【作者】 闫伟；

【导师】 何桢；

【作者基本信息】 天津大学 ， 管理科学与工程， 2012， 博士

【摘要】 复杂产品是指客户需求复杂、产品组成复杂、产品技术复杂、制造过程复杂、项目管理复杂的一类产品。在复杂产品质量控制中，质量特性监控点的有效性决定了产品质量的可控性。但是随着质量控制点数量的增多，一方面会使产品的控制成本急剧增加，另一方面会使企业的质量控制效率大幅下降。为了提高产品质量控制效率，就必须有效识别出对产品质量具有显著影响的关键质量特性，以减少控制点，提高控制效率，降低企业成本。本文主要从平衡数据集和不平衡数据集两个角度（以发动机叶片为例，如果合格叶片与不合格产品叶片数量均衡就称为平衡数据；如果合格是不合格几十倍就称为不平衡数据）对关键质量特性识别问题进行了研究：（1）在复杂产品质量类别内样本数量大概相等的平衡数据集中，将信息论中的信息熵概念引入，用信息增益来判断质量特性与所属类别之间的相关性，这样就可以绕过产品质量特性间复杂的相互依存关系，从一个新的角度对质量特性的重要性进行度量，从而识别出真正的产品CTQ。经算例验证，该法可以有效的降低质量特性维度，提高质量控制效率和控制水平，节约了大量时间和成本。（2）在真实的复杂产品生产中，质量类别内样本数量差距往往都很大，这样的不平衡数据集为关键质量特性的识别带来了比平衡数据集更大的困难。本文分别从几个角度建立了不同的不平衡数据集CTQ识别方法：第一种，对ReliefF方法的判别标准进行改进，使得类别划分标准向多数类偏移，以此降低少数类数据被作为异常值删除的风险；第二种对Wrapper方法进行修改，将SBS、SFS与代价敏感学习整合，以此建立质量特性循环选择机制，有效提高了CTQ识别效率；第三种对EM算法进行改进，通过聚类过滤掉不平衡数据中的冗余样本以此构建平衡数据集，并在此基础上进行关键质量特性识别，以此有效改善了质量特性识别性能并大幅降低第二类错误率。本文通过对复杂产品生产企业的调研，对目前质量控制的瓶颈---关键质量特性识别进行了详实的分析和研究，并通过算例进行了验证。所以，本文对未来复杂产品生产质量控制的研究具有积极的参考价值。更多还原

【Abstract】 Complex product is a kind of product which featured with complexcustomer demand, complex product composition, complex producttechnology, complex product technology and complex project management.In the quality control process of complex product, the validity ofmonitory points measuring quality characteristic determines thecontrollability of product quality. However, with the increase in thequantity of quality control points, on the one hand, the cost of productquality control will increase dramatically; on the other hand, theefficiency of enterprise quality control will drop significantly. Inorder to improve the efficiency of product quality control, it is of greatsignificance to identify Critical-to-quality Characteristics that havea significant impact on product quality, so as to reduce the quantitycontrol points, improve control efficiency and finally cut the businesscosts.In this dissertation, the study of identification ofCritical-to-quality Characteristics was carried out form twoperspectives: balanced data sets and unbalanced data sets.(1) In the balanced data sets with approximate equal sample size ofcomplex product quality category, the information entropy technology ofinformation theory was introduced and information gain was employed todecide the correlation between quality characteristics and theirrespective categories. Thus, the complex interdependent relationswithin product quality characteristics can be bypassed and the measureof importance of quality characteristics can be conducted form a newperspective, so as to identify true product CTQ. An example testify showsthat this method can effectively reduce the dimension of the qualitycharacteristics, improve the efficiency of quality control and controllevel and save a lot of time and cost in the meanwhile. (2) In the real manufacture of complex product, the difference insample size is large within quality category. And the unbalanced dataset brings about greater difficulty in the identification ofCritical-to-quality Characteristics compared with the balanced data set.In this dissertation, three different CTQ recognition methods have beendeveloped from three angles in unbalanced data set. First, improvementshave been made on ReliefF criterion to make category dividing criterionoff set towards most classes, so as to reduce the risk of minority classdata being removed as outliers. Second, base on the introduction ofmodels and algorithms of feature selection, a kind of Wrapper featureselection algorithm has been proposed on the basis of balancedclassification accuracy, thereby reducing the negative impact of theunbalanced data set on the CTQ identification. Third, improve the EMalgorithm. Filtering out the redundancy samples in the unbalanced dataset by clustering so as to build a balanced data set, based on which therecognition of Critical-to-quality Characteristics have been conducted.By this means, the performance of the quality characteristics has beenimproved effectively and error rate of the second kind has been reducedsignificantly.Based on the investigation of complex product manufacturers, thedissertation carried out a detailed analysis and research on the presentbottleneck in quality control--Critical-to-quality Characteristics,and verification has also been made through an example. The study hasa positive reference value on future study of complex product qualitycontrol.更多还原