Vibration-based damage localization in composites structures: integrating CNNs, mode shapes, and digital image correlation

Abstract

Composite sandwich structures offer high stiffness-to-weight ratios but are highly susceptible to complex internal damage mechanisms that are often difficult to detect using conventional inspection techniques. This work proposes a vibration-based Structural Health Monitoring (SHM) framework that integrates mode shape analysis, image processing techniques, Digital Image Correlation (DIC), and Convolutional Neural Networks (CNNs) for automated damage detection, localization, and sizing in composite sandwich structures. The investigated damage consists of a controlled stiffness-reduction region introduced by inserting a lower-stiffness material within the laminate, simulating internal debonding or delamination-type defects. A simulation-driven methodology is first established, in which mode shapes obtained from finite element models are transformed into image-based representations. These representations are used to train CNN models, demonstrating that image-based modal features outperform traditional approaches relying solely on global modal parameters, achieving detection accuracies above 90% under numerical conditions. The framework is further enhanced through physically informed image transformations, including residual-based representations, curvature enhancement, and multi-modal strategies that combine information from multiple vibration modes, as well as attention mechanisms embedded within the CNN architecture. These improvements lead to consistent performance gains, with accuracy increases of up to 10–15% and reductions of up to 20–30% in damage localization error compared to baseline image representations. Finally, the proposed framework is validated experimentally using mode shapes extracted from DIC measurements. Despite measurement noise, experimental uncertainty, and discrepancies between numerical and experimental domains, the models maintain robust performance, achieving damage detection accuracies above 85%, average localization errors below 10 mm, and reliable estimation of damage size under real testing conditions. The main contribution of this thesis is to demonstrate that vibration mode shapes, interpreted as spatial information, combined with advanced image representations and deep learning architectures, enable accurate localization and sizing of internal stiffness-reduction damage in real composite sandwich structures. These findings highlight the practical feasibility of the proposed methodology for real-world SHM applications.