Data dependent methods of modeling and predicting of glass quality.

摘要

The goal of this work was the development of a methodology for finding causes for flaws in flat glass. A second goal was the development of a process for analysis of large data sets. Other processes with unknown time delays between input data and output results are possible recipients for these methods.; The physical nature of the process is to be included as much as possible in the analysis steps. While the procedures are statistical, any advantage that studying the process and the data from it are to be included. In the case of flat glass, cause and effect results from large trends tend to be easily identified but small-scale changes that may lead to upsets in the form of flaws are not as easily understood.; The result of the process is not necessarily a best final model although a good model is desired. The process engineer can study the list of possible causes and the delay at which each cause occurred and use his process knowledge to better determine the real cause of the flaw. The output of the study will be a list of possible causes for each flaw and some ranking of the causes for the flaws.; The process described for finding variables that contribute to the flaw was a success. This process has never been proposed before as discussed both in literature and through anecdotal verification. Over a number of data sets, some variables were found to consistently contribute to the model. Even when the entire data range did not give a good regression model, some variables were identifiable as contributors to the flaw. The same algorithms developed for glass flaws were also used on a blind data set and successfully found hidden but causal variables in the data sets provided.; In this study is an analysis of large data sets. The same procedures can be useful in any of a number of process industries. Results of the work may also be useful to studies outside manufacturing altogether. Studies of stock-market performance have even been mentioned as a possible use of this work.