| Organizations within the paper industry have dedicated significant time and resources into the development and application of modeling and simulation techniques and prediction tools. This research looks to apply some of these software tools to make the best use of real world data from various sensor locations in a secondary fiber recovery mill.;This research explored two separate problem areas in the mill and analyzed machine learning models and prediction tools which were best suited for the challenges presented. The first part of this research explores data analysis techniques to optimize a neural network model and a subsequent sensitivity analysis to predict and analyze the sources of variability in moisture content on the paper machine wet end.;The second part of this work explores some tools which could help a paper mill operator or engineer better assess when a paper break is likely to occur. The model developed will show how to make better use of information from operator logs by incorporating this information into a predictive model. This problem was especially challenging because of the great deal of variability in the industrial environment and because of the influence of many outside factors. The methods explored in this research proved useful in identifying clusters which represent various modes of good and poor operation on the paper machine.;A data set was provided for this research by the RockTenn Paper Mill in Solvay, NY. This mill is a 100% recycled containerboard paper mill. As with other secondary fiber recovery systems, this mill has a great deal of variability in its incoming raw material which makes it even more of a challenge to develop predictive models.;Keywords. Artificial Neural Networks, K-Means, Clustering, Machine-Learning, Fault Detection, Supervised Learning, Unsupervised Learning. |
