Towards deep learning fusion of flying spot thermography and visible inspection for surface cracks detection on metallic materials
Vers une fusion d'apprentissage en profondeur de la thermographie par points volants et de l'inspection visible pour détection de fissures superficielles sur des matériaux métalliques
Résumé
"Flying spot" laser infrared thermography (FST) is a non destructive testing technique able to detect small defects through scanning surfaces with a laser heat source. Defects such as cracks can indeed be detected by the disturbance of heat propagation measured by an infrared camera. However this examination method is limited to small regions of interest and the measurement might be affected by heterogeneous surface properties. Moreover, visible spectrum enables the localisation of variations of properties on the surface and an inspection within large field of view in a single snapshot. However, such inspection could miss cracks with small dimension due to resolution issue. Deep learning is now a very efficient processing approach, first to automatically analyse and exploit context information from data, and second to conduct data fusion. Hence, we propose here to develop a new crack inspection method fusing FST and visible spectrum using deep learning. This paper presents our preliminary work towards this fusion. We first focus on IR spectrum, starting on bench settings optimization in comparison with the theory of Peclet number, for both simulated and experimental data. Then we present FST automated defect detection results using two different neural network architectures and also study the sensitivity of their performances to the physical theory of Peclet number. Finally, we analyse the defect detection results on both IR and visible spectrum separately using deep learning, to illustrate the potential of fusing these spectra.
Origine | Fichiers produits par l'(les) auteur(s) |
---|