基于SVM测井数据的火山岩岩性识别——以辽河盆地东部坳陷为例

Lithological identification of volcanic rocks from SVM well logging data:Case study in the eastern depression of Liaohe Basin

辽河盆地东部坳陷储集层由火山多期喷发形成,岩相岩性复杂,岩性以中、基性火山岩为主.本文将火山岩的岩心及岩矿鉴定资料与测井数据进行整合,应用测井数据建立支持向量机(SVM)两分类和多分类岩性识别模式.首先,深入研究支持向量机二分类及"一对一"、"一对多"和有向无环图三种经典多分类算法的基本原理及结构;然后,总结研究区域火山岩岩石特征,分析测井数据的测井响应组合特征,选择40口井中岩心分析和薄片鉴定资料完整、常规五种测井曲线(RLLD,CNL,DEN,AC,GR)齐全的1200个测井数据作为训练样本,构造三种支持向量机岩性识别模式;最后,对4测试井中800个测井数据进行岩性识别,识别结果与取心段岩心描述和岩心/岩屑薄片鉴定资料对比,实验结果表明有向无环图更适合辽河盆地火山岩的识别,识别正确率达到82.3%.

The eastern depression in Liaohe Basin has the geologic characteristic of multiphase volcanic eruption and complicate lithofacies/lithology.The reservoir is made up of intermediate and basaltic volcanic rock.The lithology of volcanic rock is the basis of precise reservoir evaluation which is also the major task of logging evaluation.In complex reservoirs,it is a challenge to classify the lithology of volcanic rock by using the existing methods based on well logging data.On the basis of core,rock-mineral determination material and well logging data,we apply binary support vector machine（SVM）and multiclass support vector machine to identify volcanic rock lithology.The classical SVM is a binary classifier,whereas we often have to solve problems involving multiclass classification.Firstly,the principle of binary and three multiclass SVM of ‘oneagainst-rest＇,‘one-against-one＇and ‘directed acyclic graph＇were analyzed.Secondly,on the basis of compensated neutron logging（CNL）,density logging（DEN）,acoustic logging（AC）,deep lateral resistivity log（RLLD）and gamma ray logging（GR）from 40 wells,with a total of1200 logging data in Liaohe Basin,China,we construct the binary support vector machine model to classify volcanic rock and non-volcanic rock.The method of cross validation and grid searching algorithm were adopted to optimize the penalty factor and kernel parameter of SVM.Then,we expend binary model to multiclass model by ‘one-against-rest＇,‘one-against-one＇and ‘directed acyclic graph＇（DAG）,and construct multiclass models to classify 6types of volcanic rocks,which consist of basalt,non-compacted basalt,trachyte,non-compacted trachyte,gabbro and diabase.According to the geological core data,we compare the three multiclass SVM models,and calculate their accuracy rate,taking 4 wells with a total of 800 logging data for example.We compare the identification result with core analysis and cutting description,the calculating result indicated the accuracy rate of DAG method reaches up to 82.3%,and the ‘one-against-rest＇method is 80.2% and the ‘one-against-one＇method is 80.3%,our experiments indicate that DAG method is more suitable for practical use in Liaohe Basin than the other two methods.In order to achieve better results of lithology identification,the part of incorrect classification of SVM is analyzed：1.The performance of SVM model depends on the penalty factor and kernel parameter,so the choice of optimization methods of these two parameters affects the classification accuracy to a certain extent;2.In the establishment of SVM sample space,the number of training samples also affects the classification accuracy;3.Due to lithological characteristics of volcanic rocks in Liaohe Basin,logging response among different lithology rocks appear cross,the classification accuracy is affected.The approach combining binary SVM with three multiclass SVM can effectively classify the basalt,non-compacted basalt,trachyte,non-compacted trachyte,gabbro and diabase located in Liaohe Basin,China.1.SVM can distinguish the boundary and thickness of different lithological formation;2.SVM can′t distinguish the lithology with similar mineral composition and different texture/structure;3.The classifiers of binary SVM and multiclass SVM were built based on conventional well logging data.In lack of imaging logging data,elemental capture spectroscopy（ECS）and other special logging data,SVM method is still applicable to lithology identification.