解盘石, 邢军军, 段思宇, 等.基于视觉检测及面向边缘计算的液压支架姿态模型研究[J].河南理工大学学报（自然科学版）,2026,45(2):107-119.

XIE P S, XING J J, DUAN S Y, et al. Research on a vision-based hydraulic support pose model for edge computing[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(2):107-119.

基于视觉检测及面向边缘计算的液压支架姿态模型研究

解盘石1,2,3, 邢军军1,2,3, 段思宇1,2,3, 李国欣1,2,3

1.西安科技大学 能源与矿业工程学院，陕西 西安  710054；2.西安科技大学 西部矿井开采及灾害防治教育部重点实验室，陕西 西安  710054；3.陕西省岩层控制重点实验室， 陕西 西安  710054

摘要: 目的 液压支架稳定性控制是大倾角煤层开采所面临的主要技术难题，受地质条件和采场环境变化影响，支架易处于非稳定的工作状态。 方法 采用改进的YOLOv11n-Pose算法训练支架姿态检测模型，实现支架整体姿态检测。为满足模型轻量化需求，在算法的主干层引入GhostNet网络结构和RepConv模块；设计C3k2-RVB-EMA注意力机制模块增加模型的鲁棒性；为提高模型的整体精度与收敛效率，对比常用的损失函数CIoU，GIoU，DIoU，EIoU，SIoU和WIoU，选择WIoU为该模型的损失函数。基于双目视觉模型和坐标转换关系，实现对支架支撑高度、顶梁俯仰角及底座偏转角的数据测量。 结果 结果表明，模型采用轻量化设计，比YOLOv11n-Pose算法模型体积减少30.26%；采用C3k2-RVB-EMA注意力机制模块和WIoU损失函数，模型准确率提升1.7%；改进算法模型生成的支架姿态曲线、关键点位置和输出的置信度数值均优于现存算法模型；在支架姿态测量实验中，所提模型测量结果最大相对误差在5.52%以内。将所提模型部署至边缘计算设备平台上，其参数量、计算量、体积和推理时间均优于常用模型。 结论 改进算法模型用于大倾角煤层支架整体姿态检测，展现出较高的准确性和鲁棒性；同时将模型部署至边缘计算设备上，实现了高效推理，有效兼顾了检测精度与轻量化部署需求，为大倾角煤层液压支架姿态检测提供了可行技术路径。

关键词:液压支架;姿态检测;YOLOv11n-Pose;轻量化模型;模型部署

doi:10.16186/j.cnki.1673-9787.2025070018

基金项目:国家自然科学基金资助项目（52174126）；陕西省杰出青年科学基金资助项目（2023-JC-JQ-42）

收稿日期:2025/07/17

修回日期:2025/11/10

出版日期:2026/01/28

Research on a vision-based hydraulic support pose model for edge computing

Xie Panshi1,2,3, Xing Junjun1,2,3, Duan Siyu1,2,3, Li Guoxin1,2,3

1.College of Energy and Mining Engineering， Xi’ an University of Science and Technology， Xi’ an  710054， Shaanxi， China；2.Key Laboratory of Western Mine Exploitation and Hazard Prevention Ministry of Education， Xi’an University of Science and Technology， Xi’an  710054， Shaanxi， China；3.Shaanxi Key Laboratory of Ground Control， Xi'an  710054， Shaanxi， China

Abstract: Objectives Stability control of hydraulic supports is a major technical challenge in steeply inclined coal seam mining. Influenced by variable geological conditions and changes in the mining environment， supports are prone to operating in unstable states.  Methods An improved YOLOv11n-Pose algorithm was employed to train the hydraulic support pose detection model， achieving holistic posture detection. To meet requirements for model lightweighting， the GhostNet network structure and RepConv modules were introduced into the algorithm's backbone. A self-designed C3k2-RVB-EMA attention mechanism module was incorporated to enhance model robustness. To improve overall model accuracy and convergence efficiency， a comparative analysis of prevalent loss functions—namely CIoU， GIoU， DIoU， EIoU， SIoU， and WIoU—was conducted. Based on this comparison， WIoU （Wise-IoU） was selected as the loss function for the model. Based on a binocular vision model and coordinate transformation relationships， data measurement for prop height， canopy pitch angle， and base deflection angle was achieved. Results The results indicate that the lightweight design reduced the model size by 30.26% compared to the original YOLOv11n-Pose algorithm model. The incorporation of the C3k2-RVB-EMA attention mechanism module and the WIoU loss function increased model accuracy by 1.7%. The improved algorithm model outperformed the existing algorithm models in generating support posture curves， keypoint positions， and output confidence values. In the posture measurement experiments， the maximum relative error of the proposed model's measurement results was within 5.52%. When deployed on an edge computing device platform， the proposed model exhibited advantages over common models in terms of parameter count， computational load， size， and inference time.  Conclusions The proposed improved algorithm model demonstrates high accuracy and robustness for holistic posture detection of supports in steeply inclined coal seams. Deploying the model on edge computing devices enables efficient inference， effectively balancing detection accuracy with lightweight deployment requirements. This provides a feasible technical pathway for posture detection of hydraulic supports in steeply inclined coal seams.

Key words: hydraulic supports; posture detection;YOLOv11n-Pose;lightweight model; model deployment