Vibration-based damage identification in concrete beams via a decomposed deep learning approach

This work proposes a hybrid data-driven strategy for structural damage identification based on dynamic structural features, in which the training data are generated through a physics-based numerical model. A cohesive fracture model is employed to simulate a wide range of damage scenarios by systematically varying both damage location and severity. This numerically generated dataset is conceived as a practical alternative to expensive and time-consuming experimental campaigns, while retaining full control over damage parameters and boundary conditions. From each simulated case, dynamic features are extracted, including modal frequencies and curvature-based indicators derived from the corresponding mode shapes. These features are used to train a deep learning framework aimed at estimating the damage state in terms of both localization and severity. Damage localization is formulated as a classification task over discrete positions along the beam, whereas damage severity is estimated through a complementary regression stage, enabling the framework to output both where damage occurs and how severe it is. The proposed approach is assessed on unseen test data and additional user-defined scenarios, demonstrating accurate damage localization and reliable severity estimation across a broad range of simulated conditions. Overall, the results highlight the effectiveness of combining physics-based dataset generation with deep learning to support vibration-based damage assessment in beam-like structures and provide a flexible foundation for future extensions to more complex structural configurations and experimental validation.