Controlling Twist on Predicted Value

Quality control of paper and paperboard is complex. A few properties can be measured online but most of the quality characteristics are relying on destructive testing where sampling can only be done at the end of every tambour that occur approximately every 45 minutes. This sample is about 30 cm in Machine Direction and during the same time a paperboard machine that runs at 600 m/min will have produced 2,7 km of paperboard. This leaves the production blind for variation in between these samples and the answer from any corrective action is very long.Twist is the property describing the lack of out-of-plane dimensional stability that is oriented neither in Machine Direction or Cross-machine Direction. The problem is well-studied in paper mechanics and is caused by a difference in fiber direction in the outer plies of multiple ply paperboard combined with the hygroexpansivity properties of cellulose fibers. Controlling the property in the ongoing production is complex for an operator since the problem is depending on multiple machine parameters.The present research attempts to find a way to predict the outcome of the Twist value continuously on the ongoing production and guide the operators when and how to adjust the process. This will be done using machine learning algorithms and domain knowledge from previous research in the occurrence of twist, combined with the traditional concept of Statistical Process Control. The focus on the research will lie in implementation and usability.