【作者】 王强；

【导师】 张凯；

【作者基本信息】 中国石油大学(华东) ， 油气田开发工程， 2023， 硕士

【摘要】 井场是油气田建设的重要生产场所,而抽油机井是井场的主要生产设施。掌握抽油机井的工作状况对于油田产量提升意义重大,是管理人员制定措施和指挥生产的依据之一,近年来受到广泛关注。随着人工智能的研究热潮和发展大势,新型机器学习算法正逐渐取代传统方法,成为抽油机井工况诊断的主流技术。通过对示功图样本进行数据收集和分析、特征总结和提取,挖掘有效信息用于判断抽油机井的不同工作状况,再将得到的结果反馈给管理者来辅助油田运营和生产。本文以视频监控为主要手段,建立低成本下的抽油机工况智能分析模型,实现人工智能在油气田开发领域的交叉应用。论文将抽油机井诊断技术和机器学习方法相结合,主要进行以下几项工作:(1)首先分析井场的抽油机视频监控数据,在此基础上提取抽油机图像样本。采用人工设计的关键点标记抽油机图像的结构,并通过自动数据增强方法扩充样本数量。从中筛选出满足条件的样本数据用于训练抽油机悬点运动模型。(2)通过关键点检测技术计算抽油机悬点运动过程,并结合高分辨率网络构建模型,能够得到连续的抽油机悬点位移数据。将数据分析结果结合悬点载荷绘制出更为准确的示功图样本。(3)针对示功图样本不均衡问题,通过将迁移知识和元学习理论相结合,构建元迁移学习模型从小样本的角度解决工况诊断问题。本方法迁移预训练知识,优化网络模型,减少网络参数量,加速梯度迭代过程,从而只凭借少量示功图样本便能实现较为理想的工况诊断效果。通过多组实验分析,本文系统地比较了主流卷积神经网络、传统小样本方法以及本研究方法的模型结果。在抽油机悬点运动分析上准确率达到93%,高于对比实验6%以上;在小样本工况诊断问题上准确率接近80%,高于对比实验9%以上,领先业内目前水平。关键点检测技术和高分辨率网络的结合,以及元迁移学习方法的使用,使得工况诊断模型预测更为准确,而且实际应用费用更少,展现了本方法在油田开发的未来潜力。更多还原

【Abstract】 As a production site,the well pad is of great significance for the construction of oil and gas fields.Pumping well is the main equipment inside the well pad,which is used to extract oil.As one of the bases for managers to formulate measures and command production,the working conditions of pumping wells can significantly affect oilfield production.Using scientific methods to reasonably grasp operating conditions has received widespread attention in recent years.With the research upsurge and development trend of artificial intelligence,new machine learning algorithms are gradually defeating traditional manual analysis methods and becoming the mainstream technology for diagnosis of pumping well conditions.Through data collection and analysis,feature summary and extraction of indicator diagram samples,effective information is mined to judge the different working conditions of pumping wells.The obtained results are then passed on to managers to assist in the operation and production of the oilfield.This thesis intends to combine machine learning method and pumping unit condition diagnosis technology to form a new data monitoring and intelligent analysis method for pumping unit wells,and realize the cross application of artificial intelligence in the field of oil and gas field development.The thesis combines diagnostic technology for pumping wells with machine learning methods,and mainly carries out the following tasks:(1)Firstly,analyze the video monitoring data of the pumping units at the well site,and extract image samples of the pumping units on this basis.The structure of the pumping unit image is marked with manually designed key points,and the sample size is expanded through automatic data augmentation methods.Select sample data that meet the conditions from them for training the suspension point motion model of the pumping unit.(2)By using key point detection technology to calculate the movement process of the pumping unit’s suspension point,and combining it with a high-resolution network to construct a model,continuous displacement data of the pumping unit’s suspension point can be obtained.Draw more accurate indicator diagram samples by combining the data analysis results with the suspension load.(3)In view of the imbalance of indicator diagram samples,a meta transfer learning model is constructed by combining transfer knowledge and meta learning theory to solve the problem of condition diagnosis from the perspective of small samples.This method transfers pre trained knowledge,optimizes the network model,reduces the number of network parameters,accelerates the gradient iteration process,and achieves ideal condition diagnosis results with only a small number of indicator diagram samples.Through multiple sets of experimental analysis,the model results of mainstream convolution neural network,traditional small sample method and this research method are systematically compared.It is proved that the designed working condition diagnosis model is very competitive,and the accuracy rate of the trajectory simulation of the pumping unit reaches93%,which is more than 6%higher than the comparison experiments;The accuracy rate of small sample working condition diagnosis is close to 80%,which is more than 9%higher than the comparison experiments,leading the current level in the industry.The combination of key point detection technology and high-resolution network,the use of meta-migration learning method,and the introduction of automatic data augmentation technology make the prediction of the working condition diagnosis model in this thesis more accurate,and the actual application cost is less,which shows the future potential of this method in oilfield development.更多还原