Study on visual inspection of fiber anomaly in 3D fabricating of composite prefabricated body

Abstract The three-dimensional weaving process for composite preforms is widely used in aerospace, national defense, and high-end equipment manufacturing, where mechanical properties depend strongly on fiber orientation and structural uniformity. During weaving, fibers are susceptible to abrasion, fracture, and knotting caused by bending, interlacing, and friction. These defects are typically subtle and random, making traditional inspection methods inadequate for real-time, non-destructive monitoring and ultimately degrading performance and mechanical properties. This paper presents a deep learning approach for the detection of abnormal fiber states. The BasicBlock network is restructured using a SAConv module to improve feature extraction. The integration of a RepC3-CBAM module into the ResNet18 backbone facilitates the enhanced capture of local details pertaining to small-scale fiber anomalies and improves attention to critical regions, all while preserving a lightweight network architecture. An SCConv module is incorporated into the detection head to facilitate lightweight feature reconstruction and enhance multi-scale feature representation. A robotic vision inspection platform, AUBO-i5, is developed using the proposed method, facilitating real-time monitoring of fiber anomalies in the weaving process. Experimental results indicate that the proposed model enhances mAP50 by 7.2%, decreases parameter count by 23.7%, and reduces FLOPs by 27.8%, while increasing detection speed from 74 frames per second (f/s) to 183 f s −1 . The results indicate that the proposed approach attains an optimal balance between accuracy and efficiency. The model’s effectiveness in complex industrial weaving environments is validated, offering reliable technical support for enhancing the structural uniformity and mechanical properties of three-dimensional woven fiber preforms.