Yves Chandon, Six Sigma Black Belt, Freescale Semiconductor
Every automotive analog integrated circuit is measured by 100 to 5,000 tests before customer delivery. Principal components analysis has been used as a variable dimension reduction technique to analyse those fairly large data files. However, for the subject matter expert, interpretation of the principal components is not always straightforward. The variable clustering procedure is a practical tool. Each cluster can be analysed, and the most representative variable for each cluster is displayed. For the subject matter expert, it is fairly convenient to fit all the variables in a cluster by the most representative with bivariate fits using confidence ellipses. This information can be useful for yield enhancement, quality improvement or cost-saving test time reduction. It is also possible to enhance the correlations between the set of most representative variables and the other variables by using stepwise regression or the Partition platform. Partial least squares can also be used to model all variables by the most representative of each cluster that are not orthogonal like principal components. The global understanding of the product variability structure is significantly enhanced by those techniques. This information is used for yield and cost optimisation, as well quality improvement.
