[2B3] AI-based defect reconstruction in lock-in thermography
M Kreutzbruck, J Rittmann and J Hufert
University of Stuttgart, Germany
In most NDT methods, the inverse problem must be solved to identify a material defect. The measurement data should be used to draw conclusions about the underlying defect in the component in terms of defect size, defect location and defect depth. There are currently various approaches to improving imaging in active thermography. Artificial intelligence (AI) is one of them and represents a future solution with great potential, which can support even inexperienced personnel in data evaluation and defect detection. Convolutional neural networks (CNNs) are often used for this purpose to analyse thermograms. Among them, U-Net is a special network architecture based on CNNs. Cross-connections between the encoder and decoder side of the network allow for high mapping accuracy, which even can be achieved with small training datasets. The encoder gradually reduces the input image to a minimum and extracts the relevant features. The subsequent decoder links the extracted features with each other and scales the information back to the original size using additional layers. This makes it possible to quickly and reliably recognise and segment impact damage in a small dataset of lock-in thermographic phase images of carbon fibre-reinforced polymer (CFRP). The recorded cooling curves are processed pixel by pixel and the depths and diameters of flat-bottomed holes in CFRP sheets are classified and quantified. The local component thickness up to the damage surface or the component thickness up to the backwall geometry is used as the output variable. This article demonstrates that a comparatively small dataset consisting of just a few hundred training images is sufficient to precisely perform an automated evaluation of 2.5D geometries of the rear wall or of damage surfaces.
