Synthetic training data for crack propagation measurement by neural networks

Abstract Crack initiation and propagation are of central importance in experimental mechanics, structural assessment, and process analysis. However, conventional optical deformation analysis techniques have thus far yielded limited accuracy in capturing these phenomena and are generally not suitable for real-time control applications due to the high necessary computation time. This study investigates the application of a fully convolutional network (FCN), originally developed for biomedical image segmentation, to the task of detecting crack initiation and tracking propagation. Supervised deep learning approaches typically require extensive, pixel-level annotated datasets, which are time-consuming and prone to human error. To address this limitation, a method is introduced for generating synthetic image data using a physics-inspired crack propagation model combined with data augmentation techniques. The FCN is trained and validated entirely on the synthetic dataset. Its performance is then evaluated using real experimental data from a metal shearing test. Results demonstrate that the FCN can accurately identify both crack initiation and propagation. This deep learning-based approach shows strong potential to enhance existing optical deformation analysis methods, particularly in the context of material failure assessment or even in real-time process control systems.