Specialists of "Severstal digital" in conjunction with the Centre for technology
development shop etching production of flat rolled Cherepovets.
metallurgical plant (part of "Severstal Russian steel")
improved the model controlling the speed of the continuous pickling unit
No. 3 (NTA-3). It was connected intelligent agent, based on the
algorithm deep machine learning with reinforcement (reiforcemet learig,
RL), performing the functions of artificial intelligence.

"Ruban" works together with the digital model of the "Adelina", which was
installed on the CPL-3 earlier. "Adelina" calculates the speed control unit,
and the agent adjusts it for optimal results. This
provides the manufacturing process with greater flexibility and security, so
as a model and the agent is able to adjust the speed of the unit every second and
respond instantly to emergency situations.

"Model "Adeline" met expectations, showing an increase of performance of NTA-3
more than 5 percent. In March 2020, we produced a record volume
etched rental on this unit is more than 130 thousand tons. After connecting
intelligent agent productivity grew by 1.5%. According to estimates,
this can provide us with more than 80 million tons of additional production in the
year. This is a significant indicator for one of the most sought after units in
the production of flat products, – said General Director
division "Severstal Russian steel" and resource assets Eugene
grapes.

a New intelligent agent "Ruban" is different from the classic models
machine learning that is trained not according to historical data, and
alone, exploring the medium of digital double CPL-3. Speed control
the technological part of the machine largely depends on the parameters passing
steel strip – length, width and thickness of steel coil, steel grade, temperature
winding metal and other. "Ruban" studies on the combinations of parameters that
creates for him the generative-adversarial network. It also specifies
the plan rental and creates a situation that did not exist in history. For more
effective learning intelligent agent was appointed, the system of incentives
and fines. By experimentation "Ruban" looking for a solution in which the amount
of rewards exceeds the maximum amount of fines.

"Technology of machine reinforcement learning to control the units still
not widespread in industry and, in particular,
metallurgy. Artificial intelligence running on NTA-3 may be a first
a similar case of the Russian practice. Productivity
recorded at NTA-3 after the introduction of digital tools, proves that
have a data drive approach has a great future in the industry, and we are moving in
the right direction", – commented General Director of "Severstal
digital" Boris Voskresenskiy.

Полная версия