Defect prediction in hot strip rolling

Abstract

In steel manufacturing, enormous amounts of data are gathered from many kinds of processes and it is difficult to distinguish useful knowledge from the resulting extensive databases. The recorded databases in the rolling process contain hundreds of features, and new methods are needed to reveal the novel and useful information. In the present study, the data from a hot strip rolling process were analysed in order to identify the rolling conditions in which some common defects, such as too thin or too narrow strip, do and do not occur. At the beginning, the dataset was reduced from its original size of over 200 features to 17% by using basic statistical analysis and linear correlation. After this, self-organising maps, parallel coordinates display, and k means clustering were used to find out the conditional probabilities of the common defect types. As a result, the method presented here revealed the rolling conditions for the four common defect types, i.e. coiling temperature close to the limits, too thin strip, too narrow strip, and torn tail end. Most important, the knowledge gained can be used to reduce the number of these defects.