粗糙集的支持向量机的优化方法研究及在资源评价中的应用

发布时间：2019-06-17 13:05

【摘要】：如今,隐伏矿产逐渐成为勘查重点,但由于其埋藏较深,只能通过各种地物化遥的数据间接预测找矿靶区,针对这种多维数据,究竟哪些属性和隐伏矿产联系密切,哪些没有联系,粗糙集的理论方法就在剔除冗余属性信息方面具有很好的筛选能力。对已知的一些隐伏矿产各属性特征通过机器学习建立分类模型,然后预测新的找矿靶区,如今已成为一个非常热门的研究方向。基于VC维和结构风险最小化的统计学习理论的支持向量机分类模型,具有很强的理论基础和分类能力,其能很好地解决小样本、非线性、过学习、维数灾难和局部极小等问题。结合粗糙集对多维数据约简降维的预处理功能,能够达到很好的分类效果和泛化能力,使得找矿靶区的预测更加准确。本文展开了如下工作:1)在云南个旧锡铜矿试验区内,基于ArcGIS对物探、化探数据用反距离插值的方法生成栅格数据,并在试验区随机提取500个随机样本点和锡铜矿矿区提取100随机样本点,用栅格数据提取至点的方法赋予其物探、化探条件属性值,在锡矿一定范围内设置一个buffer缓冲区,给缓冲区内外的样本点赋予不同的决策属性值,构建一个完整的决策属性系统。2)基于条件属性取值的连续性,在MATLAB编程实现时,主要采取邻域粗糙集的方法,设置一个合适的邻域半径,对训练样本进行离差标准化处理,对41个物化探条件属性进行属性约简,尝试通过构建模糊因子的方法优化约简算法,用基于属性重要性的方法对每一个约简的属性赋予权重,并尝试基于挑选属性的先后顺序赋予属性相应的权重。编写KNN算法排除奇异点时,设计合适的参数,并结合ArcGIS地统计分析的概率直方图、半变异协方差云分析等综合剔除噪声数据,并挑选边界域的训练样本点作为最终的SVM模型训练样本集,完成邻域粗糙集的预处理工作。3)在MATLAB中,选取高斯核函数对训练样本构建SVM模型,通过十折交叉验证的检验方法优化模型参数,得到最优分类模型,最后遍历整个个旧锡矿试验区,对整个区域做锡矿矿产资源评价分析,并通过改变邻域半径对比不同的属性约简和模型预测评价系统。
[Abstract]:Nowadays, hidden minerals have gradually become the focus of exploration, but because of their deep burial, they can only indirectly predict the prospecting target area through a variety of geophysical and remote data. Aiming at this kind of multidimensional data, which attributes are closely related to hidden minerals and which are not related, the theory and method of rough set has a good screening ability in eliminating redundant attribute information. It has become a very popular research direction to establish classification models for some known attributes of hidden minerals through machine learning and then to predict new prospecting targets. The support vector machine classification model based on VC and structural risk minimization theory has a strong theoretical basis and classification ability, which can solve the problems of small sample, nonlinear, over-learning, dimension disaster and local minima. Combined with the preprocessing function of rough set to reduce dimension of multidimensional data, it can achieve good classification effect and generalization ability, and make the prediction of prospecting target area more accurate. In this paper, the following work has been carried out: 1) in the Gejiu tin copper mine test area of Yunnan Province, the geophysical and geochemical data are generated by inverse distance interpolation based on ArcGIS, and 500 random sample points are randomly extracted from the experimental area and 100 random sample points are extracted from the tin copper mining area. The geophysical exploration is given by the method of extracting the grid data to the point, and a buffer is set up in a certain range of tin ore. Different decision attribute values are given to the sample points inside and outside the buffer, and a complete decision attribute system is constructed. 2) based on the continuity of conditional attribute values, the method of neighborhood rough set is mainly adopted in MATLAB programming, a suitable neighborhood radius is set, the deviation of training samples is standardized, and 41 geophysical and geophysical conditional attributes are reduced. This paper tries to optimize the reduction algorithm by constructing fuzzy factor, gives weight to the attributes of each reduction by the method based on the importance of attributes, and tries to give the corresponding weights to the attributes based on the sequence of selected attributes. When the KNN algorithm is written to eliminate singular points, the appropriate parameters are designed, combined with the probability histogram of ArcGIS statistical analysis and semi-variance covariance cloud analysis, the noise data are eliminated synthetically, and the training sample points in boundary domain are selected as the final training sample set of SVM model to complete the preprocessing of neighborhood rough set. 3) in MATLAB, Gao Si kernel function is selected to construct SVM model. The optimal classification model is obtained by optimizing the model parameters by the test method of ten fold cross verification. Finally, the whole test area of Gejiu tin mine is traversed, and the mineral resources of tin ore in the whole area are evaluated and analyzed, and the different attribute reduction and model prediction and evaluation systems are compared by changing the neighborhood radius.
【学位授予单位】：石家庄经济学院
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：P624

【参考文献】