On the application of domain adaptation in structural health monitoring
Paul Gardner, X. Liu, Keith Worden
Abstract
The application of machine learning within Structural Health Monitoring (SHM) has been widely successful in a variety of applications. However, most techniques are built upon the assumption that both training and test data were drawn from the same underlying distribution. This fact means that unless test data were obtained from the same system in the same operating conditions, the machine learning inferences from the training data will not provide accurate predictions when applied to the test data. Therefore, to train a robust predictor conventionally, new training data and labels must be recollected for every new structure considered, which is significantly expensive and often impossible in an SHM context. Transfer learning, in the form of domain adaptation, offers a novel solution to these problems by providing a method for mapping feature and label distributions for different structures, labelled source and unlabelled target structures, onto the same space. As a result, classifiers trained on a labelled structure in the source domain will generalise to a different unlabelled target structure. Furthermore, a holistic discussion of contexts in which domain adaptation is applicable are discussed, specifically for population-based SHM. Three domain adaptation techniques are demonstrated on four case studies providing new frameworks for approaching the problem of SHM.