赵伟, 张文康, 刘德成,等.基于YOLOv5的瓦斯抽采钻杆智能计数研究[J].河南理工大学学报（自然科学版）,2025,44(3):81-88.

ZHAO W, ZHANG W K, LIU D C, et al. Research on intelligent counting of gas extraction drill rods based on YOLOv5[J]. Journal of Henan Polytechnic University(Natural Science) , 2025, 44(3): 81-88.

基于YOLOv5的瓦斯抽采钻杆智能计数研究

赵伟1, 张文康1, 刘德成1, 王涛1, 王登科2, 夏代林3, 周礼赞3, 李志飞3

1.河南龙宇能源股份有限公司，河南 永城  476600；2.河南理工大学 河南省瓦斯地质与瓦斯治理重点实验室，河南 焦作  454000；3.武汉天宸伟业物探科技有限公司，湖北 武汉  430070

摘要: 目的 随着煤矿开采逐渐深入，瓦斯抽采作业的安全风险不断增加，钻杆计数的准确性对保障瓦斯抽采的安全和效率至关重要。传统的钻杆计数方法效率低、易出错，且难以适应复杂环境。  方法 本文基于YOLOv5深度学习模型，提出了一种结合时空信息融合的钻杆智能计数方法，通过实时处理煤矿井下的视频数据，实现钻杆的自动检测与计数。实验数据包括7组28段模拟退钻场景和10组真实退钻场景。为了增强模型的鲁棒性，采用过亮、过暗、烟尘噪声和镜像等数据增强技术。此外，对YOLOv5的不同版本（YOLOv5s、YOLOv5m、YOLOv5l、YOLOv5x）进行比较，选择最优模型进行钻杆计数。计数过程中，结合钻杆面积跳变、IOU跳变等前置更新条件，进一步优化钻杆数量更新的准确性。  结果 结果表明，4种YOLOv5模型在当前数据中均达到了99.5%的准确率和100%的召回率，因此采用YOLOv5s作为后续计数检测模型。  结论 本文方法在模拟和真实退钻场景下均实现了100%的正确计数，展现出较高的准确性和鲁棒性。通过减少人工干预，显著提高了钻杆计数的自动化水平，具有广泛的应用前景，尤其在煤矿安全生产及其他工业自动化监控领域具有重要意义。

关键词:钻杆计数;目标检测;时空融合;深度学习;瓦斯抽采

doi:10.16186/j.cnki.1673-9787.2024060009

基金项目:国家自然科学基金资助项目（52174174）

收稿日期:2024/06/06

修回日期:2014/12/30

出版日期:2025-04-18

Research on intelligent counting of gas extraction drill rods based on YOLOv5

ZHAO Wei1, ZHANG Wenkang1, LIU Decheng1, WANG Tao1, WANG Dengke2, XIA Dailin3, ZHOU Lizan3, LI Zhifei3

1.Henan Longyu Energy Co.， Ltd.， Yongcheng  476600， Henan， China；2.State Key Laboratory Cultivation Base for Gas Geology and Gas Control， Henan Polytechnic University， Jiaozuo  454000， Henan， China；3.Wuhan Tensense Geotech Co.， Ltd.， Wuhan  430070， Hubei， China

Abstract: Objectives With the increasing depth of coal mining operations， the safety risks associated with gas extraction have become more prominent. Accurate counting of drill rods is essential for ensuring the safety and efficiency of gas extraction. Traditional methods are often inefficient， error-prone， and struggle to perform reliably in complex underground environments.  Methods This study proposes an intelligent drill rod counting method based on the YOLOv5 deep learning model， enhanced by spatiotemporal information fusion. The method processes underground video data in real time to achieve automatic detection and counting of drill rods. The dataset consists of 28 simulated drill withdrawal scenarios across 7 groups and 10 real-world withdrawal scenarios. To improve model robustness， data augmentation techniques such as overexposure， underexposure， smoke interference， and mirroring were employed. Four YOLOv5 variants （YOLOv5s， YOLOv5m， YOLOv5l， YOLOv5x） were compared to identify the most suitable model. Additionally， drill rod count updates were optimized by incorporating features such as abrupt changes in area and Intersection over Union （IoU） values.  Results Experimental results demonstrated that all four YOLOv5 models achieved an accuracy of 99.5% and a recall rate of 100% on the dataset. YOLOv5s was selected for subsequent use due to its balance of accuracy and computational efficiency.  Conclusions The proposed method achieved 100% correct counting in both simulated and real drill withdrawal scenarios， demonstrating excellent accuracy and robustness. By minimizing manual intervention， it significantly enhances the automation of drill rod counting and shows strong potential for application in coal mine safety and other industrial automation monitoring fields.

Key words: drill rod counting; object detection; spatiotemporal fusion; deep learning; gas extraction