An improved automatic image labeling and classification algorithm for multi-mode damage quantification of 2.5D woven composites based on deep learning strategy

Accurately identifying and quantifying the complex multi-mode damages in woven composites is of vital importance to evaluate the service life and improve reliability of the components. However, the current advanced methods based on the deep learning framework remain mainly the manual labeling, resulting in unclear fiber/resin interfaces, easily-overlooked microcracks, and lower efficiency. To overcome the problem, this paper proposes an improved automatic image labeling and classification algorithm based on deep learning strategy to quantify the uncertainty damages of 2.5D woven composites. In detail, the original micro-computed tomography (CT) images are automatically labeled by an image algorithm that utilizes grayscale values and image boundaries to produce image datasets. Subsequently, the DCNN model is trained using the image datasets. Then, the trained deep convolutional neural networks (DCNN) model is used to identify unseen CT images and separate the damage and different sub-phases of 2.5D woven composites. Finally, the connected component analysis is introduced to classify the global cracks at the meso-scale. The results show that the proposed automatic image labeling and classification algorithm can achieve a damage identification precision of 85.87 %, surpassing that of other models. Moreover, the multi-mode damages of 2.5D woven composites are accurately captured. In the warp direction, the bending damage accumulation predominantly manifests as interface debonding, representing 51.93 % of the damage percentage. In the weft direction, it is primarily characterized by matrix cracking, representing 60.98 % of the damage percentage. It is expected that the study can provide data support for the application of large-scale and complex structural components.