On the application of domain adaptation for aiding supervised SHM methods

The lack of available damage state data is a significant challenge within the field of Structural Health Monitoring (SHM). When data are obtainable for a given system of interest, a variety of machine learning approaches have been successful in addressing a range of supervised SHM problems. However, these methods assume that the training and testing data sets are drawn from the same distribution; as a consequence damage state data must be collected for each new structure and/or damage scenario considered, which is often infeasible and/or not economically viable. In these contexts it is useful to transfer knowledge obtained from known damage state data to different, but related contexts (or domains) of interest. By utilising transfer learning, knowledge obtained from different structures and/or damage scenarios can be used to improve learners in various target domains. Domain adaptation, a subcategory of transfer learning, is concerned with scenarios where the data distributions across source and target domains are different; and is demonstrated here to be applicable to SHM in a numerical and experimental case study.