High dimensional outlier screening of small dice samples for aerospace IC reliability

Abstract

This article is based on a collaborative project between a production company in the aerospace industry (Atmel) and a statistical company (Ippon Innovation). The main objective is to develop an innovative advanced tool to detect multivariate outliers in small samples, based on measurements of thousands of parameters, which is called a high dimensional situation in statistics. After presenting the context and current computational methods used in this industry for the screening of abnormal dice, the article introduces two methods that are designed for dealing with the special case of highdimensional datasets: the ROBPCA and the GAT algorithms. The two methods are compared in a case study. GAT has a definite advantage over other methods in detecting atypical instances. Finally, the integration of this algorithm with the production tool provides the ability to go back to the real measurements involved in the revealed anomalies. The use of a sound statistical method to adress small samples and high dimension data is needed to detect reliability issues in the space industry.