Most engineering systems have multiple inputs and multiple outputs. For example, a semiconductor manufacturing system consists of thousands of fabrication steps with numerous inline production parameters affecting multiple electrical characteristics of final chips. Many-to-many analysis is thus needed to more effectively discover critical factors causing poor product qualities or a low production yield. Though methodologies of many-to-many correlation analysis have been proposed in the literature, difficulties arise, especially when there exist multicollinearity effects among features, to measure the relative importance of a feature’s contribution. Relative weight analysis offers a general framework for determining the relative importance of features in multiple linear regression models. In this article, we propose a many-to-many comprehensive relative importance analysis based on canonical correlation analysis to effectively summarize the relationship between two sets of features. Simulation and actual semiconductor yield-analysis cases are used to show the proposed method, as compared to other conventional methods, in analysis of two sets of features.

大多数工程系统具有多个输入和多个输出。例如，半导体制造系统由数千个制造步骤组成，这些制造步骤具有影响最终芯片的多种电气特性的众多在线生产参数。因此，需要进行多对多分析才能更有效地发现导致产品质量差或产量低的关键因素。尽管文献中已经提出了多对多相关分析的方法，但要衡量特征贡献的相对重要性，特别是当特征之间存在多重共线性效应时，会出现困难。相对权重分析为确定多个线性回归模型中要素的相对重要性提供了一个通用框架。在本文中，我们提出基于典范相关分析的多对多综合相对重要性分析，以有效地总结两组特征之间的关系。通过仿真和实际的半导体良率分析案例，与其他常规方法相比，该方法可用于分析两组特征。