In order to address the critical issue of defect detection in narrow gap welding, an improved YOLOv8 model was used to investigate the enhancement of weld defect detection accuracy and efficiency. The results show that by introducing the RepViT-CBAM structure to replace the downsampling module in the YOLOv8 backbone network, the feature extraction capability was significantly improved, resulting in a 3.9% increase in mean average precision (mAP) with only a 1.3% increase in computation. Additionally, the integration of the shuffle attention(SA) module enhanced the recall rate for small pores and cracks, while the use of the WIoU loss function optimized the defect localization accuracy. It is concluded that the proposed method achieves an mAP of 96.47% on mixed self-built and open-source datasets, surpassing the original YOLOv8s model by 4.55% with a frame rate of 114.59 frames/s, making it suitable for real-time weld defect detection in industrial applications.