Synthetic Data Generation Using Finite Element Method to Pre-Train an Image Segmentation Model for Defect Detection Using Infrared Thermography

The vision of a deep learning-empowered non-destructive evaluation technique aligns perfectly with the goal of zero-defect manufacturing, enabling manufacturers to detect and repair defects actively. However, the dearth of data in manufacturing is one of the biggest obstacles to realizing an intelligent defect detection system. This work presents a framework for bridging the data gap in manufacturing using the potential of synthetic datasets generated using the finite element method-based digital twin. The non-destructive technique under consideration is pulse infrared thermography. A large number of synthetic thermographic measurements were generated using 2D axisymmetric transient thermal simulations. The representativeness of synthetic data was thoroughly investigated at various steps of the framework, and the image segmentation model was trained separately on experimental and synthetic datasets. The study results reveal that when carefully rendered, synthetic datasets represent the experimental data well. When evaluated on real-world experimental samples, the segmentation model pre-trained on synthetic datasets generalizes well to the experimental samples. Furthermore, another advantage of synthetic datasets is the ease of labelling a large amount of data. Finally, the robustness assessment of the model was done on two new datasets: one where the complete experimental setup was changed, and the other was an open-source infrared thermography dataset