Severstal continues to scale computer vision model complexes that allow for automated quality control of rolled metal products. Today, 12 systems have been implemented at the site of the Cherepovets Metallurgical Plant. The solutions were developed by the Severstal Digital team (Artificial Intelligence cluster), together with the Directorate for Technical Development and Quality and the Severstal Repair Directorate.
Both neural network systems based on previously installed video surveillance cameras and new surface inspection systems based on the company's own designs have been implemented at the CherMK site.
So, at the end of 2024, four new M&A systems were put into operation: they work on polymer metal coating units No. 1,2,3 and on slitting unit No. 8. These solutions also work successfully on continuous hot-dip galvanizing units No. 1 and No. 4 and on slitting unit No. 4. They are able to recognize from 40 to 100 classes of defects in color images, which facilitates the classification of deviations. The systems use a set of computer vision models, each of which is trained to detect defects critical for a particular type of metal. When changing the type of product, for example, from cold-rolled steel to dynamic steel, the algorithm automatically switches to the desired model, and additional defect classification rules flexibly adapt to the type of rolled metal.
In addition, the developed software provides adaptive lighting settings depending on the time of day and the color tone of the coating, which allows you to get a better image for further recognition of defects by the neural network model. This is especially important for systems based on polymer coating aggregates, where up to several dozen color variants of polymer rolled metal are produced.
"Computer vision-based solutions are an important part of the digital product certification system. Due to their precision and ability to work 24/7, Severstal's customers receive higher-quality rolled metal, and unit operators can switch from routine tasks to those requiring their greater involvement. The combination of multimodel algorithms, adaptive technologies and import-independent solutions not only improves product quality, but also forms a new benchmark for the metallurgical industry. Over the past two years, we have improved the accuracy of our solutions by more than 30%. In the future, we plan to implement inspection systems at other key units, bringing entire production lines under digital control," commented Svetlana Potapova, Director of Severstal Digital.
