Severstal manages pelletizing with machine learning

2023: Implementation of a complex of machine learning models on the pelletizing line

Karelsky Okatysh (part of Severstal) has implemented a set of machine learning models on the pelletizing line. The system automatically controls the speed of rotation of the pelletizer and the dosage of bentonite and gives recommendations for the dosage of iron ore concentrate, which allows increasing the productivity of the unit while maintaining product quality. The solution was developed by a team of experts from Karelsky Okatysh and the IT function of Severstal. This was announced by the latter on December 5, 2023.

The most efficient roasting is possible when the proportion of pellets of class 10-12.5 mm prevails in the roasting machine: this ensures optimal porosity and gas permeability of the layer during heat treatment, which leads to an improvement in the quality of the finished product. Previously, operators did not have an indicator that could accurately determine this proportion: measurements were carried out visually, selectively and manually based on laboratory samples.

Now, based on the analysis of images from high-resolution cameras, the computer vision model calculates the particle size distribution of raw pellets and predicts the proportion of the required classes. Depending on this indicator, the rotation speed of the pelletizer and the dosage of bentonite and concentrate are adjusted. The solution makes it possible not only to control and control the pelletizing process, but also to standardize the operation of the roasting machine.

As a result of the use of the model, the performance of the pelletizing line increased by 11% with the preservation of product quality.

In 2018, Karelsky Okatysh had a similar project, but without using a neural network. As of December 2023, we have an order of magnitude higher reliability of the definition of grain size than it was then. The neural network more precisely defines the contour and dimensions of the pellets, including those that the first layer hides. The system allows continuous monitoring in the stream, which ensures automatic and prompt decision-making in the pelletizing line management system, "said Vladimir Lyushenko, head of the digital technologies department at the Severstal Iron Ore Business System Development Center.

The pinning process is very complex and many factors must be considered to create an adaptive control model. The solution has become a special symbiosis of physical modeling, machine learning algorithms and computer vision for the combine. When training the model, special regularizers were used, which helped in noisy data to identify the correct physical dependencies. In addition, it is continuously being refined and adjusted online based on real-time data, "said Severstal Digital Director Svetlana Potapova, head of Severstal's Artificial Intelligence cluster.