Severstal improves the quality of etched rolled products using a neural network

2023: Metal quality control at continuous etching unit No. 4

Severstal"" monitors the quality of metal on continuous etching unit No. 4 Cherepovets Steel Mill (a key asset of Severstal) using a proprietary model computer vision VERA. The solution based neural networks on helps to find deviations in the quality of rolled metal and helps to reduce the number of notifications from customers. Severstal announced this on October 10, 2023.

The computer vision model receives images from high-resolution cameras and then detects and classifies defects on rolls on a special server with high-performance graphics processors. After that, the experts carry out certification and make a decision on the compliance of the products with the client's requirements. To assess the quality of the model, two metrics of incorrect positives are monitored: "overhauling" (how often the model confuses classes of defects) and "underhauling" (how often the model does not detect defects).

The share of customer notifications about deviations in the surface of finished products decreased by about 1.8 times in the first half of this year compared to the same period in 2022. The economic effect amounted to about 2 million rubles for the same period.

Digitalization of product quality certification is one of the advantages of Severstal as a manufacturer. The VERA neural network has already proven its effectiveness at Mill 2000 and now analyzes the metal passing through the continuous etching unit No. 4 no less carefully. Subsequently, it can be scaled to other key areas. The development of our own quality control systems allows Severstal to be independent of foreign manufacturers of such equipment, as well as to adapt its own solutions to production tasks and reduce their cost, "commented Petr Mishnev, Director of Technical Development and Quality of Severstal.

The model is being continuously improved by Severstal Digital developers together with experts from Severstal's Digital Quality Development Center, who conduct further training. NTA-4 became the third unit on which this model runs. It was first introduced at Camp 2000 in 2020. In 2022, the volume of customer claims for surface defects decreased by 1.5 times compared to 2021. Then the decision was scaled to NTA-3, where over the year the volume of claims decreased by 4.8 times.