Abstract: Electromagnetic clutch is an important part in the automobile production process. Aiming at the problems of small defect size of its adsorption surface,complex background texture and existing algorithms can not achieve the diversity of defects,a lightweight target detection algorithm based on improved YOLOv8n was proposed. EMA attention and partial convolution were integrated in the backbone network,and a lightweight C2F-PE module was designed to improve the C2F structure and to enhance the feature extraction ability of the network. In order to promote richer feature fusion between the same scales,the Attention-based Intra-scale Feature Interaction (AIFI) module was introduced to replace the SPPF layer to capture more fine-grained information.A small object detection layer was added to the neck network,which effectively fused the shallow feature information and improved the model perception of small objects. The Slim-neck module was introduced to improve the neck network,which lightened the model while maintaining the detection accuracy of the network. The experimental results showed that compared with the YOLOv8n algorithm,the improved algorithm achieved an mAP@0.5 of 94.6 %,which was an increase of 4.5 %. The number of parameters was reduced by 13.3 %,and the detection speed reached 81 f/s. The algorithm effectively balanced detection accuracy and speed,meeting the needs for real-time detection in electromagnetic clutch production.

Key words: YOLOv8n, electromagnetic clutch, defect detection, lightweight network, EMA attention, intra-scale feature interaction, Slim-neck module