Automatic detection of fatigue cracks using neural networks and digital image correlation

Abstract

Fatigue cracks are an inherent part in the lightweight design of engineering structures subjected to non-constant loads. Particularly important for airframe structures are accurate design data for crack initiation, stable fatigue crack propagation (FCP) and its rapid increase until ultimate failure. Non-straight crack paths are difficult and time-consuming to detect and monitor in laboratory experiments as well as in service using traditional techniques such as direct current potential drop (DCPD) or dye penetrant inspection. Machine learning based models like Convolutional Neural Networks (CNNs) led to enormous breakthroughs in classical computer vision. Recently, they are finding their way into experimental mechanics. However, depth and complexity of these neural networks make it difficult to understand how they work. Nevertheless, explainability is a crucial prerequisite to increase acceptance amongst human experts and to justify usage in safety relevant applications. In this work, we implemented a deep CNN to detect crack paths and especially their crack tips based on full-field displacement data obtained by 3D digital image correlation (DIC). Therefore, fatigue crack propagation experiments were performed for AA2024-T3 rolled sheet material. During the experiments, several hundred datasets were acquired by DIC and labelled by optical analysis. A part of the data was then used to train the neural network. The results show that the method can accurately detect the shape and evolution of cracks in all specimens based on the x and y displacements. Moreover, we use the state-of-the-art Seg-Grad-CAM interpretability method to compute network attention heatmaps explaining the focus and reasoning of the networks. The method results in the discovery of a robust and stable model (ParallelNets), which learned the ability to focus on the physical crack tip field in order to find precise crack tip positions.