Bridging Data-Based and Model-Based SHM via Transfer Learning for Efficient Damage Detection and Localisation

Motivated by the statistical pattern-recognition paradigm, where Structural health monitoring (SHM) operators need to know in advance which damage scenarios are of interest, we propose a framework that combines data- and model-based SHM via transfer learning through domain adaptation. Here, a finite-element (FE) model generates labelled source-domain data for a monitored structure, and domain adaptation is then used to train a classifier for the monitored target structure over a set of discrete damage locations. This step is particularly challenging in practical SHM because the source domain typically covers all relevant states, whereas the target domain contains only two: healthy and one (a priori unknown) damage state. That makes it a partial domain adaptation problem. We benchmark established domain adaptation methods (transfer component analysis, joint domain adaptation, balanced domain adaptation) against dedicated partial domain adaptation approaches (normal condition alignment, normal correlation alignment) and test whether combinations of these methods further improve performance. The results show that the partial domain adaptation approaches consistently achieve the best performance in terms of accuracy and F1-score. To use the data-driven classifier’s localisation output to guide model updating, the FE model is updated only in the localised region to estimate the exact damage location. The approach is demonstrated on a cantilever beam with an added mass at varying positions, achieving both detection and localisation. This framework enables SHM practitioners to roughly localise the damage and, based on the identified region, decide whether model updating is warranted or whether alternative actions should be taken.