Improvement of 3D-Active Thermography by using Artificial Intelligence

Abstract

For the one-dimensional case of heat propagation in active thermography the thickness of the investigated specimen can be directly reconstructed using known evaluation methods such as laser flash analysis or thermographic signal reconstruction and observation of a characteristic time. Concerning the multidimensional case diffusive effects have a strong impact on the heat propagation of the thermal wave, which leads to misinterpretations when evaluating the thickness especially at deep-lying edges and inhomogeneities in a specimen. The deeper a defect is located in the component, the more diffuse it is perceived in a change in surface temperature. Within this paper, heat flux simulations and real measurements of pulse thermography of different defect geometries are used to train a neural network for depth-resolved defect interpretation. The defect geometries are based on geometries of real impact damages in composites and were realistically obtained from a micromechanical fracture simulation. The neural network is based on an encoder-decoder approach where the temperature values of the cooling curve after a pulse-shaped excitation serve as input information. Segmentation is performed as a function of the backwall geometry. By training several thousand defect geometries using an encoder-decoder network, it was possible for the first time to directly infer the backwall geometry of a component without additional information about the component. Finally, it is shown how simulations can support the inversion of thermal waves for 3D-thermography of real measurements by an artificial intelligence system.