基于主成分分析方法的海量地震数据属性降维优化

Attribute Reduction and Optimization for Massive Seismic Data Based on Principal Component Analysis

针对传统的地震数据属性降维优化方法所选取的地震数据属性特征贡献率低导致降维过程计算量大、CPU占用率高等问题,提出一种基于主成分分析的海量地震数据属性降维优化方法。首先根据地震样本特征建立地震数据特征矩阵,把矩阵中的特征进行聚类,运用降序法排列聚类结果,选取前几项数据作为地震数据属性特征选取结果,对其结果评估分类信息量;通过特征积分准则(FSC)修正分类信息量,获取海量地震数据属性特征节点;运用主成分分析方法对地震数据属性特征节点主成分添加标签,确定Fisher判别分析与PCA可变动选择不确定关系,建立半监督降维的全局最优化形式,运用特征值分解计算降维结果,克服海量地震数据属性降维过程中的过拟合问题,融合主成分分析算法与Fisher判别分析算法实现海量地震数据属性降维优化。实验结果证明,所提方法选取的属性特征精度及贡献率较高,降维过程中CPU占用率较低。

In view of the problems associated with the traditional optimization method for seismic data attribute reduction, i.e., the large amount of computation required in the reduction process and the high CPU occupancy rate, in this paper, we propose an attribute reduction and optimization method for massive seismic data based on principal component analysis (PCA). First, we establish a feature matrix of the seismic data based on the characteristics of seismic samples. The features in the matrix are then clustered and arranged in descending order. We then select the first few data as the seismic data attribute feature results and evaluate the classification information of these results. Next, the classification information is modified using the feature integral criterion to obtain the attribute feature nodes of the massive seismic data. We use PCA to label the principal components of the attribute nodes of the seismic data and establish a global optimization of the semi-supervised dimensionality reduction. The dimensionality reduction results are calculated by eigenvalue decomposition, we solved the problem of over-fitting in the attribute reduction process of massive seismic data, and realized the optimization of the attribute reduction of massive seismic data by combining the PCA algorithm with Fisher discriminant analysis. The experimental results show that the proposed method has a high accuracy and contribution rate of attribute feature selection, and the CPU occupation rate is low during the dimensionality reduction process.