基于标签传播的岩性预测半监督学习算法研究

Study of semi-supervised learning for lithology prediction based on label propagation

低渗透油气藏、致密油气藏、页岩油气藏等非常规油气藏的开发已成为全球油气开发的热点,也为测井解释带来新的挑战.为了提高测井解释精度,本文研究了岩性预测的半监督学习问题,提出了“聚类—人工标注—伪标注—分类”的岩性预测框架.首先,利用聚类算法选取待标注样本;然后,基于数据在特征空间和地理空间的相似性,利用图半监督学习方法实现人工标注样本到无标注样本的标注传播;最后,基于伪标注的置信度,采用加权支持向量机算法实现分类模型的设计与训练.本文在真实的测井数据上进行了大量的实验,发现半监督学习算法通过挖掘有标注数据和无标注数据中蕴含的分布特性,可获得更精确的岩性预测效果,即使对于不均衡的数据集,也能大幅提高分类模型在各类别上的准确率.进一步引入地理空间相似性,半监督的岩性预测模型在样本数量少的类别上的准确率得到了较大提高,从而验证了本文所提出方法的有效性.

The development of low-permeability reservoirs,tight reservoirs,shale reservoirs and other unconventional reservoirs has become a hot research topic in the society of petroleum engineering,which also brings new challenges to well logging interpretation.In order to elevate the accuracy of well logging interpretation,we studied the semi-supervised learning problem of lithology prediction in this paper,and put forward the framework of“clustering—manual labeling—pseudo labeling—classification”for lithology prediction.First,the clustering algorithm is used to select samples to be labeled.In this paper,the k-means clustering algorithm is employed.Second,based on the similarity of data in feature space and geographic space,the graph semi-supervised learning method is used to realize the label propagation from manually labeled samples to unlabeled samples.Finally,according to the confidence of pseudo-labels,the weighted support vector machine algorithm is used as the base classifier to predict the lithology classes of samples,where the weights of samples are relied on their label confidence.We have carried out experiments on real logging data.The experimental results show that the semi-supervised learning algorithm can analyse the distribution of both labeled samples and unlabeled samples,and thus improve the lithology prediction performance.Even for the imbalanced datasets,the semi-supervised learning algorithm can increase the classification accuracies on different lithology classes.The introduction of similarity in geographic space can further improve the prediction performance,especially that of lithology classes with small quantity of samples.Thus the effectiveness of the proposed method is verified.