基于前馈神经网络井控多属性融合的断裂识别方法  

A method for identifying faults based on well-controlled multi-attribute fusion using a feedforward neural network

作　　者：赵军[1] 冉琦[1] 朱博华[1] 李洋 梁舒瑗 常健强 ZHAO Jun;RAN Qi;ZHU Bo-Hua;LI Yang;LIANG Shu-Yuan;CHANG Jian-Qiang(Geophysical Research Institute,SINOPEC,Nanjing 211103,China)

机构地区：[1]中国石化石油物探技术研究院,江苏南京211103

出　　处：《物探与化探》2024年第4期1045-1053,共9页Geophysical and Geochemical Exploration

基　　金：中国石化科技攻关项目“顺北深层断溶体油藏描述及可采储量定量表征”(P21064-1)。

摘　　要：塔里木盆地碳酸盐岩断控缝洞型油气藏埋藏深度大、构造复杂,且断裂高度发育,断裂是研究区域内成藏主控因素及可能的油气运移通道,对其空间展布位置及发育强弱的预测至关重要。断裂检测属性众多,不同断裂检测属性由于计算方法不同表征的断裂尺度及特征存在一定的差异性,且常规属性检测忽视了测井信息的利用与约束,为了获取更加全面、准确的断裂预测结果,本文优选多类断裂检测属性,并结合测井数据作为先验信息,利用前馈神经网络算法进行属性融合。首先优选AFE、likelihood、倾角等多类具有不同特征的断裂属性,结合测井放空漏失数据、成像测井信息及地震同相轴错段情况作为断裂发育类型判别条件建立了断裂特征识别样本库;在样本库基础上进行深度前馈神经网络训练,对比测试了不同隐含层深度条件下的学习效果,获取预测误差最小的神经网络预测模型;最后将神经网络预测模型应用于全工区断裂预测。经对比分析,认为深度学习融合属性预测断裂与测井解释结果更为吻合,且能综合不同尺度特征的断裂信息,有效提升了预测准确度和可靠性。The fault-controlled fractured-vuggy carbonate reservoirs in the Tarim Basin exhibitconsiderable burial depths,complex structures,and highly developed faults.Faults serve asa dominant factor controlling oil and gas accumulation and possible hydrocarbon migration pathways in the study area.Hence,it is critical to predict their spatial distributions and sizes.There existvariousfault detection attributes,which characterize fault scales and features differently due totheir different calculation methods.Moreover,conventional attribute detection ignores the use and constraints of logs.For more complete and accurate fault prediction results,this study selected multiple fault detection attributes for fusion using the feedforward neural network algorithm,with logs as prior information.First of all,a sample database for fault feature identification was established using fault attributes(like AFE,likelihood,and dip angle)with distinct characteristics anddiscrimination criteria of fault types,including lost circulation data,imaging logs,and seismic event dislocations.The deep feedforward neural network was trained based on the sample database.A neural network prediction model with a minimum prediction error was obtained by comparing and testing the learning effects under different hidden layer depths.Finally,the neural network prediction model was applied to the fault prediction of the study area.The comparative analysis reveals thatthe fault prediction using deeplearning-based fused attributesyielded prediction results more consistent with the log-based interpretation results,and could synthesize the information of faults with different scale characteristics,thus effectively improving the prediction accuracy and reliability.

关 键 词：断裂检测 井控 属性融合 前馈神经网络 缝洞型油气藏 

分 类 号：P631[天文地球—地质矿产勘探]