A Pull-Winding Production Line Broken-Yarn Defect Detection System Combining Feature Enhancement and Deep Learning

ABSTRACT Broken-yarn detection in pull-winding and pultrusion processes still relies mainly on manual inspection, whereas yarn overlap, blurred defect edges, and vibration-induced background disturbances hinder reliable visual analysis. To address these problems, a broken-yarn defect detection method combining feature enhancement, three-frame differencing, and deep-learning recognition is proposed. A Gaussian-curvature-based image scaling method is used to preserve fine geometric textures, and a DGSWLTLBO-Otsu segmentation algorithm is introduced to improve broken-yarn region extraction under complex backgrounds. Based on enhanced differential images, a DF-Half-AlexNet model is developed for broken-yarn recognition. Experimental results show that the proposed method improves edge-detail preservation and broken-yarn region extraction, and the DF-Half-AlexNet model achieves an accuracy of 97.15 % ± 0.21 %. Vibration experiments further demonstrate its recognition stability under interference conditions.