改进的协同训练半监督SVM在油层识别中的应用  

Improved Co-training Semi-supervised SVM and Its Application in Oil Layer Recognition

作　　者：潘用科 贺紫平 夏克文 牛文佳 PAN Yongke;HE Ziping;XIA Kewen;NIU Wenjia(School of Electronics and Information Engineering,Hebei University of Technology,Tianjin 300401,China)

机构地区：[1]河北工业大学电子信息工程学院,天津300401

出　　处：《郑州大学学报（工学版）》2022年第1期14-19,26,共7页Journal of Zhengzhou University（Engineering Science）

基　　金：国家自然科学基金资助项目(42075129);河北省重点研发计划项目(19210404D,20351802D)。

摘　　要：实际石油测井中有标签数据获取代价昂贵,而大量低廉的无标签数据未被使用,如何利用有限的有标签样本及大量的无标签样本获取准确的油层分布有待解决。半监督学习方法因能同时利用少量有标签样本及大量无标签样本便可获取良好的分类模型而被广泛应用。因此,基于半监督支持向量机(S3VM),提出一种改进的基于量子行为粒子群优化(QPSO)的协同训练S3VM油层识别算法(QPSO-CS3VM)。首先引入多视图的协同训练策略,构造2个独立的初始分类器提高识别精度;然后为提高初始分类精度,引入了量子行为粒子群算法以优化S3VM;最后引入一种改进的近邻数据剪辑方法用于预测无标签样本伪标签的置信度,从而避免因错分样本导致的模型性能恶化的问题。通过对具有代表性的两口井的测井数据进行油层识别,结果表明:改进的协同训练半监督SVM相较于传统的协同训练算法在两口井中的识别率分别提升了5.00百分点和3.12百分点。所提算法油层识别精度较高,有一定的实际应用意义。It is expensive to obtain labeled data in actual oil logging,and a large amount of cheap unlabeled data are not used.How to use limited labeled samples and a large number of unlabeled samples to obtain accurate oil layer distribution remains to be solved.The semi-supervised learning methods were widely used because they could obtain good classification models using both a small number of labeled samples and a large number of unlabeled samples.Therefore,based on a semi-supervised support vector machine(S3 VM),an improved semi-supervised support vector machine based on co-training and quantum-behaved particle swarm optimization algorithm(QPSO-CS3 VM)was proposed for oil layer recognition.Firstly,the multi-view-based co-training strategy combined with S3 VM was used to construct two independent initial classifiers,and then exchanged and labelled unlabeled samples to improve the overall oil layer recognition accuracy.Secondly,in order to improve the initial classification accuracy of original classifiers,the quantum behavioral particle swarm algorithm was introduced to optimize S3 VM.Finally,a newly nearest neighbor data editing approach was used to predict the confidence of the pseudo-labelling of unlabeled data to reduce the deterioration of model perfor-mance caused by misclassification of data.The improved co-training semi-supervised SVM proposed in this paper improved the classification accuracy by 5.00%and 3.12%compared to the traditional co-training algorithm by performing oil layer recognition on the logging data of the two wells.The algorithm proposed in this paper had high accuracy in oil layer recognition and had practical application.

关 键 词：半监督支持向量机 协同训练 量子行为粒子群优化 数据剪辑 油层识别 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]