Similarity assessment of structures for population-based structural health monitoring via graph kernels

In population-based structural health monitoring, the aim is to make inferences about the health of structures using information from a population of other structures. It is possible to use transfer learning here, as long as the structures that are used for transfer behave similarly to each other. As a result, assessing the similarity of structures and the data collected from those structures is necessary for successful transfer. In this paper, ideas from kernel and graph theories are used to assess whether the constructional makeup of two engineering structures – a bridge and a wind turbine, for example – are similar or not. To the human brain, this notion may seem trivial because the intended use, construction and behaviours of these structures are vastly different. However, for a computer, automatically measuring these dissimilarities requires a whole host of information. In this paper, the aim is to use irreducible-element models and attributed graphs to represent engineering structures, and to use graph kernels to measure the similarity of these models. The proposed methods are able to compare discrete and continuous attributes of structures in polynomial time. Similarity assessments are provided for a group of toy structures as well as a case study of seven real operational bridges. The latter population is important in dealing with a class of highly complex real-world examples of civil infrastructure; the analysis also allows a discussion on which aspects of bridge construction might be responsible for structural similarity or dissimilarity.