On the Influence of Structural Attributes for Transferring Knowledge in Population-Based Structural Health Monitoring

The recently proposed theory of Population-Based Structural Health Monitoring (PBSHM) aims at improving diagnostic inferences, by sharing damage-state knowledge across a population of structures via transfer-learning algorithms - specifically domain adaptation. Before applying these algorithms, the similarity between structures, or substructures, should be evaluated. This assessment helps prevent negative transfer, ensuring better performance and higher robustness of data-based SHM. When structures are sufficiently similar, different transfer-learning strategies can be applied, according to the original features and the specific case study. In this framework, structural attributes play a crucial role, especially for heterogeneous populations in which the main differences can be caused by material properties, geometry or dimensions. Therefore, investigating how to consider the influence of these properties in distance metrics became necessary, and new similarity metrics have been adopted to focus on geometric features and dimensions. However, to gain a comprehensive understanding of attribute relevance, and to address it at the similarity-evaluation phase, it is necessary to evaluate the performance of transfer-learning algorithms as these structural features vary. The present work extends this research by examining the effect of material and dimension attributes on the performance of a domain adaptation method - the Transfer Component Analysis (TCA). This analysis is applied to an experimental population of laboratory-scale aircraft, comprising structures with different materials and dimensions, and similar topology. A confusion matrix is employed to compare the findings and show how these properties can influence the transfer-learning performance, especially for localised damage, thus highlighting the importance of their evaluation in the context of PBSHM.