Abstract
A methodology was developed using neural network theory to predict the occurrence of out of control process parameter conditions in a composite board manufacturing facility. Three weeks of process parameter data were collected from the manufacturing operation. Multi-variable linear regression and time series analysis techniques were utilized initially to analyze the data set. A valid regression model could not be developed due to the presence of serial correlation in the data set. Time series analysis did not result in the development of a valid model because nonstationarity was present in the data set. The nonstationarity could not be removed using differencing techniques or power and logarithmic transformations. Neural network theory was chosen as an alternative approach. Feed forward back-propagation neural networks, with one hidden layer, were successfully trained to predict the classification of bonding treatment process parameters. The bonding treatment classification was based on the operating condition of the process with respect to the statistical process control limits. The bonding treatment values were classified as one of three possible conditions: above the upper process control limit, within the process control limits, or below the lower process control limit. The inputs to the network included data representing the current process condition along with historical data on relevant parameters, including moisture contents, bulk densities, and temperatures. Two training data sets were constructed, consisting of 30 and 60 data examples, respectively. Each training data set contained equal numbers of examples of the three process operating conditions. The best networks trained using the smaller training data set correctly predicted 60 percent of the bonding treatment test values. The best networks trained using the larger training data set correctly predicted the process state of control for 73 percent of the test values. These results indicate that the neural network back-propagation learning algorithm was able to identify and extract patterns from the training data sets to allow the prediction of future values of bonding treatment.
Cook, Deborah Faye (1990). A predictive modeling system for manufacturing process parameters. Texas A&M University. Texas A&M University. Libraries. Available electronically from
https : / /hdl .handle .net /1969 .1 /DISSERTATIONS -1174767.