June 25, 2020

A team of specialists from Severstal Digital, part of PAO Severstal, one of
the world’s largest steel and steel-related mining businesses, working with
experts in flat rolled products from the Cherepovets Steel Mill have
successfully increased the productivity of a machine learning model that
controls the speed of the Mill’s continuous pickling line #3 (NTA-3).

Adelina, a digital model in use at Severstal’s continuous pickling line #3
(NTA-3) since November 2019, has now been joined by “Ruban”, a new artificial
intelligence agent based on a deep – reinforcement learning algorithm. Both
products were developed by Severstal in-house using open source

Adelina and Ruban now work in parallel with one another; Adelina controls
the speed of the unit, and Ruban adjusts the speed to achieve optimal results.
This partnership has made the production process more flexible and secure, as
the model and agent are able to adjust the speed of the unit every second and
respond instantly to any unforeseen situation.

Evgeny Vinogradov, CEO of Severstal Russian Steel Division, commented:

“The Adelina model had already met our expectations, demonstrating an
initial increase in productivity of NTA-3 by more than 5 percent. In March
2020, we produced a record volume of pickled metal at this unit – more than 130
thousand tons. After introducing the Ruban agent, we recorded a further 1.5
percent increase in productivity, and we estimate that using the two
technologies in parallel could provide more than 80 thousand tons of additional
metal each year. This is a remarkable increase for one of the most significant
units in the production of flat rolled products.”

Ruban differs from classic machine learning models, learning not just from
historical data, but independently, by exploring the digital twin of NTA-3. The
operating speed at the unit largely depends on the parameters of the passing
steel strip – the length, width and thickness of the roll, its steel grade and
temperature, among other factors. Ruban learns from combinations of different
parameters, specifically created for it by a generative adversarial network,
which uses two neural networks to generate new data. It also sets a production
plan and creates unique situations for training purposes. For effective
learning, the agent was assigned a training system based on rewards and
penalties; Ruban experiments to find a solution where the reward amount
surpasses the penalties as far as possible.

Boris Voskresenskii, Chief Digital Officer of Severstal, commented:

“The use of reinforcement learning to control production units is not
widespread, particularly in metallurgy. We believe the use of Artificial
Intelligence at NTA-3 to be the first such case in Russian practice. The
performance improvement recorded on NTA-3 following the introduction of digital
tools proves that a data-driven approach has a great future in the industry,
and we are moving in the right direction.”