基于神经网络与图形处理技术的沉积微相识别系统
详细信息    本馆镜像全文|  推荐本文 |  |   获取CNKI官网全文
摘要
天津大港港东油田的“精细油藏描述”技术提供了全新的地质模型,规范了地质研究工作的程序和方法。从根本上解决了这个老油田“措施难有成效、递减难以控制、新井难定井位、滚动难有突破”的尴尬境地。沉积微相模式识别是研究地层结构、构造和分析描述储层的基础工作,也是“精细油藏描述”技术的核心组成部分之一。但是,到目前为止,大港油田的绝大部分沉积微相识别工作都是手工进行的,对“精细油藏描述”技术的描述成果的质量造成了很大影响。
    测井沉积学是近些年来发展起来的一门新的边缘学科,它是以测井资料为主,在油区沉积学研究覆盖下,并且与其它学科和技术紧密结合的一种专门的多井测井评价技术。其核心是把测井资料用于油区沉积学研究,以获取进一步描述油气储集层的基础信息。随着测井沉积学和计算机科学的发展,利用计算机处理测井资料来自动识别沉积微相已被人们所关注。经过对多种沉积微相识别方法的分析研究,本文采用一种基于神经网络技术与图像处理技术相结合的沉积微相自动识别方法。经最终用户的使用,证明该系统具有较高的识别效率,并且在辅助地质研究人员进行单井精细解释和多井分析方面有一定的突破。
    该系统主要在三个方面具有优势:
    使用将数字化测井曲线和地层参数预处理转化为二值点阵图像模式的方式处理测井数据,避免提取曲线图像特征的困难。同时也提高了原始测井数据信息的利用和识别的准确性。
    实现通用ANN模型,使用可定制的神经网络,这使得系统具有潜在的通用性,经过简单的升级,即可推广到港东油田以外的其他油田使用。
    自动生成沉积微相图,提供较强的图像编辑功能,可综合利用已有地质图像数据,对单井精细解释和多井分析提供较好的参考信息。
    识别模型
    对于模式识别系统,提取代识别模式的特征是决定系统成败的核心问题。完整、客观地提取测井曲线形态模式特征是利用测井资料识别沉积微相的关键。
    
    本文在测井曲线特征提取上,将数字化测井曲线和地层参数预处理转化为二值点阵图像模式,经过点阵数据编码压缩,直接输入网络。该方法突破了以往利用神经网络识别沉积微相必须先提取曲线特征的做法,简化了传统预处理的过程,提高了原始资料信息的利用和识别的准确性。
    沉积微相模式识别的对象是一个区域和深度相结合的立体区域,其结构如图1所示。
    
    图1. 识别对象立体结构图
    测井数据的采集是将油井从上至下,每个0.125米取一个采样点,得到包括自然电位、2.5米电阻率、微电极、微电位和声波时差在内的多种测井数据。
    为了根据测井数据生成网络输入,以及进行数字测井曲线的绘制,需对单砂体范围内的测井数据进行转换,算法如下:
    输入:井号W、油层组OG、小层编号SL、单砂体编号S
    对于由W、OG、SL、S确定的单砂体,从数据库中取出砂层顶深STOP、砂层底深SBOTTOM、砂层厚度SHEIGHT。
    根据如下公式计算取数顶深CTOP和取数底深CBOTTOM:
    
    
    
    根据W、CTOP和CBOTTOM从数据库中获取测井数据,即取出W的从CTOP到CBOTTOM的五种类型的全部测井数据:自然电位SP、2.5米电阻率RA25、微电极MNOR、微电位MINV和声波时差AC。
    对于每种测井数据分别进行转换。
    当该范围内的数据采样点数刚好为20时,则转至第八步;否则,继续下一步。
    根据如下公式计算插值砂层顶深ITOP和插值砂层底深IBOTTOM:
    
    
    将从ITOP到IBOTTOM的深度均分19份,共得到20个深度值。将每一个深度值H与实际获取的数据的深度值比较,如果相等,这个深度的参数值就取实际获取的数据。如果在两个实际获取的数据的深度HT和HB之间,则根据线性插值计算H深度的参数值V:
    
    其中VB、VT分别为HB、HT点的实际测井数据值。
    对获得的20个数据值进行归一化处理:
    
    其中为归一化后的结果,和分别为归一化之前20个数据中的最大值和最小值。
    返回第四步,直到五种测井数据全部处理完毕。
    
    输出:二维数组,对于5类测井数据,每类存有20个数据值。
    神经网络
    为了保证系统潜在的通用性,本文设计并实现了一个通用ANN模型,此ANN模型实现的人工神经网络属于无反馈的前向ANN。所采用的学习算法为一括修正法(是基于传统误差逆传播算法的一种改良算法),即采用有教师指导的学习方式。神经元的转移函数采用sigmoid函数,误差函数采用二乘误差函数。
    本模型提供接口给开发人员,使其可以设置ANN的各个参数,包括隐层层数、隐层结点个数、计算公式参数等。可以实现一个实际的人工神经网络:设置相应参数,启动训练过程以及利用网络进行模式识别。
    该模型是作为COM组件进行设计的,包括3个类和一个枚举类型,如图2所示。
    
    
    图2. 人工神经网络模型数据结构
    
    
    
    
    由于港东油田的明化镇和馆陶镇的沉积构造有着显著不同,二者的沉积微相代识别模式是截然不同的。另外,由于早期测井技术的限制,在明化镇和馆陶镇,都有一部分油井只有两种测井数据。因此,系统利用上一节中介绍的通用ANN模型,定制并训练了4个人工神经网络,用于根据不同数量的测井数据进行不同的代识别模式的识别功能。具体情况如表1所示。
    网络
    名称
    网络
    代号
    ?
Detailed Description on Oil Resources is a new technology to solve the problem, which all of old oil fields is encountering, in GangDong Oil Field of DaGang of TianJin. It has provided new geologic models and standardized the procedure and method on geologic research. Sedimentary Micro-Facies Model Identification is the basic work which studies the structure of stratum and describes reposition. It is the main makeup of Detailed Description on Oil Resources. But, The model identification is done by hand in DaGang Oil Field at present, which plays the effect of Detailed Description on Oil Resources down.
    Measure-Well Sedimentology is a new subject recently, which is a special evaluation technology on multi-well, based on metrical data of oil wells, covered by oil-section sedimentology and combined with other subjects and technology. It uses metrical data of oil wells to research oil-section sedimentology and gather basic information on deposited layer of oil and gas. By the development of Measure-Well Sedimentology and computer science, computer is used to identify Sedimentary Micro-Facies Model with metrical data of oil wells. This thesis will realize a method of Sedimentary Micro-Facies Model Identification based on artificial neural networks (ANN) and image handling.
    The method’s goodness is following:
    Metrical data of oil wells is abstracted to series of real number, by which the difficulty of characteristic drawing is avoided and value in use of metrical data of oil wells is improved.
    A common ANN model is realized. The ANNs can be tailored. It makes the system propagable.
    The system can build the image of Sedimentary Micro-Facies automatically and provide good tools to edit the images. It can use the geologic image data and create abundant reference for the detailed explain on one well and analyse on multi wells.
    
    Identification Matrix
    The drawing of identification models is the key to any system whose work majors on model identification. So, It is the main problem to draw the characteristic of metrical data of oil wells, for using it to identify Sedimentary Micro-Facies Model.
    Identification Object is a diorama with area and depth, as Graph 1 displays.
    
    Graph 1. Diorama of Identification Objects
    Metrical data of oil wells is collected from top to bottom each 0.125 meter. For generation of net inputs, the metrical data need to be transferred.
    Input: Well Code-W, Oil Group–OG, Small Layer–SL, Sand-S
    To the Sand named by W,OG,SL and S, get the top height –STOP, bottom height-SBOTTOM, depth-SHEIGHT.
    Calculate CTOP and CBOTTOM:
    
    
    As W, CTOP and CBOTTOM, get metrical data from db: SP,
    
    
    RA25,MNOR,MINV and AC.
    Handling every metrical data.
    When number of metrical data is 20, go to Ⅷ; or go on.
    Calculate ITOP and IBOTTOM:
    
    
    Cut to 19 part from ITOP to IBOTTOM and get 20 depth value. Compare each depth value-H to the depth value with real data, if they are equaled, use the real data; or calculate as the following formula.
    
    Handle to [0,1]:
    
    Go to Ⅳ until all of five kinds of data are handled.
    Output: 2- dimension array , each kind of data have 20 values.
    Artificial Neural Networks
    A common ANN model is realized to make the system propagable. The real ANN based on it is an no-feedback ANN. The training arithmetic is an improved BP arithmetic, method of revision together which is a study way with teacher’s instruction. The transfer function is sigmoid function. The model provides interfaces to developers to set the parameters of real ANN, including the number of middle layers and their nodes, the parameters of transfer function and so on.
    The model is a COM, which includes three classes and one enumerate type, as Graph 2 displays.
    
    
    Graph 2. Data Structure of ANN Model
    
    
    
    
    Sedimentary structures are so different between MingHuaZhen and GuanTaoZhen of GangDong Oil Field that the two parts’ Sedimentary Micro-Facies Model are different too. And, some oil wells have only two kinds of metrical data because of the former measure
引文
何幼斌,《Sedimentary Petrology》,江汉石油学院,2003.
    Reading H G,《Sedimentary Environments and Facies》,Blackwell Scientific Publication,Oxford,1986.
    尹寿鹏,王贵文,测井沉积学研究综述,中图分类号:P512.2;P631.81,文章标识码:A,文章编号:1001-8166(1999)05-0440-06.
    冯增昭,《沉积岩石学》,石油工业出版社,1993.
    冯增昭,《中国沉积学》,石油工业出版社,1994.
    朱明程,徐健,李昆华,采用MCU的多层神经网络模型设计的探讨,东南大学学报,2000年第2期.
    李鸿雁等,人工神经网络及峰值识别理论及其在洪水预报中的应用[J],水利学报,2002,,(6):15~20.
    Ximin Yuan,Hongyan Li,Raising flood forecasting precision based on improving artificial neural networks algorithm, XXIX IAHR congress proceedings, Theme c, Tsinghua university press, September 2001.11.5:175-183.
    张乃尧,阎平凡,神经网络与模糊控制,清华大学出版社,1998.
    周春光,梁艳春,计算智能,吉林大学出版社,2001.
    焦翠华,神经网络和分形几何方法在识别测井沉积微相中的应用〔J〕,沉积学报,1997,15(3):62~66.
    侯庆功,神经网络测井解释系统〔A〕,《中国海上油气地质》编辑部编写的《中国石油学会测井委员会第9届测井年会论文集》〔C〕,1996.85~90.
    刘国安,武本凡,港东油田为何生机再现,中国石油报,2003-04-11.
    寥明光,模糊模式识别沉积相〔J〕,测井技术,1991,15(3):157~162.
    Rogers S J,利用神经网络由测井数据确定岩性〔J〕(钱志良译),国外油气勘探,1993,5(1):76~83.

© 2004-2018 中国地质图书馆版权所有 京ICP备05064691号 京公网安备11010802017129号

地址:北京市海淀区学院路29号 邮编:100083

电话:办公室:(+86 10)66554848;文献借阅、咨询服务、科技查新:66554700