标题: 运用资料探勘技术于分析生产良率之研究 -以矽晶圆产业的S公司为例
Applying Data Mining Techniques to Analyze Production Yield - A Case Study of S company in Silicon Wafer Industry
作者: 毛恺宏
Mao, Kai-Hung
刘敦仁
Liu, Duen-Ren
管理学院资讯管理学程
关键字: 资料探勘;Weka;简单贝氏;决策树;支持向量机;Data Mining;Weka;C4.5;Naive Bayes;SVM
公开日期: 2015
摘要: 矽晶圆是太阳能电池最重要的零件,从无到有需要较长的生产时间,生产过程有意外就无法赶不上客户的交期,良率的好坏是影响获取订单的关键,因为找出影响良率好坏的品质分析方法就更重要。传统的品质分析方法,如特性要因分析图、柏拉图、管制图、六标准差等在制造业提升良率品质的运用已经行之有年,然而这种解析方法是依据经验累积来增加效率,其作法并不稳定,过程也是缺乏效率。
资料探勘是从海量资料中发掘出隐藏于其中具有参考价值的资讯,可以提供给决策者做出决定的辅助,进而提升生产良率品质。
本研究藉由MES系统撷取制程生产履历记录及所使用的生产参数及机台别,透过专业人士分析过滤出有用的生产属性,去除无效的资料,透过资料增益的方法找出关键的生产属性,分别再以决策树C4.5、简易贝式分类及支持向量机SVM三种分类演算法训练资料、预测生产制程的良率,实际结果显示支持向量机SVM表现最佳。
As the most important component of “Solar Cell”, the lead time of “Silicon Wafer” production is considerable. Any mistake happened during production process may fail to reach promised delivery time. Under this circumstance, production yield rate is the key to increase customer satisfaction and acquire more orders. Taking into account all these factors, it will be meaningful how to discover a quality analysis methodology which will impact production yield rate. Traditional methodologies, such as Cause-and-Effect diagram、Pareto Analysis、Control Charts、Six Sigma, are applied in manufacturing industry for quite a while, nevertheless all these paradigms were built up based on personnel experience which is not considered stable.
Data mining, on the other hand, is a process to discover hidden relative information or specific models from tens of thousands of messages. The findings after data-mining process will then be a reference for decision makers to improve or stabilize production yield rate.
Through MES system, this research aims to record all production histories, production parameters, and individual machine situation during a full-production cycle. After that, professionals will sift through abovementioned information and conclude key production drivers. Finally, Decision Tree C4.5, Naive Bayes, and SVM are applied respectively to foresee production yield rate. The performance of SVM is the best among the three classification models.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT070263419
http://hdl.handle.net/11536/126527
显示于类别:Thesis