July 29, 2019
Severstal Digital LLC (a division of PJSC Severstal) has developed a neural network to detect surface defects in Metal Processing Shop No. 2 (TsOM-2), which produces flat rolled products at the Cherepovets steel mill (CherMK).
The Company has already launched a pilot project to confirm the model’s accuracy as one of the elements of automatic certification for confirming final product quality. Analysis of the first results using this model shows a significantly higher accuracy rate in detecting defects compared to the industrial equipment previously used to detect defects.
In order to both compare and enhance the model developed in-house by the Directorate for Technical Development and Quality (DTRK) and Severstal Digital with solutions offered by the world’s strongest Data Scientists teams, the Company has launched a competition on Kaggle, the world’s largest platform for machine learning competitions (part of the Google group). The competition opened on 26 July and will run over a 3-month period.
Igor Bardintsev, Chief digital officer of Severstal commented:
“Kaggle has already established itself as the premier platform for finding advanced solutions in the field of digital production. A number of leading international and Russian high-tech businesses have conducted competitions through this platform. Approximately 9,000 teams participated in one of the most recent competitions, but given the specificity of Severstal’s challenge, we expect about 1,000-3,000 teams from across the world to take part. The quality of solutions proposed by participants will be evaluated according to the Mean Dice Coefficient metric. In this way, participants’ decisions will be evaluated not only by their accuracy in classifying detected metal defects, but also by how correctly these defects were localized, and how their shapes and surface area were determined.”
The prize money amounts to a total of $120,000 thousand; for the first place the winner will receive $ 40,000 thousand, the second – $25,000, the third – $ 15,000, and the most efficient solution will receive an additional $15,000.
Igor Bardintsev added:
“Awarding an additional prize for the achieving the fastest solution allows us not just to encourage entrants to develop the most accurate models, but also to obtain solutions with a high speed of execution that are of an acceptable quality.”
