基于多分类器集成的工业品缺陷分析方法研究

发布时间：2018-01-14 20:43

本文关键词：基于多分类器集成的工业品缺陷分析方法研究　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：制造工业产品缺陷的分析是改进企业产品制造过程的重要途径之一,对于产品质量以及营销收益有着重要的研究意义和应用价值。伴随计算机技术的快速发展、自动化系统的全面部署,产品制造过程中信息采集和存储的难度大大降低。具有潜在信息和价值的数据在不断地积累。同时,机器学习、数据挖掘等方法在各行各业取得了飞速的发展和应用。然而制造业由于工业性质对这些数据的利用水平远不如其它行业,并没有真正地发挥这些数据应有的价值。为此,本文针对制造工业品数据的主要特点,总结了一般的针对工业产品缺陷分析问题的处理流程,对数据进行处理以及统计分析,并将分析产品各项质量检测结果与产品的缺陷数据之间的关系问题,转化成通过统计学习方法建立产品质量与缺陷的分类模型。然而缺陷数据同时出现多个缺陷类别以及类别样本数目不平衡的问题,这对分类算法模型的构建而言是一大阻碍。本文针对需要同时扫清该两者障碍提出了结合代价敏感与集成方法的多分类器模型,通过样本重赋权重再缩放的方法结合分类代价敏感,再集成多个决策树构建多分类模型。实验结果表明该模型可以有效地处理不平衡类别的多分类问题,同时可以平衡分类代价和预测的准确率。此外对决策树的集成拟合可以得出相关属性的重要性度量,可以作为追溯缺陷主要影响因素的一个依据。
[Abstract]:Analysis of manufacturing industrial product defect is one of the most important ways to improve the enterprise production process, and has important research significance and application value for the quality of the products and marketing revenue. With the rapid development of computer technology, the full deployment of automation system, information collection and storage products in the manufacturing process has the potential to greatly reduce the difficulty of information and value. The data is accumulated ceaselessly. Machine learning method, at the same time, data mining has achieved rapid development and application in all walks of life. However, due to the nature of these manufacturing industry data use level is far behind that of other industries, and these data did not really play its due value. Therefore, this thesis mainly manufacture of industrial products the data, summed up the general on industrial product defect analysis processing procedures, data processing and statistical analysis, The analysis and the relationship between the detection results of defect data quality products, into learning classification model based on product quality and defects by statistic method. However, the defect data appear at the same time a number of samples and the type of defect category imbalance problem, which is a major impediment to the construction of classification model according to the need to clear away the obstacles. The two also proposed multi classifier model combining cost sensitive and integration method, through the method of sample weight to weight the combination of cost sensitive classification and zoom, and integration of multiple decision tree to construct multi classification model. The experimental results show that the model can effectively deal with unbalanced classes of multi classification problems. At the same time can balance the cost of classification accuracy and prediction. In addition to measure the importance of integrated fitting decision tree can draw relevant attributes, can As a basis for the main influencing factors of retroactive defects.

【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18

【参考文献】