| 张清正,张仕海,屈重年,郭晓赛.融合RT-DETR及插值法的井口控制盘仪表参数识别[J].电子测量与仪器学报,2026,40(3):231-239 |
| 融合RT-DETR及插值法的井口控制盘仪表参数识别 |
| Parameter identification for wellhead control panel instruments byintegrating RT-DETR and interpolation method |
| |
| DOI： |
| 中文关键词: 海上平台 井口控制盘 RT-DETR模型 样条插值 目标识别 参数识别 |
| 英文关键词:offshore platform wellhead control panel RT-DETR model spline interpolation target recognition parameter identification |
| 基金项目:天津市自然科学基金（22JCYBJC01640）项目资助 |
|
|
| Author | Institution |
| Zhang Qingzheng | School of Mechanical Engineering, Tianjin University of Technology and Education, Tianjin 300222, China |
| Zhang Shihai | School of Mechanical Engineering, Tianjin University of Technology and Education, Tianjin 300222, China |
| Qu Chongnian | School of Mechanical Engineering, Tianjin University of Technology and Education, Tianjin 300222, China |
| Guo Xiaosai | School of Mechanical Engineering, Tianjin University of Technology and Education, Tianjin 300222, China |
|
| 摘要点击次数: 333 |
| 全文下载次数: 202 |
| 中文摘要: |
| 针对海上平台井口控制盘仪表智能巡检系统，提出了一种基于轻量化RT-DETR模型与插值法的仪表参数识别方法。为实现模型轻量化及目标高精度提取,重构并优化了RT-DETR的主干与颈部网络，通过引入跨阶段局部光照增强模块（CSP-IEM）与快速增强混合聚合模块（FEMAM）提升光照鲁棒性和颈部网络效率，同时设计了匹配感知损失函数（MAL）保留了匹配质量信息。消融实验表明,相比于原始的RT-DETR算法,改进后的RT-DETR算法在海上平台井口控制盘仪表数据集上的mAP@0.5达到了76.1%,参数量与计算量分别减少了30.73%和25.79%,帧率（FPS）达到了216 fps。基于仪表关键目标识别及图象处理方法，提出采用样条插值法计算仪表读数的方法。现场采集图像仪表参数的读取实验表明,所提出的样条插值法相对于局部角度法的平均相对误差与平均全局误差分别相对降低了45.5%与48.3%。综合实验证明，所提出的仪表参数识别方法满足现场计算资源受限条件下的部署，检测精度及实时性等方面的需求。 |
| 英文摘要: |
| A parameter identification method based on a lightweight RT-DETR model and a spline interpolation method is proposed for the instrument intelligent inspection system developed for the offshore platform wellhead control panel. To achieve lightweight model and high-precision targets extraction, the backbone and neck networks of RT-DETR are reconstructed and optimized. The cross-stage local illumination enhancement module (CSP-IEM) and the fast enhanced hybrid aggregation module (FEMAM) are introduced to improve illumination robustness and neck network detection efficiency. A matching-aware loss function (MAL) is designed to preserve matching quality information. The ablation experimental results show that the model achieves a mean average precision (mAP) of 76.1% at 0.5, reduces the number of parameters and computation by 30.73% and 25.79%, respectively, and achieves a frame rate of 216 fps. Based on the key targets recognition and image processing methods for instruments, a method using spline interpolation to calculate instrument readings is proposed. The experiment of reading instrument parameters from on-site collected images shows that the spline interpolation method reduces the mean relative error and mean global error by 45.5% and 48.3%, respectively compared to the local angle method. Comprehensive experiments have demonstrated that the proposed instrument parameter identification method meets the requirements for deployment under limited on-site computing resources, as well as the needs for detection accuracy and real-time performance. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
