机器学习法在碳酸盐岩岩相测井识别中应用及对比——以四川盆地MX地区龙王庙组地层为例

doi:10.13809/j.cnki.cn32-1825/te.2021.04.015

摘要/Abstract

摘要：

机器学习法是碳酸盐岩岩相测井识别的主要技术手段,针对不同地质情况和资料,选择适用的机器学习方法是高精度识别岩相的关键因素之一,然而针对机器学习识别岩相方法的适用性研究较少,为此列举了4种最常用的机器学习识别岩相方法：自组织神经网络聚类分析法（SOM）、基于图像多分辨率聚类分析法（MRGC）、K最近邻分类算法（KNN）和神经网络法（ANN）。通过对比方法的原理及方法的实际应用效果,总结出这4种机器学习法的优缺点和适用性,少量岩心样本情况下,优选MRGC法;在较多数量岩心资料情况下,优选KNN或MRGC法。其在四川盆地MX地区龙王庙组地层岩相识别应用中表明：MRGC和KNN法效果最好,其次为SOM法,ANN法效果最差。不同机器学习方法实际应用及对比分析成果,对于碳酸盐岩岩相测井识别方法在其他层组或其他工区的应用起到借鉴作用,并具有较强的实用价值。

关键词: 碳酸盐岩岩相, 测井识别, 机器学习, 龙王庙组, 四川盆地

Abstract:

The machine learning method is the main technical means of carbonate lithofacies log identification. Selecting the appropriate machine learning method according to the different geological conditions and data is one of the key factors for high-precision identification of lithofacies. However, there are few researches on the applicability of machine learning identification methods. In this paper, four most commonly used machine learning methods for identifying lithofacies are studied, including Self Organizing Maps（SOM）, Multi-Resolution Graph-based Clustering（MRGC）, K Nearest Neighbor（KNN）, and Artificial Neural Network（ANN）. By comparing the principle and practical application effects of these methods, the advantages, disadvantages and applicability of the four machine learning methods have been summarized. When there are few core samples, MRGC is preferred, while when there are more core data, KNN is preferred as well as MRGC. Their application of lithofacies identification in the Longwangmiao Formation in the MX area in Sichuan Basin shows that MRGC and KNN are the best, SOM is the second, and ANN is the worst. This study of the application effects of machine learning methods provides a guidance for the identification of carbonate rock facies in other layers and regions, and has strong practical value.

Key words: carbonate facies, logging identification, machine learning method, Longwangmiao Formation, Sichuan Basin

中图分类号:

TE19

李昌,沈安江,常少英等. 机器学习法在碳酸盐岩岩相测井识别中应用及对比——以四川盆地MX地区龙王庙组地层为例[J]. 油气藏评价与开发, 2021, 11(4): 586-596.

Chang LI,Anjiang SHEN,Shaoying CHANG, et al. Application and contrast of machine learning in carbonate lithofacies log identification: A case study of Longwangmiao Formation of MX area in Sichuan Basin[J]. Petroleum Reservoir Evaluation and Development, 2021, 11(4): 586-596.

图/表 11

表1

图1

图2

表2

图3

图4

图5

图6

图7

表3

表4

参考文献 23

[1]	沈安江, 赵文智, 胡安平, 等. 海相碳酸盐岩储集层发育主控因素[J]. 石油勘探与开发, 2015, 42(5):545-554.
	SHEN Anjiang, ZHAO Wenzhi, HU Anping, et al. Major factors controlling the development of marine carbonate reservoirs[J]. Petroleum Exploration and Development, 2015, 42(5):545-554.
[2]	周进高, 徐春春, 姚根顺, 等. 四川盆地下寒武统龙王庙组储集层形成与演化[J]. 石油勘探与开发, 2015, 42(2):158-166.
	ZHOU Jingao, XU Chunchun, YAO Genshun, et al. Genesis and evolution of Lower Cambrian Longwangmiao Formation reservoirs, Sichuan Basin, SW China[J]. Petroleum Exploration and Development, 2015, 42(2):158-166.
[3]	司马立强, 陈志强, 王亮, 等. 基于滩控岩溶型白云岩储层分类的渗透率建模方法研究——以川中磨溪—高石梯地区龙王庙组为例[J]. 岩性油气藏, 2017, 29(3):92-102.
	SIMA Liqiang, CHEN Zhiqiang, WANG Liang, et al. Permeability modeling based on the classification of beach-controlled karst dolomite reservoirs:a case from Longwangmiao Formation in Moxi-Gaoshiti area, central Sichuan Basin[J]. Lithologic Reservoirs, 2017, 29(3):92-102.
[4]	YE S J, RABLILLER P. A new tool for electro-facies analysis:multi-resolution graph-based clustering [C]// Paper SPWLA-2000-PP presented at the SPWLA 41st Annual Logging Symposium, Dallas, Texas, June 2000.
[5]	张振城, 孙建孟, 马建海, 等. 利用测井资料自动识别藻灰岩[J]. 吉林大学学报(地球科学版), 2005, 32(3):382-388.
	ZHANG Zhencheng, SUN Jianmeng, MA Jianhai, et al. Automatic identification of algae limestone with well logging data[J]. Journal of Jilin University(Earth Science Edition), 2005, 32(3):382-388.
[6]	AGHCHELOU M, SHAHVAR M B. Lithofacies estimation by multi-resolution graph-based clustering of petrophysical well logs: Case study of South Pars Gas Field of Iran [C]// Paper SPE-162991-MS presented at the Nigeria Annual International Conference and Exhibition, Lagos, Nigeria, August 2012.
[7]	于建华. 应用人工神经网络自动识别岩性[J]. 石油地球物理勘探, 1993, 28(1):59-66.
	YU Jianhua. Automatic lithology identification using artficail nervous netwlork[J]. Oil Geophysical Prospecting, 1993, 28(1):59-66.
[8]	ALPANA B, HANS B H. Determination of facies from well logs using modular neural networks[J]. Petroleum Geoscience, 2002, 8:217-228. doi: 10.1144/petgeo.8.3.217
[9]	HONG T, NIALL T, MEDDAUGH W S. Successful carbonate well log facies prediction using an artificial neural network method: Wafra Maastrichtian reservoir, partitioned neutral zone(PNZ), Saudi Arabia and Kuwait [C]// Paper SPE-123988-MS presented at the SPE Annual Technical Conference and Exhibition, New Orleans, Louisiana, October 2009.
[10]	HASSAN M, MOOSA E, ELHAM H A. Application of artificial neural networks for prediction of Sarvak Formation lithofacies based on well log data, Marun oil field, SW Iran[J]. Geopersia, 2015, 5(2):111-123.
[11]	桑应宾. 基于K近邻的分类算法研究[D]. 重庆:重庆大学, 2009.
	SANG Yingbin. Research on classification algorithm based on K-nearest neighbor[D]. Chongqing: Chongqing University, 2009.
[12]	周少华, 付略, 梁宝鎏. 基于SOM神经网络的古代青瓷聚类分析[J]. 中国科学E辑, 2008, 38(7):1089-1096.
	ZHOU Shaohua, FU Lue, LIANG Baoliu. Clustering analysis of Ancient Celadon based on SOM neural network[J]. Science in China(Series E), 2008, 38(7):1089-1096.
[13]	付小青, 张爱明. 基于SOM的入侵检测算法的特征选择[J]. 华中科技大学学报(自然科学版), 2007, 35(7):5-7.
	FU Xiaoqing, ZHANG Aiming. Feature selection of SOM-based intrusion detection algorithm[J]. Journal of Huazhong University of Science and Technology(Natural Science Edition), 2007, 35(7):5-7.
[14]	KOHONEN T. Self-organized formation of topologically correct feature maps[J]. Biological Cybernetics, 1982, 43:59-69. doi: 10.1007/BF00337288
[15]	CHANGA H C, KOPASKA-MERKELB D C, CHEN H C. Identification of lithofacies using Kohonen self-organizing maps[J]. Computers & Geosciences, 2002, 28:223-229. doi: 10.1016/S0098-3004(01)00067-X
[16]	STUNDNER M, OBERWINKLER C. Self-Organizing maps for lithofacies identification and permeability prediction [C]// Paper SPE-90720-MS presented at the SPE Annual Technical Conference and Exhibition, Houston, Texas, September 2004.
[17]	张猛. 地层岩性的多层次自组织自动分类方法研究[D]. 武汉:长江大学, 2018.
	ZHANG Meng. Study on multi-level self-organization automatic classcification method for stratigraphy lithology[D]. Wuhan: Yangtze University, 2018
[18]	吴进波. MRGC方法在储层分类及渗透率精细评价中的应用[J]. 海洋石油, 2017, 37(1):49-53.
	WU Jinbo. Application of reservoir classification and permeability elaborate evaluation with MRGC method[J]. Offshore Oil, 2017, 37(1):49-53.
[19]	寇莎莎, 魏振军. K-最近邻的改进及其在文本分类中的应用[J]. 河南师范大学学报(自然科学版), 2005, 33(3):134-136.
	KOU Shasha, WEI Zhenjun. Improvement of K nearest neighbors and applications in text classification[J]. Journal of Henan Normal University(Natural Science Edition), 2005, 33(3):134-136.
[20]	孙岩, 吕世聘, 王秀坤, 等. 基于结构学习的KNN分类算法[J]. 计算机科学, 2007, 34(12):184-186.
	SUN Yan, LYU Shiping, WANG Xiukun, et al. K-nearest neighbor algorithm based on learning structure[J]. Computer Science, 2007, 34(12):184-186.
[21]	张涛. 用人工神经网络研究元素与岩浆岩、沉积岩岩性及沉积岩矿物的关系[D]. 西安:西北大学, 2016.
	ZHANG Tao. The relationships between rock elements and the igneous rocks, the lithologic discrimination and mineral identification of sedimentary rocks: A study based on the method of artificial neural network[D]. Xi’an: Northwest University, 2016.
[22]	金民东, 曾伟, 谭秀成, 等. 四川磨溪—高石梯地区龙王庙组滩控岩溶型储集层特征及控制因素[J]. 石油勘探与开发, 2014, 41(6):650-660.
	JIN Mindong, ZENG Wei, TAN Xiucheng, et al. Characteristics and controlling factors of beach-controlled karst reservoirs in Cambrian Longwangmiao Formation[J]. Petroleum Exploration and Development, 2014, 41(6):650-660.
[23]	张建勇, 罗文军, 周进高, 等. 四川盆地安岳特大型气田下寒武统龙王庙组优质储层形成的主控因素[J]. 天然气地球科学, 2015, 26(11):2063-2074.
	ZHANG Jianyong, LUO Wenjun, ZHOU Jingao, et al. Main origins of high quality reservoir of Lower Cambrian Longwangmiao Formation in the Giant Anyue Gas Field, Sichuan Basin, SW China[J]. Natural Gas Geoscience, 2015, 26(11):2063-2074.

SOM法测井相类型	岩心标定概率百分比					MRGC法测井相类型	岩心标定概率百分比
SOM法测井相类型	颗粒云岩		粉晶云岩		泥晶云岩	MRGC法测井相类型	颗粒云岩		粉晶云岩	泥晶云岩
测井相1	86.4	13.6		0		测井相1	80.6	19.4		0
测井相2	0	100.0		0		测井相2	0	100.0		0
测井相3	0	0		100.0		测井相3	100.0	0		0
测井相4	100.0	0		0		测井相4	0	100.0		0
测井相5	0	93.3		6.7		测井相5	84.6	15.4		0
测井相6	0	100		0		测井相6	0	0		100.0
测井相7	95.5	0		4.5		测井相7	11.8	70.6		17.6
测井相8	16.7	83.3		0		测井相8	73.7	26.3		0
测井相9	0	0		100.0		测井相9	0	0		100.0

岩相	岩性
颗粒云岩相	颗粒云岩,砾屑云岩、砂屑云岩,细晶云岩等粗结构的白云岩
粉晶云岩相	粉晶云岩、细—粉晶云岩等粉晶级别的白云岩
泥晶云岩相	泥晶云岩,泥质泥晶云岩,泥质泥—粉晶云岩等

序号	岩相	厚度（m）	声波时差（AC）（μs/ft）	自然伽马（GR）（API）	中子（NPHI）（V/V）	密度（DEN）（g/cm³）	深电阻率（RT）（Ω·m）
1	颗粒云岩	4.25	48.24	17.71	0.082	2.77	103.59
2	颗粒云岩	5.49	47.92	19.10	0.091	2.79	51.91
3	颗粒云岩	1.24	49.05	17.63	0.064	2.74	425.84
4	颗粒云岩	16.99	49.35	17.60	0.052	2.73	809.46
5	颗粒云岩	5.91	50.08	18.04	0.061	2.70	595.41
6	颗粒云岩	0.93	47.98	20.53	0.050	2.75	818.09
7	颗粒云岩	0.81	47.46	19.45	0.041	2.83	1 649.04
8	颗粒云岩	1.04	45.08	19.17	0.034	2.83	4 071.69
9	粉晶云岩	1.17	43.62	32.17	0.036	2.87	1 036.42
10	粉晶云岩	4.75	45.91	18.58	0.070	2.81	381.87
11	粉晶云岩	1.76	45.08	28.45	0.054	2.82	777.23
12	粉晶云岩	0.70	44.14	22.85	0.041	2.82	1 258.25
13	粉晶云岩	0.70	44.22	26.39	0.030	2.87	13 848.18
14	粉晶云岩	0.58	44.06	16.49	0.035	2.83	8 054.69
15	粉晶云岩	0.81	43.86	19.15	0.035	2.80	9 937.23
16	粉晶云岩	0.73	48.18	18.53	0.047	2.79	506.02
17	粉晶云岩	1.97	50.59	18.58	0.055	2.72	726.43
18	粉晶云岩	1.85	49.47	18.28	0.064	2.72	469.56
19	粉晶云岩	2.55	47.49	19.94	0.040	2.80	2 024.54
20	粉晶云岩	5.56	46.51	19.49	0.041	2.81	1 858.67
21	粉晶云岩	6.60	45.57	18.94	0.036	2.81	7 662.91
22	泥晶云岩	2.00	46.45	25.89	0.060	2.80	210.82
23	泥晶云岩	3.89	44.62	26.17	0.045	2.84	530.76
24	泥晶云岩	2.66	43.65	22.57	0.032	2.86	10 060.70
25	泥晶云岩	3.36	43.77	22.41	0.038	2.83	7 761.66
26	泥晶云岩	4.29	43.49	20.58	0.029	2.83	12 356.61
27	泥晶云岩	1.85	43.64	19.02	0.027	2.86	7 996.26
28	泥晶云岩	13.78	46.15	32.67	0.042	2.82	2 495.68

方法	优缺点		应用条件		方法关键点
方法	优点	缺点	测井资料要求	岩心资料要求	方法关键点
自组织映射神经网络聚类分析（SOM）	容错性强,稳定性好, 具有自联想性	人工经验主导聚类测井相数目,步骤较多	至少5项常规测井资料（中子、密度、声波、自然伽马、电阻率）	少量岩心资料	岩相和测井相对应关系
基于图像多分辨率聚类分析（MRGC）	善于解决类域交叉或重叠较多的问题	人工经验主导聚类测井相数目,人为影响因素大,步骤多		少量岩心资料	岩相和测井相对应关系
K最近邻分类算法（KNN）	善于解决类域交叉或重叠较多的问题, 无需训练,方法简单	样本数量不均衡或较少时,预测偏差大		中等数量岩心资料,且每类岩相类型样本数量均衡	不同类型样本数量要均衡
神经网络方法（ANN）	较强容错能力及自学习自适应能力	容易陷入局部极小化问题,对样本数量依赖性强		大量岩心样本资料	样本数据量大而且要具有代表性