Studies in Industrial Statistics

NSF DMS-9704711 Topics in Industrial Statistics Dennis K.J. Lin Penn State University PROJECT SUMMARY A major objective of the National Science Foundation is to improve the nation's capacity for intellectual and economic growth. It does this by supporting the discovery of new knowledge and the enhancement of a skilled workforce. American industry is becoming increasingly aware of the benefits of running statistically designed experiments. In complex experimental work, the number of potential factors is large, but often only a few active factors are expected. A problem frequently encountered is how to identify the factors that matter, and consequently improve the quality and productivity of the products. The basic results we hope to obtain here will be extremely useful to practitioners, since new designs will be furnished that will save considerable cost when real world issues are involved. When the number of potential factors is large and only a few active factors are expected, we do not need unbiased estimates to detect those active factors. In this project, we propose the use of supersaturated designs -- allowing for a slight bias in estimation but significantly reducing the design size in experimental work. Some initial results were actually utilized in electronic and chemical industries (specifically, IBM and Lonza). The proposed methodology helps extract useful information that may not be provided by 'traditional wisdom.' From these practical experiences, it is anticipated that a much wider application is possible. On the other hand, this research has broad theoretical implications. Some theoretical results have been obtained, and it is shown that those results can be applied to other statistical areas, such as linear models, information theory and optimal design. Another statistical area that has received a great deal of attention in industry involves reliability problems. The estimation of system (or component) reliability is a common problem for many industries. When a system consists of several components, and when the cause of failure is unspecified, the estimation problem is unresolved. This project proposes the use of a Bayesian approach. Some initial results applied to IBM PS/2 personal computer data have been obtained. More general fundamental research will be further studied. Problems given here are very specific to be realistically solved during the funding period. However, it has a much wider research impact for both industrial applications or academia type research.