A Transfer Learning approach for damage identification in operational viaducts

A major limitation in data-driven Structural Health Monitoring is the scarcity of labeled data for training machine learning models. Transfer Learning addresses this by enabling knowledge sharing across similar structures, reducing datasets distribution shift. This study proposes a novel Transfer Learning framework for damage identification in operational viaducts with similar spans, using modal frequencies as damage-sensitive features. Domain Adaptation is performed via Normal Condition Alignment, to map source and target features in a shared latent space. A baseline normal condition is established on source features through a linear regression model. Gaussian Mixture Models are trained on source residuals, and used to detect anomalies in the target domain, based on residual distributions. A real viaduct for which long-term monitoring data are available is used as a case study. The structure comprises two homogeneous datasets collected on the deck of similar spans. Source data pertain to a deck with extensive measurements, whereas target data refer to a second deck with a reduced dataset, due to sensor malfunctions. Damage is simulated in the target dataset by reducing the measured frequencies. Validation using data from real damaged scenarios will enable future scaling of the proposed framework to operational conditions, providing a practical tool for data-driven SHM of viaducts, enabling damage detection in under-instrumented areas by leveraging data from other spans.