One of the problems of the use of drum pelletizers in metallurgy is the lining wear, as well as the economic costs associated with it, including increased energy costs during operation and the need to periodically stop the units and then replace the lining. Most significantly the trajectory of particle motion affects the lining wear profile and wear intensity. It is assumed that during the implementation of the technological process, a monodisperse occurs, which has the greatest effect on the wear profile. In addition, the lining wear is influenced by the impact of particles at an acute angle, with a maximum impact caused by a collision at an angle of 39°18′. At present there are no universal solutions for determining the degree of lining wear in real time with a corresponding adjustment of the pelletizing process parameters. Creating a system for monitoring the lining wear is necessary for timely repair and maintenance of equipment to prevent an emergency situation, as well as increase the service life of the aggregates. This article proposes a concept of a method that allows to evaluate the trajectory of the charge during the technological process according to the coordinates of the movement and acceleration of the probe in the unit during the implementation of the technological process. The digitization and analysis of data obtained from the probe will allow to assess the integrity of the lining surface, the degree of lining wear and places with increased wear rate in real time with the possibility of adjusting technological processes to increase the lining service life. The obtained data will allow to clarify the computer model of the process by assessing the behavior of the charge in the unit and create a reserve for the further implementation of digital twin equipment.

The paper presents a methodology for training neural networks for vision tasks on synthe-sized data on the example of steel defect recognition in automated production control systems. The article describes the process of dataset procedural generation of steel slab defects with a symmetrical distribution. The results of training two neural networks Unet and Xception on a generated data grid and testing them on real data are presented. The performance of these neural networks was assessed using real data from the Severstal: Steel Defect Detection set. In both cases, the neural networks showed good results in the classification and segmentation of surface defects of steel workpieces in the image. Dice score on synthetic data reaches 0.62, and accuracy—0.81.

A lot of researchers are developing new DEM parameters calibration approaches based on an experiment plan or the use of learning algorithms. This research is aimed at improving iterative algorithms frequently used for calibration. The big time consumption as a main problem of iterative algorithms is questioned. It is proposed to use Random forest algorithm to determine DEM parameters impact on the measured bulk responses. Measured responses are the parameters obtained by image processing using a technical vision system. As a result of 200 experiments processing, DEM parameters impact values on each bulk response were generated and presented as histograms. Obtained results were interpreted on the basis of the bulk material behavior and its physical properties. There is a discussion on the possibility of developing a universal DEM parameters calibration method based on the iterative algorithm.