Speaker
Description
In Germany, more than 130,000 m³ of low- and medium-level radioactive waste – approximately 90% of the national inventory – is stored in 200-liter steel drums at interim storage facilities. At present, the integrity of these drums is assessed primarily through manual visual inspection of the outer surface. While this approach can identify surface-visible corrosion, it cannot detect corrosion caused internal material loss threatening the integrity of the drum and is inherently limited in terms of objectivity and information depth.
This study investigates infrared thermography as a non-contact remote non-destructive testing (NDT) method for the detection of internal defects in metallic radioactive waste drums. In practical inspection scenarios, thermographic measurements are strongly influenced by surface-related artifacts, including scratches, dirt, labels, multiple paint layers with low thermal conductivity, and the curvature of the drum surface. These factors distort heat propagation and reduce defect contrast, significantly limiting the effectiveness of conventional thermographic post-processing.
To overcome these limitations, advanced thermal signal processing methods – specifically principal component thermography (PCT), pulse phase thermography (PPT), and thermal signal reconstruction (TSR) – are combined with machine learning techniques to enhance defect detectability under realistic conditions. Instead of relying on individual post-processing outputs, multiple thermographic representations are used jointly to extract complementary spatial and temporal features that are more robust to surface artifacts.
Different machine learning strategies are investigated for defect identification and segmentation. Classical decision-tree-based methods, including decision trees and random forests, are evaluated using feature vectors derived from processed thermographic data. In parallel, several neural network architectures are explored, ranging from shallow convolutional neural networks with a limited number of layers to more advanced encoder–decoder architectures such as U-Nets. These models are trained to exploit spatial context and multi-channel thermographic inputs obtained from combined post-processing methods. The results demonstrate that integrating thermographic signal processing with data-driven learning improves the visibility and separability of defect-related signals across samples with varying surface conditions, compared to thermographic post-processing alone.
The study highlights the potential of combining infrared thermography with machine learning to extend the capabilities of current inspection practices beyond purely visual assessment. The presented approach provides a foundation for more reliable thermographic evaluation of radioactive waste drums and supports the development of automated NDT workflows for challenging industrial inspection scenarios.
