Damage detection and localization in composite plates using mode shapes and artificial intelligence: from numerical models to experimental validation

This study presents a simulation-driven deep learning framework for vibration-based damage detection and localization in composite structures using mode shape images. CNNs are trained exclusively on numerically simulated data and directly validated against experimental measurements, addressing the domain gap between numerical models and real-world structural behavior without requiring experimentally labeled data for training. The proposed approach is evaluated on the challenging task of detecting and localizing small damages in a large composite plate, with defect-to-structure area ratios as low as $$0.004%.$$ A binary classification model is first developed to distinguish between damaged and undamaged conditions, achieving 100% accuracy on simulated data and 90% accuracy on experimental data. In addition, regression-based CNN models are employed to estimate damage location and size under two distinct boundary conditions: free–free and fully clamped. The models demonstrate consistent and accurate localization performance across both scenarios, with improved results in the free–free condition. Despite being trained solely on simulated data, the models generalize effectively to experimental mode shapes, capturing spatial damage features and providing physically meaningful predictions. Overall, the results demonstrate that simulation-driven CNN frameworks can achieve robust cross-domain performance. This significantly reduces the dependence on costly experimental datasets and highlights the practical potential of the proposed approach for real-world Structural Health Monitoring applications in composite structures.