Doctoral Dissertation Announcement
Candidate: Jason S. Trahan
Doctor of Philosophy
Department: Industrial and Manufacturing Engineering
Title: Decision Support Systems in Product Development: A Bayesian Network Approach for Injection-Molded Products
Dr. Paul Engelmann, Chair
Dr. Steven Butt
Dr. Abdolazim Houshyar
Dr. Leonard Lamberson
Dr. Daniel Mihalko
Date: Thursday, November 2, 2006 1:00 p.m.- 3:00 p.m. CEAS, Room C-258
Abstract: Attribute defects cause a majority of the scrap for many manufacturers injection-molding aesthetic parts, yet these defects receive little attention during product development. Methods for preventing attribute defects have been hampered by the inability to quantify complex cause-effect relationships. However, the qualities of a burgeoning branch of knowledge-based systems, called Bayesian networks, have benefited various fields faced with similar circumstances. Bayesian networks have the capability of presenting causal relationships between variables in a graphical structure and quantitatively explaining their effects throughout that structure. These effects can be characterized by probabilities derived from sources, such as equations, historical data, rules-of-thumb or expert knowledge.
Thus, a Bayesian network was selected to model the decisions and uncertainty involved in preventing shear splay in polypropylene products. Shear splay is an attribute defect that occurs when shear generated during processing causes the polymer to degrade and release volatiles, which appear as silvery streaks on a part’s surface. The shear splay network consisted of 17 conditional probability tables, 859 probabilities and 37 variables from areas of materials, part and mold design and processing.
A computer-based system was created to elicit 296 subjective probabilities from 25 experts in the field of plastics. A user interface based on past research was designed to interact with experts to efficiently provide relatively unbiased and consistent probabilities. A bootstrap method using trimmed means of their responses was employed to develop a statistic for the probabilities, make inferences about the sample and assess the variability. Interpolation and modified shear rate equations were used to calculate the remaining probabilities.
Both the qualitative and quantitative aspects of the network were evaluated by introducing evidence to validate relationships, comparing data of existing products against actual results and allowing several experts to review the model first-hand. Overall, the network performed as expected. It displayed the ability to be used as a decision-support system in preventing shear splay and explained cause-effect relationships that until now have never been quantified. In addition, the elicitation software and method for analyzing response data showed potential for becoming a standard in building Bayesian networks.
(Now payable by CC Online)