基于LLE和SVM的地震断层自动识别方法  

Automatic identification method of seismic fault based on LLE and SVM

作　　者：邹冠贵[1,2] 丁建宇 任珂 殷裁云 董青山 ZOU Guangui;DING Jianyu;REN Ke;YIN Caiyun;DONG Qingshan(College of Geoscience and Surveying Engineering,China University of Mining and Technology-Beijing,Beijing 100083,China;State Key Laboratory of Coal Resource and Safety Mining,China University of Mining and Technology-Beijing,Beijing 100083,China;Huaneng Coal Technology Research Co.,Ltd.,Beijing 101100,China)

机构地区：[1]中国矿业大学(北京)地球科学与测绘工程学院,北京100083 [2]中国矿业大学(北京)煤炭资源与安全开采国家重点实验室,北京100083 [3]华能煤炭技术研究有限公司,北京101100

出　　处：《煤炭学报》2023年第4期1634-1644,共11页Journal of China Coal Society

基　　金：国家重点研发计划资助项目(2018YFC0807803)。

摘　　要：传统地震资料的断层解释主要依靠解释者的知识和经验,存在工作量大、效率低的问题。基于机器学习的断层识别方法,可以融合已有的地质资料、解释人员的知识和经验,构建高质量的数据集,增加解释的准确率。为了提高机器学习方法断层解释的准确率,构建基于局部线性嵌入(LLE)和支持向量机(SVM)算法的断层识别方法。首先,介绍了LLE和SVM算法的基本原理,说明各算法的计算过程和主要参数;然后建立断层正演模型,分析不同属性的断层响应特征,针对训练数据集中多种地震属性之间的信息冗余,分别通过LLE和主成分分析(PCA)2种算法对地震属性数据进行降维,引入的量化指标计算结果表明LLE算法对于非线性数据体有较好的降维效果;利用西上庄井田6条巷道、5口钻井揭露的11854个已知构造信息的数据点,分别训练SVM,PCA-SVM和LLE-SVM断层识别模型;以准确率A、查全率R、查准率P、F作为模型的衡量标准,对比各模型在工区数据上的预测分类性能;其中,LLE-SVM模型综合表现最佳,查准率可达94.4%,远高于其他模型;最后,利用构建的各模型对整个工区进行预测,并结合实际揭露情况和人机交互解释结果进行分析。综合结果表明,基于LLE和SVM的断层识别方法在去除冗余信息的同时能够有效突出断层响应特征,减少主观人为因素的影响,提高断层解释的效率。The fault interpretation of traditional seismic data mainly relies on the knowledge and experience of the interpreter,which has the problems of heavy workload and low efficiency.In order to construct high-quality data sets and increase the accuracy of interpretation,machine learning can integrate the existing geological data,the knowledge and experience of the interpreter.A fault recognition method based on Local Linear Embedding(LLE)and Support Vector Machine(SVM)algorithms is constructed to improve the accuracy of fault interpretation by machine learning methods.First,the basic principles of LLE and SVM algorithms are introduced to illustrate the calculation process and main parameters of algorithms.Then a fault forward modeling model is established to analyze the fault response characteristics of different attributes.Aiming at the information redundancy among various seismic attributes in the training data set,the seismic attrib-ute data are dimensionally reduced by LLE and principal component analysis(PCA).The intersection diagram shows that the LLE algorithm has a better dimensionality reduction effect for nonlinear data volumes.The SVM,PCA-SVM and LLE-SVM recognition models of fault were trained by using 11854 known structural information data points revealed by six roadways and five drilled wells in the Xishangzhuang Coalfield.Accuracy rate A,recall rate R,precision rate P and F value were used as the measurement standards to compare the prediction and classification performance of each model in the research area.Among them,the LLE-SVM model has the best overall performance,with a precision rate of 94.4%,much higher than those of other models.Finally,the whole research area is predicted by using the models,and analyzed by combining the actual disclosure and artificial interpretation results.The comprehensive results show that the fault identific-ation method based on LLE and SVM can effectively highlight the fault response characteristics while removing redund-ant information,reduce the influence of subjec

关 键 词：断层识别 地震属性优化 煤田三维地震 局部线性嵌入 支持向量机 

分 类 号：P631.4[天文地球—地质矿产勘探]