基于机器学习方法的测井岩相分类研究  

Study on well-log lithofacies classification based on machine learning methods

作　　者：高飞 曲志鹏 魏震[3] 朱剑兵 程远锋 GAO Fei;QU ZhiPeng;WEI Zhen;ZHU JianBing;CHENG YuanFeng(College of Geology and Mining Engineering in Xinjiang University,Vrümqi 830017,China;Geophysical Research Insitute of Sinopec,Dongying 257000,China;Laboratory of Continental Dynamics and Metallogenic Prognosis of Central Asian Orogenic Belt in Xinjiang University,Ürümqi 830017,China)

机构地区：[1]新疆大学地质与矿业工程学院,乌鲁木齐830017 [2]中国石化胜利油田物探研究院,东营257000 [3]新疆大学中亚造山带大陆动力学与成矿预测实验室,乌鲁木齐830017

出　　处：《地球物理学进展》2024年第3期1173-1192,共20页Progress in Geophysics

基　　金：新疆维吾尔自治区基金项目“基于物理数据双驱的薄储层地震智能识别方法”(2022D01C422);新疆维吾尔自治区“天池英才”引进计划项目“含流体孔隙介质波动力学及油气识别”联合资助。

摘　　要：目前,在基于测井数据的岩相分类研究中,机器学习算法是一个研究热点.然而,实际上钻井取心极少,岩相样本匮乏,机器学习算法就会遇到过拟合问题,从而导致测井岩相分类效果不佳.因此,本文研究小样本环境下不同机器学习算法在测井岩相分类工作中的预测效果.以北美Panoma油气田数据集为例,通过逐步减少训练样本数量,建立了四种训练模式.同时,选用三类有代表性的有监督学习算法,以评估小样本环境下不同算法的预测效果,包括基于一般梯度下降的线性回归分类算法、支持向量机算法和一维卷积神经网络算法.在综合评价算法的预测效果时,选用了岩相分类准确率、总体岩相分类F1值、各岩相分类F1值以及有效识别最大相数.结果表明,随着训练样本数量的减少,三种算法的预测效果并未呈现线性下降趋势,且一维卷积神经网络较另外两种算法表现更为稳健.Machine learning algorithms have been widelyapplied in the study of lithofacies classification based onwell-log data at present.However,with very few drilledcores and a scarcity of lithofacies samples,machinelearning algorithms encounter overfitting problems.Thisleads to poor well-log lithofacies classification.Therefore,the prediction effectiveness of different machine learningalgorithms were studied in small sample environment,taking the North American Panoma oil and gas field datasetas an example.Four training models are established bygradually reducing the number of training samples.Meanwhile,three different types of supervised learningalgorithms were selected to evaluate the predictioneffectiveness,including Linear Regression Classificationbased on General Gradient Descent(GGD-LRC),SupportVector Machine(SVM),and One-Dimensional ConvolutionalNeural Networks( 1D-CNN).In the comprehensiveevaluation of the prediction effectiveness of the algorithms,selected the lithofacies classification accuracy,the overalllithofacies classification F1 value,the individuallithofacies classification F1 value,and the maximumnumber of effectively identified lithofacies.The resultsshow that the prediction effectiveness of the threealgorithms does not present a linear downward trend as thenumber of training samples decreased,and the 1D-CNNalgrithm is more robust than the others.

关 键 词：测井岩相分类 机器学习 线性回归分类 支持向量机 一维卷积神经网络 

分 类 号：P631[天文地球—地质矿产勘探]