摘要
针对小样本情形下的相关多质量特性优化设计问题,在运用田口(Taguchi)质量损失函数度量多质量特性的稳健性基础上,本文提出了使用多变量偏最小二乘(Partial Least-Squares,PLS)回归模型处理此类问题的新方法,并结合实际的工业案例进行了分析。结果表明,本文所提出的新方法能够有效地处理相关多质量特性的优化设计问题。
Widely used in the optimization of product or process design,robust parameters can effectively improve product quality and economic benefits.A company needs to consider multiple quality attributes in product and process design in order to successfully compete with rivals.Although design optimization of multiple quality characteristics plays a central role in continuous quality improvement activities,it needs to resolve some problems.For instance,the current literature usually ignores the correlation between quality characteristics.This paper tries to resolve this major problem by using the partial least-square（PLS） method and Taguchi quality loss function to measure the robustness of multiple quality characteristics.The PLS method integrates the principal component analysis,canonical correlation analysis and multiple linear regression analysis in the modeling process.The MPLS method considers independent and dependent variables.Hence,the PLS method can effectively solve multicollinearity problems between independent variables and dependent variables,and has no certain sample size requirements.Moreover,the PLS method can implement visual analysis and observe using various auxiliary analysis techniques via mapping multidimensional data in two dimensions.The correlated quality characteristics are usually converted into new synthetic variables（i.e.extracted components） with the best explanative power,which can effectively reduce dimensions of independent and dependent variables,and make it easier to deal with design optimization problems.We discussed the concrete procedure of design optimization with the correlated multiple quality characteristics.The first section of the paper used Taguchi＇s quality loss function to compute the quality loss of each quality attribute.The second section of the paper created a regression model to analyze the relationship between independent（i.e.controllable and noise factors） and dependent variables（i.e.quality loss of each quality attribute）.The importance of each factor was also investigated via the variable importance in projection（VIP） index.We then identified significant factors by the degree of their importance.The third section of the paper established an objective function using optimal conditions,and calculated the effects of different factors using the normalized value of the objective function.The last section of the paper computed the normalized optimization loss residual（NOLR） of different factors and their combinations.Our research results validate research findings of previous studies and improve the optimization of factor combination.Different from the principal component analysis and improved principal component analysis,this paper uses a novel approach to effectively reduce the number of dimensions for multiple responses and conflicts caused by multi-response optimization problems.Moreover,the design optimization method of multiple quality characteristics based on PLS can take full advantage of various auxiliary analysis technologies to obtain more precise information.We were able to obtain a better predictive model using the cross validation analysis.Significant factors were identified using VIP indices.The discriminant analysis of specific samples was analyzed using the T2 ellipse diagram to reduce the influence of outliers on the regression model.More importantly,the manufacturing process of complex products involves a high degree of correlation between variables.This paper makes theoretical and practical contributions to the quality management field.
出处
《管理工程学报》
CSSCI
北大核心
2011年第2期66-73,共8页
Journal of Industrial Engineering and Engineering Management
基金
国家自然科学基金资助项目(70672088
70931002)
国家自然科学基金国际交流资助项目(70711140386)
关键词
多质量特性
相关性
偏最小二乘回归
优化设计
multiple quality characteristics
optimization design
partial least-squares
quality loss function
