On-line Classification of Surface Defects in Hot Rolling Processes

This paper develops a methodology for the online classification of different types of surface defects in rolling processes. The method developed consists of three major steps: feature generation, feature selection, and classifier design. Based on the engineering knowledge of rolling processes, a set of potential features are generated from the local regions of sensing images that contain the suspicious surface defects. Then, a feature selection method based on the Genetic Algorithm (GA) and a linear classifier is developed to eliminate redundant features. Finally, a new classifier, named weighted hierarchical SVM, is proposed to classify different types of surface defects. This newly designed classifier can improve the classification accuracy for those defects that have similar shapes or/and have small sample sizes. A case study is conducted in a real world rolling process to demonstrate the implementation procedures and the effectiveness of the developed methods.