On the Influence of Structural Attributes for Assessing Similarity in Population-Based Structural Health Monitoring

Abstract

The viability of many machine learning methods within Structural Health Monitoring (SHM) is often limited by the lack, or the incompleteness, of the data required for implementing these algorithms. Indeed, learning a data-based SHM predictive model usually requires the dynamic response availability for undamaged and damaged states, and the assumption that both training and test data refer to the same domain. In this framework, the population-based approach to Structural Health Monitoring (PBSHM) aims at improving the performance and the robustness of diagnostic inferences, exploiting the transfer of damage-state knowledge across a population of structures. However, sharing these data produces a meaningful inference only if the structures, and their datasets, are sufficiently similar. Therefore, an initial phase of similarity assessment becomes essential before being able to apply transfer learning algorithms. This phase shows which structures are suitable for knowledge sharing, if any, reducing the possibility of negative transfer. Some distance metrics have been proposed, exploiting abstract representations of structures, such as Irreducible Element (IE) models and Attributed Graphs (AGs). Although these metrics can consider the structure attributes, many performed comparisons mainly concern structural topology. This study aims at broadening the application of similarity assessment, focussing on the geometrical and material differences in the distance metrics. Therefore, a heterogeneous population of laboratory-scale aircraft is analysed. These structures predominantly follow the geometry of a benchmark study conducted by the Structures and Materials Action Group (SM-AG19) of the Group for Aeronautical Research Technology in EURope (GARTEUR). The IE models of these aircraft are produced. Subsequently, Graph Matching Network (GMNs) are used to determine the similarity matrix. The structures in the Garteur population are topologically homogeneous, which enables a more accurate investigation of how attributes can influence distance metrics. This paper constitutes the first step in the Garteur structures population investigation. The similarity assessment results will establish which population members are most suitable for applying transfer learning algorithms. It will enable the subsequent development, and experimental validation, of a population-based strategy for damage identification across heterogeneous structures.