer the available labelled damage-state data across the members of the investigated population. Given possible differences exhibited by each structure, a domain adaptation technique in the field of statistic alignment, called Normal Condition Alignment (NCA), is applied to match different distributions in a shared feature space. The methodology is validated on a heterogeneous population composed of two numerical bridges of different geometry and materials, representing the Z24 and the S101 benchmark bridges. Finite Element Models are built to simulate healthy conditions and several damage cases. The natural frequencies describing such scenarios are considered as damage-sensitive features and thus employed to characterise the two domains and fed to a supervised learning-based classifier. The presented approach is deemed effective to provide mappings that allow the exchange of health-state information from source to target datasets, becoming a promising approach to be applied within a population of real bridges.