Deep learning-based real-time detection of fiber anomalies in the three-dimensional braiding process

Authors:
Publisher website
Google Scholar
Abstract
Three-dimensional (3D) braided composites have considerable application potential in aerospace, defense, transportation, and marine engineering owing to their excellent structural integrity and mechanical properties. However, during the 3D braiding process, yarns are susceptible to fluctuations in creel tension, wear of yarn-guiding components, and variations in process parameters, causing fiber anomalies such as fiber breakage and fuzzing. Owing to their random and abrupt occurrence, these anomalies are difficult to identify accurately in real time using conventional inspection methods, thereby affecting process stability and potentially compromising preform quality. To address these challenges, this study proposes YOLOv13-SFC, a deep-learning-based fiber anomaly detection model for the real-time detection of fiber breakage and fuzzing during the 3D braiding of composite preforms. A DySample-based adaptive upsampling mechanism is introduced into the neck to enhance adaptive sampling during multi-scale feature fusion. A DSC3k2-FSAS module is proposed to synergistically integrate local convolutional representations with frequency-domain feature interactions within local patches, thereby enhancing the discriminative representation capability of the network for fine-grained textures, small anomalous regions, and complex backgrounds. Furthermore, a lightweight detection head, termed Detect-DyHead, is developed to improve the dynamic modeling and prediction of anomalies of varying scales and morphologies. Experimental results show that YOLOv13-SFC achieves an mAP50 of 92.1%, 2.7 percentage points higher than the YOLOv13n baseline. Meanwhile, the parameter count and computational cost are reduced by 8.03% and 22.58%, respectively, while the model-level inference speed reaches 49.8 frames per second. Based on the proposed model, a vision-based detection platform integrating YOLOv13-SFC with a UR5 collaborative robot is developed for online monitoring during the 3D braiding process. The end-to-end processing latency of the developed platform is 101–126 ms. Continuous-braiding experiments further demonstrate that, at a preform linear traction speed of 300 mm min −1 , the developed system is capable of online monitoring of fiber anomalies.
Funding Information
  • Natural Science Foundation of Shandong Province (ZR2022ME185)

This publication has 30 references indexed in Scilit: