Speaker
Description
This study evaluates the effectiveness of an Automated Defect Recognition (ADR) system for aeronautical composite structures through the integration of active infrared thermography and Deep Learning–based image analysis. The research aims to assess not only the performance of the computational model but also the reliability of the inspection technique itself when applied to diverse and increasingly complex composite material configurations. Given the extensive use of composite materials in modern aircraft structures, ensuring the robustness and repeatability of automated non-destructive evaluation (NDE) systems across different laminate architectures is essential for guaranteeing structural integrity and operational safety. The experimental campaign focused on carbon fiber reinforced polymer (CFRP) laminates, as well as hybrid composite configurations incorporating fiberglass layers and metallic lightning protection foils. These hybrid structures introduce additional electrothermal and optical complexities that challenge conventional inspection and automated analysis methods. To establish a comprehensive and controlled testing baseline, both internal and surface discontinuities were intentionally introduced into the specimens. These defects were simulated using foreign object inclusions embedded between laminate plies and precision machining techniques applied at predefined depths and geometries. Data acquisition was performed using active infrared thermography in both reflection and transmission modes, allowing the assessment of defect detectability under different heat flow conditions. This dual-mode approach enabled a more complete evaluation of thermal contrast variations associated with subsurface defects, particularly in the presence of heterogeneous material interfaces and metallic layers. For the computational analysis, a U-Net convolutional neural network architecture was employed to perform semantic segmentation of defects. The model was trained and validated on a dataset comprising approximately 3,000 thermographic images, collected from the different laminate configurations and inspection setups. Special emphasis was placed on evaluating the robustness of the network in maintaining spatial accuracy and precise defect boundary delineation despite the variability in thermal signatures caused by distinct material combinations and structural arrangements. The results, quantified using the Intersection over Union (IoU) metric, demonstrated a high segmentation efficiency of 97.72% for standard CFRP laminates. Importantly, the system maintained strong performance when applied to hybrid composite structures, achieving an IoU of 93.98%, along with a Recall of 98.86% and an F1-Score of 96.90%. These results indicate that the proposed ADR system is capable of accurately identifying and localizing defects even under challenging electrothermal interference conditions introduced by metallic foils. Overall, the findings validate the effectiveness of combining active thermography with Deep Learning–based semantic segmentation for automated defect detection in aeronautical composites. The demonstrated robustness across multiple laminate architectures supports the potential standardization of this integrated NDE approach for aerospace inspection applications.
