Big Data Analytics: Prediction of Surface Defects on Steel Slabs Based on One Class Support Vector Machine

Abstract

Visible surface defects are common in steel products, such as crack or scratch defects on steel slabs (a main product of the upstream production line in a steel production line). In order to prevent propagation of defects from the upstream to the downstream production lines, it is important to predict or detect the defects in earlier stage of a steel production line, especially for the defects on steel slabs. In this paper, we address the problem regarding the prediction of surface defects on continuous casting steel slabs. The main goal of this paper is to accurately predict the occurrence of surface defects on steel slabs based on the online collected data from the production line. Accurate prediction of surface defects would be helpful for online adjusting the process and environmental factors to promote producibility and reduce the occurrence of defects, which should be more useful than only inspection of defects. The major challenge here is that the amounts of samples for normal cases and defects are usually unbalanced, where the number of defective samples is usually much fewer than that of normal cases. To cope with the problem, we formulate the problem as a one class classification problem, where only normal training data are used. To solve the problem, we propose to learn a one-class SVM (support vector machine) classifier based on online collected process data and environmental factors for only normal cases to predict the occurrence of defects for steel slabs. Our experimental results have demonstrated that the learned one class SVM (OCSVM) classifier performs better prediction accuracy than the traditional two-class SVM classifier (relying on both positive and negative training samples) used for comparisons.