On the application of kernelised Bayesian transfer learning to population-based structural health monitoring

Data-driven approaches to Structural Health Monitoring (SHM) generally suffer from a lack of available health-state data. In particular, for most structures, it is not possible to obtain a comprehensive set of labelled damage data – even covering the most common damage types – due to impracticalities and economic considerations in observing the structure in a range of damage states. One solution to this problem is to utilise labelled data from a set of ‘similar’ structures. The assumption is that, as a population, the group may have a shared label set that covers a wider range of damage states, which can be used in labelling a different structure of interest. These goals, producing a model that generalises for a population of structures, and transferring label information between structures, are part of a population-based view of SHM — known as population-based SHM (PBSHM). By considering data from a population, it is possible to make data-driven SHM practical in industrial contexts beyond unsupervised learning, i.e. novelty detection. In order to realise the potential of PBSHM, this paper applies a heterogeneous transfer learning method – kernelised Bayesian transfer learning (KBTL) – which is a sparse Bayesian method that infers a discriminative classifier from inconsistent and heterogeneous feature data, i.e. the dataset from each member of the population may refer to different quantities in different dimensions. The technique infers a shared latent space where data from each member of the population are mapped on top of each other, meaning a single classifier can jointly be inferred that generalises to the complete population. As a consequence, label information can be transferred in this shared latent space between members of the population. The ability to infer a mapping from inconsistent and heterogeneous feature data make the approach a heterogeneous transfer learning method. To the best of the authors knowledge, this is the first time a heterogeneous transfer learning method has been applied in an SHM context. •For the first time heterogeneous transfer learning is performed in SHM.•PBSHM is applicable to populations with heterogeneous, inconsistent feature spaces.•KBTL is benchmarked on several populations of experimental and numerical structures.