Autoencoder-Assisted Domain Adaptation via Procrustes-Based Latent Alignment for Structural Health Monitoring
Wellington De lima nogueira, S. D. da Silva, Eloi Figueiredo
TL;DR: A framework that combines unsupervised learning and domain adaptation to enhance model transferability under limited data, reducing dependence on labeled datasets while preserving sensitivity to structural and operational changes is proposed.
Abstract: The scarcity of long-term vibration data real-world structures remains a significant barrier to the application of machine learning in structural health monitoring (SHM). Available datasets are often short, unlabeled, and affected by operational and environmental variability, limiting the generalization of data-driven models. This paper proposes a framework that combines unsupervised learning and domain adaptation to enhance model transferability under limited data. An autoencoder is trained on commissioning data from a source structure to extract latent features that represent key dynamic characteristics. These embeddings act as unsupervised, damage-sensitive indicators. To adapt across domains, two complementary strategies are introduced: (i) latent-space realignment using the Procrustes method for geometric alignment, and (ii) selective decoder retraining with unlabeled target data. This enables efficient adaptation without full retraining. The framework is validated on the Z24 bridge benchmark, successfully adapting to seasonal and damage variations using only output data, and later transferred to the New Bridge, one of the twin bridges over the Itacaíunas River in Brazil. Results show robust latent-space alignment, accurate reconstruction, and effective generalization under domain shifts. Overall, the method offers a computationally efficient approach for transfer learning in SHM, reducing dependence on labeled datasets while preserving sensitivity to structural and operational changes.