基于MultiRes-Unet神经网络的三维断层识别研究

Three-D fault identification based on MultiRes-Unet neural network

(１．中国地质大学(武汉)地球物理与空间信息学院,湖北武汉４３００７４;２．东方地球物理勘探有限责任公司研究
院,河北涿州０７２７５０;３．黄河勘测规划设计研究院有限公司,河南郑州４５０００３)

(１．ChinaUniversityofGeosciences(Wuhan),Wuhan４３００７４,China;２．GeophysicalResearchInstitute,BGPInc．,CNPC,
Zhuozhou０７２７５０,China;３．YellowRiverEngineeringConsultingCo．,Ltd．,Zhengzhou４５０００３,China)

摘要:三维地震数据中的断层规模从米级到千米级不等,断距从米级到数十米级不等.断层通常表现为同相轴错断、突然增减、消
失或扭曲等显著地震响应特征.断层在三维地震数据体中占的体积比例很小,使得利用常规断层识别方法得到的断层识别结果存
在不连续、识别率低等问题.针对断层的多分辨率特征,充分考虑了断层点或线在整个地震数据的占比小等特点,提出了一种三维
神经网络MultiRes-Unet３D断层识别方法,该方法在网络学习过程中使用加权交叉熵损失函数解决了普通交叉熵损失函数不同项
之间的平衡问题,使得神经网络具有了较为可靠的断层识别能力.首先,利用正演模拟方法生成三维合成地震数据集和断层标签,
然后基于Tensorflow搭建、训练与测试MultiRes-Unet３D神经网络,再将训练好的网络模型迁移到实际三维地震数据的断层识别
中.该神经网络断层识别方法在实际地震数据中的应用表明,断层识别结果空间连续性好,识别结果客观,断层边界更为准确,网
络模型泛化性能良好,适用于具有不同断层构造特征的实际地震数据,节约了断层解释的时间成本与人工成本.

Abstract:Faults,whichextendfordozensofmeterstodozensofkilometerswithfaultthrowchangingfromafewmeterstotensof
meters,mayexhibitquitedifferentseismicresponses,e．g．discontinuousreflections,suddenlyincreasingordecreasingevents,and
blankordistortedreflections．Faultresponsesmerelyaccountforatinypercentageoftotalseismicresponses;thismeansthatfault
predictionsmaybequitesnatchyandsomewhatinaccurate．A３Dneuralnetwork,MultiRes-Unet３D,isaplausiblesolutiontomulti-
resolutionfaultcharacterization．Inviewofthesmallproportionoffaultresponses,aweightedcross-entropylossfunctionisusedin
thelearningprocesstobalanceamongdifferenttermsandimprovethecredibilityoffaultdetection．３Dsyntheticseismicdatasets
andfaultlabelsaregeneratedthroughforwardmodeling．TheMultiRes-Unet３Disbuilt,trained,andvalidatedbasedonTensor-
flow,andthenthenetworkmodeltrainedisappliedto３Dseismicdataforfaultidentification．Theresultsshowgoodspatialconti-
nuityoffaultidentificationandcrediblefaultboundarydetection．TheMultiRes-Unet３Dhasgoodgeneralizationperformanceand
couldbeappliedtoseismicdatawithdifferentfaultfeatures．Thistechniquecansavethecostintimeandlaboroffaultinterpreta-
tionandyieldobjectiveresults．