In response to the issues of low efficiency, susceptibility to missed and false detections in traditional manual inspection of spherical shut-off valve bolts, as well as the insufficient capability of standard object detection algorithms in identifying subtle defects, this paper proposes an improved YOLOv5sbased model for bolt state detection. The model enhances feature focus on critical regions by incorporating the convolutional block attention module (CBAM） attention mechanism, optimizes detail preservation in multi-scale feature fusion through the content-aware ReAssembly of features (CARAFE） upsampling operator, and employs depthwise separable convolution (DWConv） to achieve lightweight network design. These improvements collectively enhance the identification accuracy of four bolt states: normal, loose, dislodged and offset. Experimental results on a dataset collected from hydropower plant sites show that the proposed model achieves mAP@0.5 of 93.54%, representing a 2.91% improvement over the original YOLOv5s model. Furthermore, it outperforms mainstream detection algorithms such as faster regions with convolutional neural networks (Faster R-CNN） and single shot MultiBox detector (SSD） in overall performance, demonstrating its capability to perform accurate and efficient abnormal state detection.