On the Application of Partial Domain Adaptation for PBSHM

To address limitations presented by the unavailability of labelled data in structural health monitoring, transfer learning – in the form of domain adaptation – can facilitate leveraging information from a population of physical or numerical structures, by inferring a mapping that aligns the feature spaces. Thus, data from a similar source structure can be used to train a classifier that generalises to a target structure. Conventional unsupervised domain adaptation methods would assume that the data available from both structures belongs to the same number of damage-states or disparate environmental conditions; however, in practical scenarios, the data in the target structure may pertain to a subset of the available classes in the source structure; this is a partial domain adaptation problem. Conventional domain adaptation methods are prone to performance degradation in this scenario. To address this issue, this paper proposes a novel statistic alignment method and instance-weighting strategy. A numerical population of structures demonstrates that these methods facilitate transfer where a number of state-of-the-art domain adaptation algorithms cannot.