Application of hierarchical neural networks to pattern recognition for quality control analysis in steel-industry plants

Our paper focuses on the classification of surface defects in flat rolled strips in steel industry. Since this work aims at the classification of samples organized in a hierarchical way it seems natural to use a hierarchical approach. We choose a hierarchical neural architecture, based on the multilayer perceptron, which, to some extent, combines classification trees with neural network approaches. We exhaustively tested the proposed architecture in the classification of surface defects in flat rolled strips on real plant data, obtaining a higher classification accuracy with respect to the state-of-the-art technologies. This approach can be generalized to many other industrial classification problems.