新疆石油天然气 ›› 2025, Vol. 21 ›› Issue (2): 24-34.DOI: 10.12388/j.issn.1673-2677.2025.02.003

• 油气勘探 • 上一篇    下一篇

基于有监督与无监督算法的卡钻智能诊断分析

  

  1. 1.中国石油大学(北京),北京昌平 102249;
    2.油气资源与工程全国重点实验室,中国石油大学(北京),北京昌平 102249;
    3.智能钻完井技术与装备研究中心,中国石油大学(北京),北京昌平 102249
  • 出版日期:2025-06-30 发布日期:2025-06-30
  • 通讯作者: 祝兆鹏(1993-),2021年毕业于中国石油大学(北京)油气井工程专业,博士,副教授,目前从事井筒多相流理论、智能钻井基础理论与工具装备的教学与科研工作。(Tel)010-89732176(E-mail)zhuzp@cup.edu.cn
  • 作者简介:宋先知(1982-),2010年毕业于中国石油大学(北京)油气井工程专业,博士,教授,目前从事油气井流体力学与工程、智能钻完井理论与技术的教学与科研工作。(Tel)010-89735895(E-mail)songxz@cup.edu.cn
  • 基金资助:

    中国石油创新基金“油气钻井破岩智能监测与优化调控技术”(2022DQ02-0308)

Intelligent Diagnosis and Analysis of Stuck Pipe Based on Supervised and Unsupervised Algorithms

  1. 1.China University of Petroleum (Beijing),Changping 102249,Beijing,China;
    2. State Key Laboratory of Petroleum Resources and Engineering,China University of Petroleum (Beijing),Changping 102249,Beijing,China;
    3.Research Center of Intelligent Drilling and Completion Technology and Equipment,China University of Petroleum (Beijing),Changping 102249,Beijing,China
  • Online:2025-06-30 Published:2025-06-30

摘要:

在钻井工程中,卡钻作为常见的井下复杂工况之一严重影响钻井效率,卡钻监测对保障钻井作业的安全与效率至关重要。近年来,人工智能技术的快速发展为卡钻监测提供了新的思路,而现有的卡钻智能监测研究多聚焦单一无监督或有监督算法的优化与应用,针对两类算法在卡钻监测场景中的系统性对比研究仍存空白。通过距离相关系数筛选钻头位置、大钩高度、转盘扭矩等多维度钻井参数为研究对象,构建包含经典无监督算法模型(自编码器、K均值聚类、DBSCAN聚类)和有监督算法模型(支持向量机、随机森林、长短期记忆网络)的对比评估体系,对两类算法的卡钻趋势判断能力开展性能分析。结果表明,与有监督算法模型相比,无监督算法模型的卡钻监测平均准确率提高12.5%、平均虚警率降低37.1%、平均漏警率降低27.6%,无监督算法模型在小样本及无机理约束情况下更有优势。研究结果可为钻井工程卡钻智能监测的模型选型与优化提供参考,推动无监督算法模型在卡钻风险监测中的实际应用。

关键词:

卡钻监测, 机器学习, 无监督算法模型, 有监督算法模型, 钻井

Abstract:

In drilling engineering,stuck pipe,as one of the common downhole complex issues,seriously affects drilling efficiency. Stuck pipe monitoring is crucial for ensuring the safety and efficiency of drilling operations. With the rapid development of artificial intelligence technology in recent years,new approaches have emerged for stuck pipe monitoring. However,the existing research on intelligent stuck pipe monitoring mainly focuses on the optimization and application of single unsupervised or supervised algorithm,leaving a gap of systematic comparative studies on these two types of algorithms with respect to stuck pipe monitoring. This study selects multi-dimensional drilling parameters such as bit position,hook height,and torque of ratary table as the study targets,according to distance correlation coefficients,constructs a comparative evaluation system involving classic unsupervised algorithms (AE,K-means,DBSCAN) and supervised algorithms (SVM,RF,LSTM),and analyzes the performance of these algorithms in estimating stuck pipe trends. The results show that compared with supervised algorithm,unsupervised algorithm delivers a increase of 12.5% in monitoring average accuracy and reductions of 37.1% and 27.6% in average false and missed alarm rates respectively. Unsupervised algorithm demonstrates greater advantages in cases of small sample sizes and absence of mechanistic constraints. The findings of this research provide references for model selection and optimization in intelligent stuck pipe monitoring for drilling engineering and promote the practical application of unsupervised algorithm in stuck pipe risk monitoring.

Key words:

stuck pipe monitoring, machine learning, unsupervised algorithm model, supervised algorithm model, drilling

中图分类号: