The Directorate for Technical Development and Quality of the Severstal-Russian Steel Division and Severstal Digital LLC (part of PJSC Severstal) are improving the software used at the industrial site of the Cherepovets Metallurgical Plant (CherMK, part of the Severstal-Russian Steel Division) for quality control of the surface of rolled metal. A new proprietary development of the company's specialists - the VERA neural network. It is being implemented on a high-tech and most productive hot rolling mill in Russia - Mill 2000.
VERA is an improved analogue of the EVE neural network, which Severstal employees developed in 2019 and is used to classify surface defects of rolled metal on cut-to-length line No. 4 in the metal finishing workshop No.2 of the flat-rolled steel production of CherMK.
EVE is able to find four types of deviations - captivity, crack, mechanical defect and cavity - on digital images of the metal surface. The program became the first in-house development of the company to detect metal surface defects and replaced the original defect classification algorithm of the Parsytec solution. The network receives images from Parsytec cameras, and then on a special server with high-performance graphics processors detects and classifies defects, determines their parameters. Information about the found defects is displayed on the operator's screen. EVE finds three times more real defects in comparison with its foreign counterpart, as well as 13 times less common defects. The network also sorts out many small non-critical and false defects, reducing the workload on staff and increasing the productivity of the shop.
The new program - the VERA neural network - is a similar, but more complex algorithm for finding and classifying more deviations. She will be able to find already 19 classes of defects. The program includes an anomaly detector that will allow you to identify atypical images. Also, Severstal Digital employees have updated the user interface, making it more convenient and adding new functions, for example, the concept of "like-dislike", which allows the operator to point out the models to errors and correct solutions in non-standard situations. This approach allows us to adapt to changing conditions: the model is periodically retrained on the basis of operator corrections and adapts to the technological process.
volume of information. At the same time, our neural networks allow us to achieve a higher accuracy of results and make it possible to customize algorithms for specific technological needs, ”notes Petr Mishnev, Director of Technical Development and Quality of the Severstal Russian Steel Division.
“ We continue to develop the digital solutions for product quality control. The EVE model has proven its worth and has laid a solid foundation for future implementations. VERA has become even more accurate in finding defects and has a more complex architecture. Computer vision is in great demand in production, and we hope to extend this technology to other units of the plant, ”comments Boris Voskresensky, director of Severstal Digital.