Assessing the Data-Centric Robustness of a Crack Tip Detection Model through Artificial Data Pollution

Abstract

Ensuring the early detection of fatigue cracks is critical for maintaining the structural integrity of components in high-risk environments such as aerospace engineering. Optical methods that measure material deformation without physical contact have made it possible to track crack development with high spatial precision. Recent advances in automated analysis tools based on machine learning have enabled efficient interpretation of deformation data to identify crack tips. However, real world experiments often yield imperfect input images due to sensor noise, motion blur, or other artifacts, which may reduce the reliability of these algorithms. This thesis investigates how the accuracy of a crack detection model is affected by such degradation in data quality. Controlled distortions were applied to input images at varying intensity levels, and the Performance of the model was evaluated based on its ability to localize the crack tip under each condition. The results reveal which forms of degradation are most detrimental, at which threshold the algorithm begins to fail and how does training with noisy inputs improve performance. The study concludes with practical recommendations for improving the reliability of crack detection tools when used in imperfect laboratory environments, thereby contributing to more robust fracture analysis in aerospace applications.