石油化工高等学校学报 ›› 2026, Vol. 39 ›› Issue (5): 65-73.DOI: 10.12422/j.issn.1006-396X.2026.05.008

• 石油工程 • 上一篇    下一篇

基于目标检测与模糊度判断的油藏注水悬浮颗粒物在线检测方法及其应用

林晨1(), 刘铭辉1(), 王新乐1, 王妮1, 毛邓添1,2   

  1. 1.中海油能源发展股份有限公司 安全环保分公司,天津 300450
    2.安睿人工智能创新研发中心,天津 300450
  • 收稿日期:2026-03-01 修回日期:2026-03-17 出版日期:2026-10-25 发布日期:2026-11-02
  • 通讯作者: 刘铭辉
  • 作者简介:林晨(1993-),男,硕士,工程师,从事油气田污染物治理与环境监测方面的研究;E⁃mail:252588829@qq.com。
  • 基金资助:
    中国海油科技项目(SSKJ?YY?AHY?2024?14)

A Suspended Particle Detection Method for Petroleum Produced Water Reservoir Water Injection Based on Object Detection and Blur Assessment and Its Application

Chen LIN1(), Minghui LIU1(), Xinle WANG1, Ni WANG1, Dengtian MAO1,2   

  1. 1.Safety and Environmental Protection Branch,CNOOC Energy Development Co. ,Ltd. ,Tianjin 300450,China
    2.Anrui AI R&D Center,Tianjin 300450,China
  • Received:2026-03-01 Revised:2026-03-17 Published:2026-10-25 Online:2026-11-02
  • Contact: Minghui LIU

摘要:

随着环保生产技术的不断升级,悬浮颗粒物的精准监测已成为提升地质生态环境质量的关键环节。提出一种面向石油生产水微小颗粒的实时检测范式:在YOLOv8框架内,将GSConv与上下文锚点注意力(CAA)进行轻量化⁃注意力耦合,形成GSConv⁃CAA联合单元。首先,以改进YOLOv8定位颗粒并裁剪感兴趣区域;继而设计频域⁃梯度混合模糊度指标,自适应筛除低质量子图;最后,结合形态学算法快速拟合颗粒等效直径。为了确保模型在实现轻量化的同时保持特征表达能力,将CBS模块中的标准卷积替换为GSConv,并在检测头嵌入CAA,以显式建模像素级长程依赖,从而进一步强化中心区域特征。在自建油田水样数据集上的实验表明,所提方法的精确率达81.6%,AP50(交并比为0.5时计算得到的平均精度)达到82.2%、F1分数达到77.6%,推理速度为214 FPS,为石油生产水悬浮颗粒的实时、高精度监测提供了可行方案。

关键词: 颗粒物检测, 目标检测, 图像处理, 深度学习, YOLO

Abstract:

With the continuous upgrading of environmentally friendly production technologies, accurate monitoring of suspended particulates has become a pivotal step toward improving geological and ecological environmental quality. A real⁃time detection paradigm for micro⁃particles in petroleum produced water is proposed.Within the YOLOv8 framework, GSConv is lightweightly coupled with Contextual Anchor Attention (CAA) to form a GSConv⁃CAA joint module. First, an improved YOLOv8 localizes particles and crops the corresponding regions of interest. Next, a frequency⁃domain⁃gradient hybrid blur metric is designed to automatically discard low⁃quality sub⁃images. Finally, a morphological algorithm is employed to rapidly fit the equivalent diameters of the particles. To maintain representational capacity while reducing computational burden, we replace the standard convolutions in the CBS modules with GSConv and embed a CAA module into the detection head, explicitly modeling long⁃range pixel⁃wise dependencies and further enhancing central⁃region features. Experiments on a self⁃built oilfield water sample dataset show that the proposed method achieves a precision of 81.6%, AP50 (average precision calculated at an intersection over union of 0.5) of 82.2%, and an F1 score of 77.6%, with an inference speed of 214 FPS. This provides a feasible solution for real⁃time, high⁃precision monitoring of suspended particles in petroleum produced water.

Key words: Particulate matter detection, Object detection, Image processing, Deep learning, YOLO

中图分类号: 

引用本文

林晨, 刘铭辉, 王新乐, 王妮, 毛邓添. 基于目标检测与模糊度判断的油藏注水悬浮颗粒物在线检测方法及其应用[J]. 石油化工高等学校学报, 2026, 39(5): 65-73.

Chen LIN, Minghui LIU, Xinle WANG, Ni WANG, Dengtian MAO. A Suspended Particle Detection Method for Petroleum Produced Water Reservoir Water Injection Based on Object Detection and Blur Assessment and Its Application[J]. Journal of Petrochemical Universities, 2026, 39(5): 65-73.

使用本文