Speaker
Description
Improving aircraft availability while ensuring structural integrity has made the optimization of aeronautical maintenance a major industrial challenge. Current inspection methods do not reliably detect the most critical defects and are often too time-consuming to significantly reduce aircraft ground time.
Since 2018, an extensive experimental campaign has been conducted on lightning-struck composite plates protected by various lightning strike protection systems. These specimens were inspected using flash infrared thermography to detect subsurface defects, primarily delaminations. The resulting database, built over several years, includes a wide variety of composite materials (thermoset and thermoplastic matrices), different resin systems, and configurations with or without protective coatings.
In continuity with recent studies carried out at ONERA, the objective of this work is to contribute to the integration of artificial intelligence techniques to enhance infrared image processing algorithms developed at ONERA. The focus is placed on the detection of delamination defects in composite materials through the implementation of unsupervised learning algorithms, with the long-term objective of automating non-destructive testing (NDT) and enabling its future robotization.
Most segmentation approaches reported in the literature rely on supervised deep learning architectures such as U-Net and primarily exploit spatial information. However, strong thermal contrasts not related to delamination can lead to false detections. To overcome this limitation, the present study focuses on the temporal thermal signal rather than spatial supervision.
Two types of unsupervised algorithms are investigated: a one-dimensional Variational Autoencoder (1D VAE) and diffusion-based models. The 1D VAE is designed to model the temporal dynamics of each pixel, allowing the learning of normal thermal behavior while remaining sensitive to deviations induced by subsurface defects. Diffusion-based models are the State-of-the-Art approach for image generation, which iteratively learn to denoise random samples for signal/image synthesis. The capability of these models to capture complex data distribution makes them promising in the context of infrared thermal physics. Consequently, they are explored for the task of anomaly detection on thermographic images.
These algorithms were first validated on a semi-analytical synthetic database generated from quadrupole simulations. Virtual infrared sequences were created for 2 mm-thick composite plates, including both sound areas and centrally located defects at various depths and signal-to-noise ratios. The methods were then applied to experimental infrared thermography sequences acquired after lightning strike impacts. The experimental database contains more than one hundred infrared sequences.
The performance evaluation focuses on detection-oriented criteria rather than classical segmentation accuracy. Although the F1-score is adopted to quantify the ability of the proposed methods to distinguish defective from sound areas, commonly used segmentation metrics such as Intersection over Union are deliberately not considered. Indeed, IoU is poorly suited to infrared thermography data, as the reference annotations are affected by heat diffusion after thermal excitation and therefore do not accurately represent the physical extent of the defects. By relying on the F1-score, the evaluation framework emphasizes detection reliability while reducing sensitivity to thermally induced spatial spreading, enabling a more meaningful comparison of the proposed unsupervised approaches.
