Production Multivariate Outlier Detection Using Principal Components

Various aspects of using principal component and related analyses to detect outliers in multiple analog measurements made on digital CMOS circuits were investigated. The focus was on implementing practical production reliability screens with an extension to analog performance tests. Experimentally examined were outlier criteria, the reproducibility of the principal component signature, and simplifications to the characterization and test flow. It was found that the best of the 5 outlier criteria examined depended on the variability of, and the degree of correlation among, the variables. It is important to perform the analysis on a sample that covers the significant variance components. It was also determined that the principal component decomposition (the signature) of a chip design is repeatable enough to be performed in characterization prior to production for I_DDQ but not for a collection of other analog measurements.