Machine Vision and Deep Learning for Enhanced Grading and Classification of Surface Wear on Hot-Rolling Work Rolls

The wear of work rolls significantly affects the production efficiency and product quality. However, existing methods for wear assessment fail to effectively quantify work roll surface wear conditions, thereby affecting the quality control of steel strips and maintenance strategies for rolls. To accurately assess the wear conditions of hot-rolling work rolls, this study initially established an apparatus for capturing high-precision roll surfaces images. Subsequently, a quantitative assessment of common surface wear morphologies was conducted, and a hot-rolling work roll surface wear dataset was constructed. The MobileNetV2 convolutional neural network (CNN), augmented by transfer learning, was employed to develop a MobileNetV2-wear detection and classification (WDC) surface wear grading model. A comparison with mainstream CNN models revealed that the MobileNetV2-WDC model achieved high-speed (21.92 ms) and accurate (96.86%) grading with minimal model parameters (2.27 M) and size (27 M), meeting the industrial efficiency and practicality requirements. A visual analysis of the model classification errors was conducted, outlining paths for further optimization. This study provides an efficient and accurate solution for detecting and grading surface wear on hot-rolling work rolls, enhancing product quality and extending the lifespan of rolls.