To address the issues of missed and false detections in industrial safety device target detection within factory environments, the SAG-YOLOv8 improved architecture is proposed. This architecture replaces the conventional convolution in the original C2f module with shifted convolution, forming the new C2f-SWC module to enhance multi-scale feature representation. Additionally, the AIFI module is employed instead of the traditional spatial pyramid pooling structure to strengthen semantic understanding. The GFPN network architecture is introduced to improve small target feature propagation by enhancing cross-layer multi-scale interactions and same-layer lateral connections. Experimental results show that the SAG-YOLOv8 algorithm improves the mAP@0.5 metric by 3.4% compared to the original YOLOv8, with noticeable enhancements in precision and recall. This method significantly enhances the accuracy and stability of safety device target detection in chemical plants, providing strong technical support for safe operations.