Detection of Manufacturing Defects in Steel Using Machine Learning

Abstract

Guaranteeing steel quality is a crucial step in the steel manufacturing process. Many manufacturing industries still resort to manual visual inspection which is inefficient and time-consuming. Industries have not fully adopted automated visual inspection due to inaccuracies, the variability of real-world manufacturing environments, and a lack of familiarity with the decisions output by the automated technology. Nonetheless, the implementation of automated defect detection systems can substantially enhance the quality of the end product. In particular, Convolutional Neural Networks have demonstrated exceptional abilities in tasks related to image classification and segmentation. There are still large rooms for improvement in terms of detection and localization accuracy of the algorithms, their practicality of use, and robustness of the algorithms. This paper employs and evaluates different semantic segmentation approaches with U-net and FPN architecture utilizing different CNN backbones. Additionally, the study enhances the model’s robustness by utilizing various data augmentation techniques. This work contributes to the field by improving the practicality, and robustness of CNN-based algorithms for steel surface defect detection and segmentation. Moreover, the study incorporates explainable artificial intelligence to provide insights into the decision-making processes of deep neural networks, bridging the gap in understanding. The research also includes a thorough assessment of the model’s transferability to different datasets that encompass a wide range of defects commonly observed in various industrial settings. The contributions of this work are in improving the practicality, robustness, and adaptability of CNN-based algorithms for steel surface defect detection and segmentation.

Date of Award	Aug 2023
Original language	American English
Supervisor	U Zeyar Aung (Supervisor)