Multi-view deep learning for CT-based damage quantification in 3D woven composites

Accurate identification and quantification of damage in 3D woven composites (3DWC) are essential for understanding damage evolution and predicting mechanical performance. However, the morphological characteristics of cracks vary significantly across different viewing directions, leading to weaker identification accuracy when relying on a single view. To address this, we propose a multi-view deep learning approach that identifies cracks from three mutually orthogonal planes, thereby improving segmentation accuracy. In addition, a data-driven predictive analysis based on quantitative crack features is conducted to estimate shear stiffness degradation. To apply and evaluate the proposed method in practice, we conducted a stepwise experimental program as described below. Initially, a stepwise out-of-plane shear test was performed to enable progressive CT data acquisition during deformation. Subsequently, the proposed approach was adopted to automatically identify cracks from CT data. Based on damage visualization and quantitative analysis, interfacial cracking was identified as the dominant damage mechanism, and the shear failure process was classified into three distinct stages. Finally, CT-derived crack descriptors were used to predict shear stiffness degradation. Comparative evaluation with different regression models showed that the CT-derived crack descriptors contain useful predictive information for stiffness degradation, and that Gradient Boosting provided the best predictive performance under the present limited-data condition.