许焱平, 陈华天, 王满利,等.基于改进YOLOv8n的矿井下钻机退钻检测算法[J].河南理工大学学报（自然科学版）,2026,45(3):59-68.

XU Y P, CHEN H T, WANG M L, et al.Detection algorithm for underground mine drilling rig withdrawal based on improved YOLOv8n[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(3):59-68.

基于改进YOLOv8n的矿井下钻机退钻检测算法

许焱平1, 陈华天1, 王满利1, 张善文2, 郭妮3

1.河南理工大学 物理与电子信息学院 河南 焦作  454003；2.西京大学 电子信息学院 陕西 西安  710000；3.北京联合大学 应用科技学院 北京  100000

摘要:目的 为了实现对钻杆的精确计数，提出一种改进YOLOv8n的矿井下钻机退钻检测算法YOLOv8_SDI。  方法 首先，针对成像环境复杂、图像特征信息较少问题，引入通道注意力模块SEAttention，该模块利用自学习方式基于特征通道的重要程度赋予权重值，提高算法关注重要信息的能力，提升丰富信息通道的特征、抑制无用信息通道特征，克服YOLOv8n边缘特征信息提取较差的问题；其次，使用C2f_DBB替换网络中C2f模块，该模块通过转化多尺度分支，增加网络的空间变形适应性、丰富特征空间的多样性，进一步提高模型的特征表达力，增加模型的泛化性能；最后，为解决低质量图像的梯度损失大和收敛慢问题，引入Inner_IoU损失函数，通过动态调整辅助锚框，根据评估回归状态进行梯度调整，克服原损失函数CIoU在梯度传递方面的缺陷，提高锚框预测精度，加快网络收敛。  结果 分别采用2种算法在同一退钻数据集上验证，YOLOv8n模型平均检测精度为97.1%，精确度为95.2%，召回率为94.5%；YOLOv8n_SDI模型平均检测精度为98%，精确度为96.4%，召回率为96.1%。  结论 相比于YOLOv8n算法，改进后的YOLOv8n_SDI算法退钻检测效果更好，在工程应用方面具有一定价值。

关键词:深度学习;目标检测;注意力机制;钻机退钻;钻杆识别

doi:10.16186/j.cnki.1673-9787.2024090049

基金项目:国家自然科学基金资助项目（62172338）；河南省科技攻关项目（242102221006）；北京市教委科技发展基金项目（12213611611）

收稿日期:2025/02/25

修回日期:2025/05/03

出版日期:2026/04/28

Detection algorithm for underground mine drilling rig withdrawal based on improved YOLOv8n

Xu Yanping1, Chen Huatian1, Wang Manli1, Zhang Shanwen2, Guo Ni3

1.School of Physics & Electronic Information Engineering， Henan Polytechnic University， Jiaozuo  454003， Henan， China；2.School of Electronic Information， Xijing University， Xi’an  710000， Shaanxi， China；3.School of Applied Science and Technology， Beijing Union University， Beijing  100000， China

Abstract: Objectives To eliminate explosion hazards and achieve accurate counting of drill pipes， an improved YOLOv8_SDI drilling rig withdrawal detection algorithm was proposed.  Methods First，to address the issues of a complex imaging environment and limited feature information in the images，the channel attention module SEAttention was introduced. This module employs a self-learning method to assign weight values based on the importance of feature channels，thereby enhancing the algorithm's ability to focus on critical information. Features of information-rich channels were enhanced，while features of irrelevant information channels were suppressed， effectively overcoming the poor extraction of edge features by YOLOv8n. Second，the C2f module in the network was replaced by C2f_DBB，which increases the spatial deformation adaptability of the network， enriches the diversity of the feature space，and further improves the feature expression of the model. The generalization performance of the model is enhanced through the transformation of multi-scale branches. Finally，to address the issues of large gradient loss and slow convergence associated with low-quality images，the Inner_IoU loss function was introduced. The auxiliary anchor box is dynamically adjusted，and the gradient is modified according to the evaluated regression state to overcome the deficiencies of the original CIoU loss function in gradient transfer，thereby improving the prediction accuracy of the anchor box and accelerating the network convergence.  Results The two algorithms were verified on the same drilling withdrawal dataset. The YOLOv8n model  achieved a mean average precision of 97.1%， an accuracy of 95.2%， and a recall of 94.5%. In contrast，the YOLOv8n_SDI model achieved a mean average precision of 98%， an accuracy of 96.4%， and a recall of 96.1%.  Conclusions Compared with the YOLOv8n algorithm， the improved YOLOv8n_SDI algorithm demonstrates better performance in drilling  withdrawal detection， showcasing its value in engineering applications.

Key words:deep learning;object detection;attention mechanism;drill pipe withdrawal;drill rod identification