Challenges for SHM from structural repairs : an outlier-informed domain adaptation approach

Data-based approaches to structural health monitoring are typically constructed on the assumption that the underlying data distributions in training will be the same as those experienced when the method is deployed. However, structural repairs alter the physical properties of the system, leading to a change in structural response. This change in response leads to a shift in the data distributions from the pre- to post-repair states—known as domain shift—invalidating the assumption that training, and subsequent operational data come from the same underlying distribution. As a result, structural repairs represent a significant challenge to data-based approaches to structural health monitoring (SHM). Not only will domain shift cause an algorithm trained on the pre-repair data to fail to generalise, but it will also make labels acquired from the pre-repair state redundant for building conventional data-based methods on the post-repair data. Transfer learning, in the form of domain adaptation, provides a solution to this problem, allowing knowledge from the pre-repair labels to be transferred to the post-repair dataset by forming a shared latent space where the pre- and post-repair dataset distributions are approximately equal. This paper presents a novel modification of a domain adaptation technique—joint domain adaptation—in creating outlier-informed joint domain adaptation, which can be used in transferring knowledge from pre- to post-repair states, forming a post-repair classifier that utilises all the pre-repair knowledge and generalises to post-repair data. The algorithm is demonstrated on an experimental dataset from a Gnat aircraft wing, where it is shown to outperform conventional data-based approaches and existing domain adaptation techniques.