Accurate segmentation and quantitative evaluation of Cf/SiC fiber fracture defects using an enhanced deep learning method

Abstract

The complex preparation process and demanding operating conditions of ceramic matrix composites (CMCs) frequently result in fiber fracture defects, posing significant safety risks. Accurate characterization of these defects and evaluation of their impact on the mechanical properties of CMCs are crucial. X-ray computed tomography, a widely used nondestructive testing method for CMCs, faces challenges in accurately segmenting and quantifying fiber fracture defects due to their complex spatial structures and low grayscale contrast in large datasets. This paper proposes a Transformer-based deep neural network for segmenting fiber fracture defects. By incorporating a semantic enhancement module in the decoder, the model achieves accurate defect segmentation, outperforming existing image segmentation networks while reducing computational costs. Three-dimensional visualization and quantitative analysis of the defects helped clarify the failure mechanisms of CMCs. In addition, mechanical tests reveal a progressive decline in both tensile and compressive properties with aggravating defects. The final retention rates of tensile and compressive strength are 60.65 % and 57.38 %, respectively, compared with defect-free samples. Fiber fracture defects alter the material's fracture surface direction and microstructure, inducing delamination and cracks. The proposed method offers valuable insights for the intelligent nondestructive evaluation of CMC components with fiber fracture defects.