Convolutional neural network for automated surface crack detection in inductive thermography

This work presents a convolutional neural network (CNN), trained on simulated data and used for the detection of cracks resulted by inductive thermography measurements. In inductive thermography the sample under study is heated with a short heating pulse and an infrared (IR) camera records the emitted surface radiation during both heating and cooling. The recorded IR sequence is then evaluated to a phase image using Fourier transform. In phase images, short surface cracks become visible due to the hot spots around the defect tips and due to the low phase value along the crack line. For the training of a deep neural network many images are necessary, which should be different from the images to be evaluated. This is why FEM simulations have been carried out varying crack length, depth and inclination angle. Additional Gaussian noise and augmentation have been added to these simulated images before using them to train a CNN. Samples with real cracks along a weld have been created in Inconel 718, and the CNN, trained on the simulation results, has been used for semantic segmentation of these real samples' phase images, in order to identify the defects. Additionally, the samples were investigated by computer tomography, and this 3D information of the cracks is compared to the phase image results.