Transfer Learning of Multi-Fidelity Models for Damage Identification: A Case Study of Laboratory Steel Frames

Vibration-based monitoring through Operational Modal Analysis (OMA) is one of the most widely adopted approaches in Structural Health Monitoring (SHM), as it allows the extraction of modal parameters from ambient vibrations in a non-intrusive and cost-effective way. When a finite element (FE) model is available, these identified modal data can also be used to update it, enabling variations in model parameters to be related to damage. However, model-based schemes generally require a high-fidelity (HF) numerical model, which is often unavailable in practice because design or assessment models include simplifications, such as idealized boundary conditions, rigid joints, or lumped masses, that limit their predictive capability. To address this limitation, this work proposes a two-stage multi-fidelity transfer learning framework for zero-shot structural damage detection. A dual-encoder neural network is first pre-trained on HF simulations from a source structure and then fine-tuned on calibrated low-fidelity (LF) simulations of the target structure using differential learning rates. The framework performs floor-level damage localisation and enables damage classification from frequency drop signatures without requiring experimental training data. Validation on experimental measurements from two damage scenarios on a five-storey steel frame benchmark shows 99.3% zero-shot floor localisation accuracy. Moreover, t-SNE visualisation of the frequency encoder’s learned embeddings reveals distinct clusters for the two experimental damage classes, despite the absence of damage class labels during training.