罗伟立1 陈贵龙2 刘桂雄2 陈建勋3
（1.珠海市安粤科技有限公司，广东 珠海 519080
2.华南理工大学机械与汽车工程学院，广东 广州 510640
3.广东省特种设备检测研究院珠海检测院，广东 珠海 519002）
摘要：针对人工审查电梯检验视频耗费人力物力，且难以保证客观性、一致性的问题，提出基于YOLOv11n的电梯检验视频关键要素合规性审查方法。该方法通过正交实验设计原则构建初始训练数据集，并引入半监督学习策略减少人工标注成本、提升模型泛化能力；针对不同关键要素特性，设计差异化合规性审查算法，实现检测、审查全流程自动化。实验结果表明，该方法在图像级目标检测中P=0.9586、R=0.9783、mAP@50=0.9795、mAP@50:95=0.825 6；在视频级合规审查中P=Sp=1.00，确保审查过程无假阳性样本。该方法不仅在检测精度上具有优势，还同时兼顾了实时性、轻量化，为电梯检验视频关键要素合规性审查提供切实可行的技术路径。
关键词：半监督学习；YOLOv11n模型；电梯检验；合规性审查；目标检测
中图分类号：TP391.4; TQ171.1; TU857 文献标志码：A 文章编号：1674-2605(2025)05-0007-09
DOI：10.12475/aie.20250507 开放获取
Application of Semi-supervised YOLO in Compliance Review of Elevator Inspection Videos
LUO Weili1 CHEN Guilong2 LIU Guixiong2 CHEN Jianxun3
(1.Zhuhai Anyue Technology Co., Ltd., Zhuhai 519080, China2.School of Mechanical and Automotive Engineering, South China University of Technology, Guangzhou510640,China
3.Zhuhai Branch, Guangdong Institute of Special Equipment Inspection and Research, Zhuhai 519002, China )
Abstract: Addressing the issues of high labor and resource costs associated with the manual review of elevator inspection videos, as well as the difficulty in ensuring objectivity and consistency, this paper proposes an automatic compliance review method for key elements in elevator inspection videos based on YOLOv11n. The method constructs an initial training dataset following the principles of orthogonal experimental design and introduces a semi-supervised learning strategy to reduce manual annotation costs and enhance model generalization capability. Differentiated compliance review algorithms are designed according to the characteristics of various key elements, achieving full automation of the detection and review pipeline. Experimental results indicate that the proposed method achieveP=0.9586、R=0.9783、mAP@50=0.9795、mAP@50:95=0.825 6 in image-level object detection; and in video-level compliance review,P=Sp=1.00, ensuring a review process with zero false positives. The method not only demonstrates advantages in detection accuracy but also balances real-time performance and a lightweight design, providing a practical and feasible technical pathway for the automatic compliance review of key elements in elevator inspection videos.
Keywords: semi-supervised learning; YOLOv11n model; elevator inspection; compliance review; object detection
