Background/Introduction <p>Composite sandwich structures are widely used in aerospace, marine, and civil engineering applications due to their highstrength-to-weight and stiffness-to-weight ratios. However, these materials are susceptible to barely visible damage,which can significantly compromise structural integrity and is difficult to detect using conventional inspection methods.This challenge has motivated the development of advanced Structural Health Monitoring (SHM) techniques.</p> Purpose <p>This study aims to develop and validate an intelligent SHM framework capable of detecting, classifying, and localizingdifferent types of damage (skin, interface, and core) in composite sandwich panels by combining vibration-based modeshape analysis with Convolutional Neural Networks (CNNs).</p> Methods <p>Finite element models of composite sandwich panels were developed to simulate undamaged conditions and threedamage scenarios (skin, interface, and core) under four different boundary conditions. The first ten non-null mode shapeswere extracted and transformed into contour plot images. These images were used to train multi-head CNN architecturesfor damage classification and damage localization. Measurement noise was introduced to the modal data to emulaterealistic operational conditions. Bayesian optimization was employed to select optimal hyperparameters for each network.</p> Results <p>The classification network achieved nearly 100% accuracy for undamaged and core damage cases, while skin andinterface damage were identified with approximately 80% accuracy, even in the presence of noise. The damagelocalization network achieved an average mean absolute error of 10–12 mm, corresponding to a relative error ofapproximately 4–5% with respect to the panel dimensions. Boundary conditions with partial constraints (CFFF and CCFF)yielded the most accurate classification and localization results.</p> Conclusions <p>The proposed methodology demonstrates strong robustness and accuracy in detecting, classifying, and localizingdamage in composite sandwich panels using mode shapes and CNNs. While damage position can be reliably predicted,estimating damage characteristics such as size and orientation remains challenging due to problem complexity. Overall,the results highlight the potential of integrating vibration-based SHM with deep learning techniques for reliable andautomated damage assessment in composite structures under realistic noise conditions.</p>