On the Use of Graph Kernels for Assessing Similarity of Structures in Population-Based Structural Health Monitoring

Population-based structural health monitoring (PBSHM), expands the implementation of structural health monitoring concepts from a single structure to a group of structures. Within the populations of interest, it is useful to assess the similarity of structures, in order to form communities and networks; i.e., to establish clusters. By doing so, it is possible to infer whether transfer learning is likely to be applicable across a population. To address this, structures are represented in graphical form via irreducible element (IE) models or attributed graphs (AG), that encompass the topology, material properties and geometry of the structures. Kernel-based methods - also known as graph kernels - can then be applied to these IEs and AGs to assess the similarities within the population. In this paper, a number of comparisons are made between graphical representations of structures at different levels of resolution; these vary from assessing similarity at a purely topological level, to an increase in complexity, where discrete and continuous node labels are evaluated via graph kernels.