3D Deep Learning Segmentation for Fiber Break Analysis of Carbon Fiber Reinforced Polymer Tomograms

Abstract

Carbon fiber reinforced polymers (CFRPs) find extensive use in modern aerospace structures due to their adaptability in stiffness and strength for specialized applications. However, their heterogeneous composition and the microscopic scale of fiber and matrix lead to complex damage mechanisms, as well as make failure difficult to predict, slowing the progress of CFRP adoption over traditional engineering materials such as metals. X-ray computed tomography allows for non-destructive, volumetric imaging of CFRPs under stress, enabling real-time 3D observation of the material’s failure. Due to the data-rich nature of the 3D scans at each time step, such experiments can result in thousands of 2D images per scan and multiple scans at different time- or loading-steps per test. This accumulates to hundreds of thousands of images following a typical test campaign. Human analysis, of these scans is time- and resource-intensive, creating the need for an automated way to segment and analyze these images. Recent work has shown that the application of 2D convolutional deep learning models to the identification of damage types in CFRP yields accuracy levels exceeding those of humans and requires a fraction of the working time. However, similar research with deep learning models applied to medical images (MRIs, X-rays, etc.) has noted 3D convolution as strictly better and is now the standard.

Here, a 2D vs. 3D deep learning model comparison of the segmentation of carbon fiber breaks, an imbalanced classification problem with less than 0.01% of the data being fiber breaks of interest, shows overall similar performance between 2D and 3D segmentation (e.g., IoU scores of 67.5% and 70.7%, respectively). Qualitative and quantitative analysis reveals that the 3D model has the ability to embed the third dimension of spatial information, such that 3D segmentation is evaluated as improved over 2D for the fiber break problem, which is also desirable for future applications in composite damage segmentation beyond fiber breaks, suggesting 3D will be strictly better in the composite damage classification problem space.