On the influence of attributes for assessing similarity and sharing knowledge in heterogeneous populations of structures

The effectiveness of data-driven Structural Health Monitoring methods, which require dynamic response data from undamaged and damaged states within the same domain, is often hindered by the scarcity of data. The population-based approach to Structural Health Monitoring (PBSHM) aims to overcome this challenge by exploiting the transfer of damage-state knowledge across a population of structures via transfer-learning algorithms — specifically domain adaptation. However, meaningful inference through data sharing is possible only when the structures and their datasets are sufficiently similar. This study advances PBSHM by focusing on its two main phases: assessing structural similarity and developing effective transfer-learning strategies for heterogeneous populations. Furthermore, it explores the interdependence of these phases by addressing the role of structural attributes – such as material properties, geometry, and dimensions – in the performance of domain adaptation and similarity metrics. The proposed methodology employs Normal Condition Alignment (NCA) and Joint Distribution Adaptation (JDA) as domain-adaptation techniques across a population of laboratory-scale aircraft, considering tasks of increasing complexity. Additionally, the effect of material properties, geometry or dimensional differences in similarity assessment is investigated, comparing Graph Matching Networks and correlation-based distance metrics.