Adaptive multi-resolution graph-based clustering algorithm for electrofacies analysis

一种自适应多分辨率图聚类测井相分析方法

Logging facies analysis is a significant aspect of reservoir description.In particular,as a commonly used method for logging facies identification,Multi-Resolution Graph-based Clustering(MRGC)can perform depth analysis on multidimensional logging curves to predict logging facies.However,this method is very time-consuming and highly dependent on the initial parameters in the propagation process,which limits the practical application effect of the method.In this paper,an Adaptive Multi-Resolution Graph-based Clustering(AMRGC)is proposed,which is capable of both improving the efficiency of calculation process and achieving a stable propagation result.More specifically,the proposed method,1)presents a light kernel representative index(LKRI)algorithm which is proved to need less calculation resource than those kernel selection methods in the literature by exclusively considering those"free attractor"points;2)builds a Multi-Layer Perceptron(MLP)network with back propagation algorithm(BP)so as to avoid the uncertain results brought by uncertain parameter initializations which often happened by only using the K nearest neighbors(KNN)method.Compared with those clustering methods often used in image-based sedimentary phase analysis,such as Self Organizing Map(SOM),Dynamic Clustering(DYN)and Ascendant Hierarchical Clustering(AHC),etc.,the AMRGC performs much better without the prior knowledge of data structure.Eventually,the experimental results illustrate that the proposed method also outperformed the original MRGC method on the task of clustering and propagation prediction,with a higher efficiency and stability.

测井相分析是通过自动聚类方法对多维测井曲线进行分析,进而进行相聚类与预测。基于图的多分辨率聚类(MultiResolution Graph-based Clustering,MRGC)方法是一种常用的测井相分析方法,然而MRGC算法非常耗时,并且在传播过程中高度依赖初始参数,实际应用效益差。本文提出了一种自适应多分辨率图聚类(Adaptive Multi Resolution Graph based Clustering,AMRGC)分析方法。该方法不仅能提高测井相计算效率,而且能获得稳定的测井相传播结果。本文方法的两个核心算法是:1)轻核代表指数(L-KRI)算法只需计算少量"自由吸引"点,有效提高了计算效率;2)采用了反向传播算法(BP)与多层感知器(MLP)神经网络,有效避免了传统K近邻算法因随机初始化参数导致的不稳定结果。实验结果表明,本文方法在聚类和传播预测任务上优于传统的MRGC方法,具有更高的运行效率和稳定性;同时,在没有数据先验知识的条件下效果明显优于自组织映射(SOM)、动态聚类(DYN)和自底向上的层次聚类(AHC)等其它常用聚类方法。