Semi-supervised learning for steel surface inspection using magnetic flux leakage signal

This paper proposes a semi-supervised learning model for detecting multi-defect classification and localization on the steel surface for industries with limited labeled datasets. This study uses 1-D data from magnetic flux leakage (MFL) testing, a powerful and cost-effective nondestructive inspection method for steel bars. Most steel surface defect systems are based on supervised learning classification with 2-D image datasets. However, acquiring labeled datasets for developing supervised learning models is practically limited in the actual steel manufacturing process. Furthermore, due to the frequent occurrence of multiple defect classes on the same steel bar, the problem of multi-defect classification and localization needs to be addressed. Therefore, this paper proposes a steel bar surface inspection system for multi-defect classification and localization based on a semi-supervised learning model and MFL signals. The proposed system solves the multi-defect classification and localization problem by reducing the feature dimension with an autoencoder. Then, it classifies the defects based on the semi-supervised support vector machines that require only a small portion of the labeled dataset. Also, the classification process is repeated on the overlapped small steel section to address the multi-defect classification and localization issue. When it was evaluated on an industry MFL inspection dataset, the accuracy ranged from 81% to 90% when the labeled data ratio varied from 2% to 90%.