Transfer learning for population-based SHM

Structural health monitoring (SHM) systems aim to proactively identify damage and provide diagnostic information to support maintenance decisions in mechanical, aerospace, and civil infrastructure. A critical challenge for the application of SHM systems -- particularly those that provide contextual information -- is the feasibility and cost of acquiring comprehensive data. Population-based SHM (PBSHM) presents a potential solution by leveraging data from related structures. However, differences between structures often prevent conventional machine learning models from generalising across domains. This issue motivates the use of transfer learning, which seeks to improve predictive performance in a target domain using data from a related source domain.

In PBSHM, target structures will often only have data for a limited range of health states. Therefore, to enable transfer when target labels are sparse, this thesis presents novel statistic alignment (SA) methods that require only undamaged target data. These methods are shown to facilitate the generalisation of models learnt using only labelled source data.

Quantifying similarity between structures and their features is essential to ensure that transfer learning will yield positive results. This thesis investigates using physics knowledge to address limitations with data-based similarity measures in sparse-data scenarios. This approach is incorporated into a feature-selection criterion to identify transferable, damage-sensitive features. Subsequently, it is used within a regression framework to predict the quality of predictions when transferring between a specific source/target pair, supporting decisions about when transfer is appropriate.

Previous work has not considered how to incorporate transfer learning into an online framework that updates as labels are collected during a monitoring campaign. Thus, a Bayesian model is proposed that uses the SA methods to define mappings early in the monitoring campaign and updates sequentially as labels are obtained. This model is integrated into an active-sampling strategy that guides inspections by selecting the most informative observations to label.