【作者】 吕琳；

【导师】 范继璋； 王明君；

【作者基本信息】 吉林大学 ， 地球探测与信息技术， 2011， 博士

【摘要】 目前天然气水合物已被公认为未来能源的替代品,存在巨大的经济效益和潜在的战略意义,很多国家政府和科研单位为此投入了巨大资金进行水合物研究并已取得了一定的成果。在天然气水合物研究中,天然气水合物的识别占有极其重要的地位,可以为天然气水合物储量估算、钻探、开采提供科学的参考依据。为此,探索一条天然气水合物识别的新方法无疑具有重要的理论意义和实践意义。本文通过借助于神经网络方法,对测井、地震数据进行处理,实现天然气水合物的识别。本研究得到了国家863高新技术研究与发展计划项目“天然气水合物勘探开发关键技术”课题(编号2006AA09202)的子课题“天然气水合物矿体的高精度地震定量评价软件开发”的资助,重点研究了天然气水合物神经网络识别与软件开发。论文以神经网络算法为天然气水合物识别和预测的主要算法,通过建立神经网络识别体系结构,搭建于分析、处理、解释和预测功能为一体的平台,设计神经网络识别体系模型,进而开发出具有实际应用价值的天然水合物神经网络识别系统。研究思路是在对测井数据与地震数据预处理、地震属性提取基础上,选用自组织神经网络进行分类,达到岩性识别和矿体边界识别的目的,选用BP神经网络方法,达到对水合物储量参数估算和预测的结果,从测井曲线、平面、地质体三种模式识别结果进行比对,为将来的勘探研究提供科学的数据资料。天然气水合物只有在特定环境才能存在。本文阐述了天然气水合物存在三个必要特征：充足的烃类物质、温压条件及水合物聚集和移动的稳定空间(地质构造环境)。目前水合物的研究识别主要方法是地震特征、地球化学特征及地貌标志。根据这些识别方法,本文对国内外已展开的水合物研究进行简单的叙述,对目前水合物研究的关键问题进行分析。神经网络能有效地解决地质数据关系复杂、非线性求解问题。本文阐述了神经网络的概念、特征、结构及基本原理,重点介绍了BP神经网络和自组织神经网络。两种神经网络有不同的网络结构、算法描述。BP神经网络是有导师学习方式,需要有网络输出与期望输出对比,而自组织神经网络属于无导师学习型,根据输入自动调整,主要完成聚类操作。在对它们的优点和不足进行反复研究、分析后,我们认为BP神经网络更适用于储量参数预测和估算,自组织神经网络更适用于岩性分类、矿体识别。测井数据包含丰富的地质纵向信息,能较好的反映地层、岩性和储量参数间的差异性。利用测井数据进行储量参数估计和预测,所采用的方法是不同的。储量参数估算的方法是通过取测井的数据实测数据建模,并对该井测井曲线上的储量参数进行估算。储量参数预测则是通过取模型井的实测数据为网络训练的样本数据进行建模,将其它井的测井曲线输入到此模型中,得到该井的储量参数预测结果实现的。岩性分类选用自组织神经网络方法,测井曲线上的样本数据输入到模型中,得到岩性分类结果。分类的个数,由使用者提前定好。地震数据进行预处理,提取地震属性数据。地震属性数据能更好的反映出地质信息,更利于水合物BSR及振幅空白带识别。地震属性数据进行主成分成析,选取部分属性,利用组织神经网络对地震属性数据进行分类,结果与测井结果反复比对,得出水合物矿体边界图及雕刻图。地震—测井联合反演能将测井数据频带宽与地震数据横向信息相结合。计算并选取合适的子波,求得的合成记录与地震数据反复比对,最终确定层位。利用BP神经网络,井旁地震数据为输入,测井输出,建立波阻抗模型。将整个地震体数据输入,得到波阻抗数据体。用同样的方法,将神狐海域地震属性数据作为输入,储量参数作为输出,预测储量参数结果。天然气水合物神经网络识别系统(SNET)采用QT4.3为开发平台,结合INT公司的图形处理插件开发而成。QT是一个跨平台的C++图形用户界面应用程序框架,使程序开发后可移植性强。它是完全面向对象,易扩展并允许真正的组件编程。INT公司的图形插件软件CarnacGeo,很好地解决了地震数据和测井数据的显示问题。本软件采用中国南海神狐海域的地震、测井资料进行测试,包括3个地震剖面数据、1个地震体3D体数和8个站位测井曲线数据。测试结果得到了测井分类图8张、测井预测图40张。地震属性图30张,地震分类图5张,雕刻图1张,地震平面储量参数图2张,波阻抗图1张。研究成果在“863”项目成果验收发挥了重要作用,表明其方法具有实际应用价值,可用于天然气水合物识别。天然气水合物地球物理勘探和神经网络算法仍然处于研究阶段,还有很多地方仍需改进。水合物识别系统整合了三种方法来进行识别：测井识别、地震识别、地震—测井联合识别。结果的显示方式可以以剖面形式显示,也可以以空间平面方式显示。更多还原

【Abstract】 Currently, natural gas hydrate has been recognized as an alternative energy in the future. There is a huge potential economic and strategic significance, so governments and research institutions in many countries invested huge funds for this hydrate research and have achieved certain results. In gas hydrate research, the identification occupies an extremely important position, and it can provide scientific references for estimating gas hydrate reserves, drilling and mining. To this end, exploring a new method for identification of gas hydrate is undoubtedly of great theoretical and practical significance.The thesis discusses the identification of the natural gas hydrate through the application of neural networks and processing of logging and seismic data. This research was funded by the sub-project "high-precision seismic quantitative evaluation software development of gas hydrate ore " under national 863 high-tech research and development project "key technologies of natural gas hydrate exploration and development" (No.2006AA09202), focusing on neural network necognition of gas hydrates and software development.In this thesis, a platform integrating analysis, processing, interpretation and forecast was built and a neural network system model was designed, through establishment of neural network identification architecture and with neural network algorithm as the main algorithm for the identification and prediction of gas hydrate. Based on above work, a practical natural gas hydrate neural network recognition system was developed. The research logic is selecting self-organizing neural network for classification based on the pre-processing of logging & seismic data and extraction of seismic attributes, to identify the lithologic and ore body boundary identification; selecting BP neural network to estimate and predict parameters of hydrate reserves; then comparing the results of three identifications from log presentation, surface, geological bodies as to provide scientific data for future exploration research.Gas hydrates can exist only in specific environment. This thesis presents three prerequisites for the existence of gas hydrate:sufficient hydrocarbons, temperature and pressure conditions and the stability of hydrate accumulation and moving space (geological environment). The main methods for hydrate identification research currently focus on seismic features, geochemistry and landscape signs. Based on these identification methods, the thesis provides a brief description on undergoing natural gas hydrate researches both at home and abroad, and analyzes the primary challenges on hydrate research.Neural network can effectively deal with the complex relationship and non-linear equation of geological data. The thesis discusses the concept, characteristics, structure and basic principles of neural networks with focus on the BP and self-organizing neural networks which have different network structures and algorithm descriptions. BP neural network is instructed learning, which includes the comparison of network output with expected output, while self-organizing neural network is unsupervised learning, which mainly completes clustering operation according to automatic input adjustment. After repeated study on their advantages and disadvantages, analysis, we believe that BP neural network is more suitable for forecasting and estimating reserves parameters, while self-organizing neural network is more suitable for rock classification, ore body identification.Logging data contains a wealth of longitudinal geological information, which can better reflect the discrepancy among stratum, lithology and reserve parameters. Different methods are applied when logging data is used to estimate and predict reserve parameters. To estimate reserve parameters needs build a model based on the measured logging data, then make estimation of reserve parameters located on log presentation; while to predict reserve parameters needs build a model based on network training sample data which is obtained from the measurement of model well, then input the log presentation of other wells into the newly-established model and make prediction of reserve parameters. Lithological classification uses self-organizing neural network method:input the sample data on the log presentation into the model to get classification results. The number of classification shall be preset by the user.Seismic data is pre-processed to extract seismic attribute data which can better reflect the geological information as to facilitate the identification of hydrate BSR and amplitude blanking belts. Seismic attribute data is analyzed for its principal ingredients and partial attributes are selected, then the self-organizing neural network is used to classify the selected data. Finally gas hydrate ore body boundary maps and carving figures are obtained by repeated comparison of classification results with logging results.Seismic—logging joint inversion can combine the frequency bandwidth of log data and horizontal information of seismic data. Then appropriate wavelet is calculated and selected, and synthetic seismogram is compared with the seismic data repeatedly to finally determine the layer. BP neural network is used to set up the wave impedance model with well seismic data as input and logging as output, and the entire seismic volume data is input into the model to get wave impedance data volume. The same method is used to predict reserve parameter results in Shenhu sea area by taking seismic attribute data as input and reserve parameter as output.Neural network recognition system of natural gas hydrate (SNET) takes QT 4.3 as its development platform and integrates INT company’s graphics processing plug-in module. QT is a cross-platform C++graphical user interface application framework that enables easy portability after program development. It is fully object-oriented, easily extensible and allows true component programming. INT’s graphics plug-in software CarnacGeo is a good solution to the display problems for seismic data and well log data.The software is tested with seismic and logging data from Shenhu area of China South Sea, which contains three SEGY files, one seismic 3D body data and eight stations logging data. The results are eight logging classification maps, forty logging forecast figure, thirty seismic attributes figures, five seismic classification figure, one engraving figure, two seismic surface storage parameter figures and one wave impedance figure. The research results play an important role in the acceptance inspection of "863" project, which indicates that the method is a practical way to identify gas hydrate. Geophysical exploration and neural network algorithm for gas hydrate are under development, and many areas still need improvement. Hydrate recognition system incorporates three methods:well logging, seismic identification and seismic—logging joint identification. The results can be displayed both in profile and space plane.更多还原