基于随机森林的火山岩岩性识别方法

Lithology Identification Method of Volcanic Rocks Based on Random Forest

摘要: 随机森林模型是在决策树模型基础上发展而来的一种模式识别技术。与其他方法相比，该模型不仅能解决决策树模型分类过程中出现的过拟合现象，提升模型的泛化能力，还具备更高的分类精度。岩性识别本质上属于分类问题，本文利用随机森林泛化能力强以及分类准确性高的优势进行火山岩岩性识别，首先通过薄片分析、井壁取心、钻井取心等资料确定了X断陷火石岭二段133个岩性样本点以及岩性识别标准，以GR、CNL、DEN、AC、RLLD、RLLS测井曲线为样本属性，然后通过随机森林模型中的最大特征数、最大深度、决策树分类器的个数变化对岩性识别精度影响的数值模拟，建立了基于最大特征数、最大深度、决策树分类器个数的最优随机森林岩性识别模型，最后对15口重点井岩性进行测井识别，可识别安山岩、安山质(沉)火山角砾岩、安山质(沉)凝灰岩、流纹岩、流纹质沉火山角砾岩、流纹质(沉)凝灰岩6种岩性。将识别结果与FMI以及取心岩性进行对比验证，平均符合率为80.2%，结果表明：利用该方法识别岩性比录井确定的岩性更准确，满足X断陷岩性识别的要求。

Abstract: Random forests model is a pattern recognition technology developed on the basis of decision tree model. Compared with other methods, it can not only solve the over-fitting phenomenon of decision tree model classification, improving the generalization of the model, but also have higher classification accuracy. Lithology identification is essentially a classification problem. Therefore, the advantages of strong generalization of random forest and high classification accuracy are used in this paper. Firstly, 133 lithology samples and lithology identification standard of the second member of Huoshiling formation in X fault depression are determined by thin section analysis, borehole coring and drilling coring data. GR, CNL, DEN, AC, RLLD and RLLS are used to identify lithology as the sample attribute. Then, by simulating the influence of the maximum characteristic number, the maximum depth and the number of decision tree classifiers on lithology recognition accuracy in random forest model, a lithology recognition model based on random forest is established. Finally, the lithology of 15 key wells is identified. Six kinds of lithology including andesite, andesitic volcanic breccia, andesitic tuff, rhyolite, rhyolitic volcanic breccia and rhyolitic tuff are recognised. Compared with FMI and core lithology, the average coincidence rate is 80.2%. The results show that the lithology identification method is more accurate than that determined by mud logging and meets the requirements of lithology identification in X fault depression.