An integrated autoencoder-based framework for unsupervised domain adaptation in structural health monitoring of bridges

The limited availability of long-term bridge-monitoring data presents a significant obstacle to the application of machine learning in structural health monitoring (SHM). Domain adaptation offers a means of addressing this data scarcity, but existing methods often have limited generalization and are sensitive to hyperparameter choices. This study proposes a reconstruction-based domain adaptation framework in which target-domain observations are geometrically aligned with the latent space learned by an autoencoder trained on data from a source bridge and, subsequently, reconstructed using its decoder. Unlike conventional feature-alignment approaches, the proposed formulation transfers the reconstruction model itself, enabling target-domain evaluation without labels, classifier retraining, or threshold recalibration. The autoencoder is trained on baseline data collected during commissioning of the source bridge, using observations composed of natural frequencies as damage-sensitive features. Adaptation is performed through orthogonal Procrustes alignment in the latent space, followed by reconstruction with the source-trained decoder. Validation using the Z24 Bridge dataset demonstrates that the framework mitigates seasonal domain shift while preserving sensitivity to controlled damage. The approach is then transferred to a twin bridge over the Itacaiúnas River using only latent-space alignment and no model retraining. The results demonstrate that the proposed framework enables damage assessment across bridges while substantially reducing the data and training requirements for practical deployment.