Eintrag in der Universitätsbibliographie der TU Chemnitz
Volltext zugänglich unter
URN: urn:nbn:de:bsz:ch1-qucosa2-969172
Pareek, Kaushal Arun
Wunderle, Bernhard ; Bailey, Christopher (Gutachter)
Physics-Informed Synthetic Data Generation for Deep Learning-Based Defect Detection in Infrared Thermography
Kurzfassung in englisch
The vision of a deep learning-empowered, non-destructive evaluation technique aligns with zero-hour quality control and preventive maintenance goals. Integrating deep learning algorithms and non-destructive evaluation techniques can enhance defect detection and characterization accuracy, automate analysis processes, reduce inspection time and costs, and gain valuable insights into manufacturing processes; however, this integration process is data-driven, which poses a challenge, especially in manufacturing where there is an inherent dearth of data let alone the need for a labeled dataset. This work presents a framework to bridge the data gap in manufacturing by leveraging the prior knowledge of the underlying physics and domain-specific knowledge to aid the application of data-driven deep learning algorithms. The non-destructive technique under consideration is pulsed infrared thermography. This work begins with an extensive introduction to infrared thermography as a non-destructive testing technique, followed by the basics of deep learning. Later, results from several infrared thermography post-processing techniques are evaluated based on the ease of finding the images with the optimal information on the sub-surface defects, which will be the input for the deep learning algorithm ahead. The work then explores developing a finite-element-based simulation framework to generate synthetic pulsed thermography datasets. The results from the finite-element simulations are evaluated against the experimental setup to validate its fidelity, and solutions are provided to the computational and parameter estimation challenges in the development of a simulation model. The calibrated finite-element simulation model is then scaled to generate a large amount of synthetic pulsed thermography data and is post-processed based on the earlier findings. A novel noise addition technique is introduced to improve the representativeness of the synthetic dataset. The representativeness of synthetic data is thoroughly investigated at various steps of the framework, and the image segmentation model is trained separately on experimental and synthetic datasets. The research reveals that synthetic datasets represent the experimental data well when carefully rendered. When evaluated on experimental samples, the segmentation model pre-trained on synthetic datasets generalizes well to the experimental samples. Furthermore, the advantage of the suggested framework for physics-based synthetic dataset generation is the ease of labeling large amounts of data. The effectiveness of the suggested approach is demonstrated by evaluating the models on two new datasets: one where the complete setup is changed and the other being an open-source infrared thermography dataset.
| Universität: | Technische Universität Chemnitz | |
| Institut: | Professur Werkstoffe und Zuverlässigkeit mikrotechnischer Systeme | |
| Fakultät: | Fakultät für Elektrotechnik und Informationstechnik | |
| Dokumentart: | Dissertation | |
| Betreuer: | Wunderle, Bernhard | |
| DOI: | doi:10.60687/2025-0079 | |
| SWD-Schlagwörter: | Infrarotthermografie , Zerstörungsfreie Werkstoffprüfung , Finite-Elemente-Methode , Künstliche Intelligenz , Maschinelles Sehen , Zuverlässigkeit | |
| Freie Schlagwörter (Englisch): | Flaw detection , deep learning , data augmentation , pre-training , zero defect manufacturing , inline inspection , synthetic data , finite element method , image segmentation , infrared thermography , non-destructive testing | |
| DDC-Sachgruppe: | 006.32, 006.37, 621.362, 620.1127, 004.0151, 620.00452 | |
| Sprache: | englisch | |
| Tag der mündlichen Prüfung | 11.04.2025 | |
| OA-Lizenz | CC BY 4.0 |
