输电塔作为电力输送网络的关键基础设施，其结构安全性直接关系到电网的稳定运行。在长期服役过程中，输电塔螺栓受风荷载、温差效应及材料老化等多因素耦合作用，易逐渐发生松动。该文提出了一种基于改进YOLOv5s的输电塔螺栓松动智能检测模型（CCSGS-YOLO）：采用坐标卷积替代主干网络中的标准卷积层，增强模型对目标位置信息的获取能力；引入卷积注意力模块（CBAM），通过通道与空间双重注意力机制，强化模型在复杂背景下的特征鉴别能力；构建Slim-Neck特征融合结构，借助跨阶段部分连接与深度可分离卷积的优化组合，在维持检测精度的同时降低计算复杂度；采用GIoU损失函数与Soft-NMS的联合优化策略，通过考虑预测框与真实框的重叠几何特性，提升目标检测的定位精度。实验结果表明：CCSGS-YOLO的精确率达91.7%，召回率为89.4%，平均精度均值达到95.3%，F₁分数提升至90.0%，较基准模型YOLOv5s分别提高了1.6、3.0、1.4和1.0个百分点；在计算效率方面，CCSGS-YOLO模型检测速度达74.8 f/s，推理时延降低至13.4 ms，较YOLOv5s模型提升11.6%。此外，该文通过现场实验验证了CCSGS-YOLO在不同场景下的检测鲁棒性，为输电塔螺栓松动的智能巡检提供了一种新思路。

As the critical infrastructure in power transmission networks, the structural stability of transmission towers directly impacts the safe and reliable operation of the power grid. During long-term service, bolts in the tower structure are prone to gradual loosening under the coupled effects of multiple factors such as wind loads, temperature variations, and material aging. This paper proposed an intelligent detection model for bolt looseness in transmission towers based on an improved YOLOv5s (named CCSGS-YOLO). The model incorporates several key enhancements: coordinate convolution replaces standard convolutional layers in the backbone network to strengthen the model’s ability to capture positional information of targets; a convolutional block attention module (CBAM) is introduced to strengthen the model’s feature discrimination capability in complex backgrounds through dual channel and spatial attention mechanisms; a slim-neck feature fusion architecture is constructed, leveraging an optimized combination of cross-stage partial connections and depthwise separable convolutions to reduce computational complexity while maintaining detection accuracy; a joint optimization strategy employing the Generalized Intersection over Union (GIoU) loss function and Soft Non-Maximum Suppression (Soft-NMS) improves localization accuracy by considering the geometric overlap characteristics between predicted and ground-truth bounding boxes. Experimental results demonstrate that CCSGS-YOLO achieves a precision of 91.7%, a recall of 89.4%, a mean average precision (mAP) of 95.3%, and an F₁ score of 90.0%. These metrics represent improvements of 1.6, 3.0, 1.4, and 1.0 percentage points, respectively, over the baseline YOLOv5s model. In terms of computational efficiency, the model achieves a detection speed of 74.8 frames per second (FPS), reducing the inference latency to 13.4 ms, which represents an 11.6% improvement compared to the YOLOv5s model. Furthermore, this paper validates the detection robustness of CCSGS-YOLO across various scenarios through field experiments, providing a novel approach for intelligent inspection of loose bolts on transmission towers.