In the modern world, when technology is developing at an incredible rate, computers have gained the ability to "see" and perceive the world around them like a human. This has led to a revolution in visual data analysis and processing. One of the key achievements was the use of computer vision to search for objects in photographs and videos. Thanks to these technologies, it is possible not only to find objects such as people, cars or animals, but also to accurately indicate their position using bounding boxes or masks for segmentation. This article discusses in detail modern models of deep neural networks used to detect humans in images and videos taken from a height and a long distance against a complex background. The architectures of the Faster Region-based Convolutional Neural Network (Faster R-CNN), Mask Region-based Convolutional Neural Network (Mask R-CNN), Single Shot Detector (SSD) and You Only Look Once (YOLO) are analyzed, their accuracy, speed and ability to effectively detect objects in conditions of a heterogeneous background are compared. Special attention is paid to studying the features of each model in specific practical situations, where both high-quality target object detection and image processing speed are important.

Keywords: machine learning, artificial intelligence, deep learning, convolutional neural networks, human detection, computer vision, object detection, image processing

This article presents the development of a combined method for summarizing Russian-language texts, integrating extractive and abstractive approaches to overcome the limitations of existing methods. The proposed method is preceded by the following stages: text preprocessing, comprehensive linguistic analysis using RuBERT, and semantic similarity-based clustering. The method involves extractive summarization via the TextRank algorithm and abstractive refinement using the RuT5 neural network model. Experiments conducted on the Gazeta.Ru news corpus confirmed the method's superiority in terms of precision, recall, F-score, and ROUGE metrics. The results demonstrated the superiority of the combined approach over purely extractive methods (such as TF-IDF and statistical methods) and abstractive methods (such as RuT5 and mBART).

Keywords: combined method, summarization, Russian-language texts, TextRank, RuT5

The paper considers a stochastic model of the operation of an automatic information processing system, which is described by a system of differential equations of the Kolmogorov distribution of state probabilities, assuming that the flow of requests is Poisson, including the simplest one. A scheme for solving a system of differential equations of high dimensionality with slowly changing initial data is proposed, and the parameters of the presented model are compared with the parameters of the simulation model of the Apache HTTP Server. To compare the simulation and stochastic models, a test server was used to generate requests and simulate their processing using the Apache JMeter program, which was used to estimate the parameters of the incoming and processed request flows. The presented model does not contradict the simulation model and allows us to evaluate the system's states under different operating conditions and calculate the load on the web server when there is a large amount of data.

Keywords: stochastic modeling, simulation model, Kolmogorov equations, sweep method, queuing system, performance characteristics, test server, request flow, service channels, queue

The paper considers the effect of particle size on the dynamics of suspended sediments in a riverbed. The EcoGIS-Simulation computing complex is used to simulate the joint dynamics of surface waters and sediments in the Volga River model below the Volga hydroelectric dam. The most important factor in the variability of the riverbed is the spring releases of water from the Volgograd reservoir, when water consumption increases fivefold. Some integral and local characteristics of the riverbed are calculated depending on the particle size coefficient.

Keywords: suspended sediment, soil particle size, sediment dynamics, diffusion, bottom sediments, channel morphology, relief, particle gravitational settling velocity, EcoGIS-Simulation software and hardware complex, Wexler formula, water flow

The article examines the influence of the data processing direction on the results of the discrete cosine transform (DCT). Based on the theory of groups, the symmetries of the basic functions of the DCT are considered, and the changes that occur when the direction of signal processing is changed are analyzed. It is shown that the antisymmetric components of the basis change sign in the reverse order of counts, while the symmetric ones remain unchanged. Modified expressions for block PREP are proposed, taking into account the change in the processing direction. The invariance of the frequency composition of the transform to the data processing direction has been experimentally confirmed. The results demonstrate the possibility of applying the proposed approach to the analysis of arbitrary signals, including image processing and data compression.

Keywords: discrete transforms, basic functions, invariance, symmetry, processing direction, matrix representation, correlation

Modern engineering equipment operation necessitates solving optimal control problems based on measurement data from numerous physical and technological process parameters. The analysis of multidimensional data arrays for their approximation with analytical dependencies represents both current and practically significant challenges. Existing software solutions demonstrate limitations when working with multidimensional data or provide only fixed sets of basis functions. Objectives. The aim of this study is to develop software for multidimensional regression based on the least squares method and a library of constructible basis functions, enabling users to create and utilize diverse basis functions for approximating multidimensional data. Methods. The development employs a generalized least squares method model with loss function minimization in the form of a multidimensional elliptical paraboloid. LASSO (L1), ridge regression (L2), and Elastic Net regularization mechanisms enhance model generalization and numerical stability. A precomputation strategy reduces asymptotic complexity from O(b²·N·f·log₂(p)) to O(b·N·(b+f·log₂(p))). The software architecture includes recursive algorithms for basis function generation, WebAssembly for computationally intensive operations, and modern web technologies including Vue3, TypeScript, and visualization libraries. Results. The developed web application provides efficient approximation of multidimensional data with 2D and 3D visualization capabilities. Quality assessment employs MSE, R², and AIC metrics. The software supports XLSX data loading and intuitive basis function construction through a user-friendly interface. Conclusion. The practical value lies in creating a publicly accessible tool at https://datapprox.com for analyzing and modeling complex multidimensional dependencies without requiring additional software installation.

Keywords: approximation, least squares method, basic functions, multidimensional regression, L1/L2 regularization, web-based

The study addresses the problem of short-term forecasting of ice temperature in engineering systems with high sensitivity to thermal loads. A transformer-based architecture is proposed, enhanced with a physics-informed loss function derived from the heat balance equation. This approach accounts for the inertial properties of the system and aligns the predicted temperature dynamics with the supplied power and external conditions. The model is tested on data from an ice rink, sampled at one-minute intervals. A comparative analysis is conducted against baseline architectures including LSTM, GRU, and Transformer using MSE, MAE, and MAPE metrics. The results demonstrate a significant improvement in accuracy during transitional regimes, as well as robustness to sharp temperature fluctuations—particularly following ice resurfacing. The proposed method can be integrated into intelligent control loops for engineering systems, providing not only high predictive accuracy but also physical interpretability. The study confirms the effectiveness of incorporating physical knowledge into neural forecasting models.

Keywords: short-term forecasting, time series analysis, transformer architecture, machine learning, physics-informed modeling, predictive control

This paper is devoted to the theoretical analysis of the methods used in verifying the dynamics of a signature obtained from a graphic tablet. A classification of three fundamental approaches to solving this problem is carried out: matching with a standard; stochastic modeling and discriminative classification. Each approach in this paper is considered using a specific method as an example: dynamic transformation of the time scale; hidden Markov models; support vector machine. For each method, the theoretical foundations are disclosed, the mathematical apparatus is presented, the main advantages and disadvantages are identified. The results of the comparative analysis can be used as the necessary theoretical basis for developing modern signature dynamics verification systems.

Keywords: verification, biometric authentication, signature dynamics, graphic tablet, classification of methods, matching with a standard, stochastic modeling, discriminative classification, hidden Markov models, dynamic transformation of the time scale

The characteristics of a submersible induction motor are described with sufficient reliability for practice by the theory of multi-motor electric drive. In this case, the classical circuit of a submersible induction motor is a coupled system of several equivalent-T circuits. In turn, this significantly increases its computational complexity and reduces the speed of ACS. It is proposed to construct a mathematical model of the submersible electric motor in the form of polynomials with significantly higher speed using the methods of experiment planning. In the area of applicability, the differences in the estimation of energy performance do not exceed 1.1%, between the proposed models and classical equivalent-T circuits.

Keywords: automated control system, mathematical model, polynomial, mean absolute percentage error, computational complexity, design of experiment, scatter diagram, modal interval, submersible electrical motor, rotor package

The article is devoted to the study of the possibilities of automatic transcription and analysis of audio recordings of telephone conversations of sales department employees with clients. The relevance of the study is associated with the growth of the volume of voice data and the need for their rapid processing in organizations whose activities are closely related to the sale of their products or services to clients. Automatic processing of audio recordings will allow checking the quality of work of call center employees, identifying violations in the scripts of conversations with clients. The proposed software solution is based on the use of the Whisper model for speech recognition, the pyannote.audio library for speaker diarization, and the RapidFuzz library for organizing fuzzy search when analyzing strings. In the course of an experimental study conducted on the basis of the developed software solution, it was confirmed that the use of modern language models and algorithms allows achieving a high degree of automation of audio recordings processing and can be used as a preliminary control tool without the participation of a specialist. The results confirm the practical applicability of the approach used by the authors for solving quality control problems in sales departments or call centers.

Keywords: call center, audio file, speech recognition, transcription, speaker diarization, replica classification, audio recording processing, Whisper, pyannote.audio, RapidFuzz

The article presents a mathematical model that formalizes the process of managing the scientific activities of an organization. The model based on the theory of queuing. The principle of death - reproduction used in the construction. For a special case, a graph of states and a system of Kolmogorov differential equations are given. The intensity of the input and output streams are time-dependent non-stationary streams. The model allows us to consider various structures and schemes of interaction between scientific departments and various sce-narios for setting scientific tasks and the intensity of their solution by employees of the organization. A software package for decision-making has developed for the model for optimal management of the scientific activities of the department. The article presents one of the results of an experimental and model study of the influence of the motivational component and the level of competence of employees. Graphs of the system states given for the resulting solution. The research can used for comprehensive evaluation of results, planning, resource allocation and management of scientific activities.

Keywords: diversification of management, production diversification, financial and economic purposes of a diversification, technological purposes of ensuring flexibility of production

The article discusses the structure and principle of operation of an improved centrifugal unit for mixing bulk materials. A special feature of which is the ability to control mixing modes. Due to its design, the selection of a rational position of the bump makes it possible to provide such conditions for the impact interaction of particle flows, in which a high-quality homogeneous mixture of components is formed, the particles of which have different sizes, shapes and other parameters. To characterize the resulting mixture, the coefficient of heterogeneity was used, the conclusion of which is based on a probabilistic approach. A computational scheme of the rarefied flow formation process is given. An expression is derived for calculating the coefficient of heterogeneity when mixing bulk media, the particles of which have different sizes, shapes and other parameters. The research conducted in the article allows not only to predict the quality of the resulting mixture, but also to identify the factors that have the greatest impact on achieving the required uniformity.

Keywords: aggregate, bulk media, mixing, coefficient of heterogeneity, concentration, design scheme, particle size

The article presents a novel approach for adaptive control of genetic algorithm parameters using reinforcement learning methods. The use of the Q-learning algorithm enables dynamic adjustment of mutation and crossover probabilities based on the current population state and the evolutionary process progress. Experimental results demonstrate that this method offers a more efficient solution for optimization problems compared to the classical genetic algorithm and previously developed approaches employing artificial neural networks. Tests conducted on the Rastrigin and Shaffer functions confirm the advantages of the new method in problems characterized by a large number of local extrema and high dimensionality. The article details the theoretical foundations, describes the implementation of the proposed hybrid model, and thoroughly analyzes experimental results. Conclusions highlight the method's adaptability, efficiency, and potential for application in complex optimization scenarios.

Keywords: genetic algorithm, reinforcement learning, adaptive control, Q-learning, global optimization, Rastrigin function, Shaffer function

The article discusses a software module developed by the authors for automatic generation of program code based on UML diagrams. The relevance of developing this module is due to the limitations of existing foreign code generation tools related to functionality, ease of use, support for modern technologies, as well as their unavailability in Russian Federation. The module analyzes JSON files obtained by exporting UML diagrams from the draw.io online service and converts them into code in a selected programming language (Python, C++, Java) or DDL scripts for DBMS (PostgreSQL, Oracle, MySQL). The Python language and the Jinja2 template engine were used as the main development tools. The operation of the software module is demonstrated using the example of a small project "Library Management System". During the study, a series of tests were conducted on automatic code generation based on the architectures of software information systems developed by students of the Software Engineering bachelor's degree program in the discipline "Design and Architecture of Software Systems". The test results showed that the code generated using the developed module fully complies with the original UML diagrams, including the structure of classes, relationships between them, as well as the configuration of the database and infrastructure (Docker Compose). The practical significance of the investigation is that the proposed concept of generating program code based on visual models of UML diagrams built in the popular online editor draw.io significantly simplifies the development of software information systems, and can be used for educational purposes.

Keywords: code generation, automation, python, jinja2, uml diagram, json, template engine, parsing, class diagram, database, deployment diagram

The article studies possibilities for analyzing geopolitical processes within the framework of situational analysis methodology using cognitive modeling. Situational analysis description is given, and scenario for developing events is presented where two stages are distinguished: a preparatory stage (a pre-scenario stage) which is essential for performing descriptive and explanatory functions of predictive research, and a scenario stage intended for substantive and formal research as well as for description of predicted processes, construction of system models and preparation of all significant information for scenario synthesis. Furthermore, a method for applying situational analysis is proposed to be used within the framework of the cognitive modeling toolkit of a “future scenario” option and its analysis with account of new “main” factors, relationships, feedbacks and dynamics of their alterations. When forming a scenario for a specific geopolitical situation within the framework of cognitive modeling, this method can be presented by causal (functional) and logical-semantic relation between the elements/agents of actions and counteractions. By interpreting the logical-semantic as structural, and the causal as dynamic, we obtain a structural-dynamic systemic description of geopolitical confrontation using the language of cognitive graphs, i.e. presenting a graphical expression of causal relationships between the concepts (factors) that characterize a particular geopolitical process. Thus, within the framework of a scenario stage the following procedures are conducted: analyzing the initial geopolitical situation, namely: determining key factors that build up the scheme of internal connections and external relationships, and their structuring; defining factors that make an impact; determining impact directions and force (positive and negative effect); choosing basic stereotypes or generalized models of interactions that correspond to the initial situation; constructing cognitive models of the current state of a situation; studying trends for the situation’s development and its dynamics analysis; transferring a scenario onto a practical basis.

Keywords: geopolitical processes, situational analysis, cognitive modeling, methodology for constructing predictive scenarios

Traditional marketing methods of promoting goods and services are aimed at a wide audience and do not take into account the individual characteristics of consumers, which can lead to a small percentage of positive responses and even to negative responses (loss of customers). Wide audience coverage leads to an increase in the cost of marketing interactions and does not guarantee the achievement of the goals of marketing campaigns. In such a situation, the task is to minimize excess costs through a more rational organization of marketing interactions aimed at obtaining maximum profit from each target client. To implement such a strategy, tools are needed that can identify customer segments, marketing interaction with which will lead to a positive response. One of the technologies for building such tools is uplift modeling, which is a section of machine learning and is considered a promising direction in data-driven marketing. In this article, based on the open data X5 RetailHero Uplift Modeling Dataset, provided by X5 Retail Group, a comparative analysis of the effectiveness of various uplift modeling approaches is conducted to identify the segment of customers who are most susceptible to target impact. Various uplift metrics and visual technologies are used to conduct the comparative analysis.

Keywords: effective marketing communications with customers, customer segmentation, machine learning methods, uplift modeling, uplift quality metrics

Changes in external conditions, parameters of object functioning, relationships between system elements and system connections with the supersystem lead to a decrease in the accuracy of the artificial intelligence models results, which is called model degradation. Reducing the risk of model degradation is relevant for electric power engineering tasks, the peculiarity of which is multifactor dependencies in complex technical systems and the influence of meteorological parameters. Therefore, automatic updating of models over time is a necessary condition for building user confidence in forecasting systems in power engineering tasks and industry implementations of such systems. There are various methods used to prevent degradation, including an algorithm for detecting data drift, an algorithm for updating models, their retraining, additional training, and fine-tuning. This article presents the results of a study of drift types, their systematization and classification by various features. The solution options that developers need to make when creating intelligent forecasting systems to determine a strategy for updating forecast models are formalized, including update trigger criteria, model selection, hyperparameter optimization, and the choice of an update method and data set formation. An algorithm for forming a strategy for automatic updating of artificial intelligence models is proposed and practical recommendations are given for developers of models in problems of forecasting time series in the power industry, such as forecasting electricity consumption, forecasting the output of solar, wind and hydroelectric power plants.

Keywords: time series forecasting, artificial intelligence, machine learning, trusted AI system, model degradation, data drift, concept drift

This paper provides a survey of metrics used to assess the quality of images generated by generative models. Specialized metrics are required to objectively evaluate image quality. A comparative analysis showed that a combination of different metrics is necessary for a comprehensive evaluation of generation quality. Perceptual metrics are effective for assessing image quality from the perspective of machine systems, while metrics evaluating structure and details are useful for analyzing human perception. Text-based metrics allow for the assessment of image-text alignment but cannot replace metrics focused on visual or structural evaluation. The results of this study will be beneficial for specialists in machine learning and computer vision, as well as contribute to the improvement of generative algorithms and the expansion of diffusion model applications.

Keywords: deep learning, metric, generative model, image quality, image

The article is devoted to the application of modern methods of generative image compression using variational autoencoders and neural network architectures. Special attention is paid to the analysis of existing approaches to image generation and restoration, as well as a comparative assessment of compression quality in terms of visual perception and metric indicators. The aim of the study is to systematize deep image compression methods and identify the most effective solutions based on the variational Bayesian approach. The paper considers various architectures, including conditional autoencoders and hypernetwork models, as well as methods for evaluating the quality of the data obtained. The main research methods used were the analysis of scientific literature, a comparative experiment on the architectures of generative models and a computational estimation of compression based on metrics. The results of the study showed that the use of variational autoencoders in combination with recurrent and convolutional layers makes it possible to achieve high-quality image recovery with a significant reduction in data volume. The conclusion is made about the prospects of using conditional variational autoencoders in image compression tasks, especially in the presence of additional information (for example, metadata). The presented approaches can be useful for developing efficient systems for storing and transmitting visual data.

Keywords: variational autoencoders, generative models, image compression, deep learning, neural network architectures, data recovery, conditional models

The article focuses on developing data clustering algorithms using asymmetric similarity measures, which are relevant in tasks involving directed interactions. Two algorithms are proposed: stepwise cluster formation and a modified version with iterative center refinement. Experiments were conducted, including a comparison with the k-medoids method. The results showed that the fixed-center algorithm is efficient for small datasets, while the center-recalculation algorithm provides more accurate clustering. The choice of algorithm depends on the requirements for speed and quality.

Keywords: clustering, asymmetric similarity measures, clustering algorithms, iterative refinement, k-medoids, directed interactions, adaptive methods

This paper considers the modeling of user work with a multi-server database developed on the basis of microservice architecture. The subject area was analyzed, the main entities of the system were described, and the mechanisms of data transfer and service interaction using Docker and Apache Kafka were implemented. It was revealed that the development of a multi-server database allowed to achieve high scalability and fault tolerance of the system. The implementation of replication and sharding mechanisms provided even load distribution, and the use of Kafka message broker facilitated efficient data exchange between services. The testing confirmed the system's reliability under high load, as well as revealed its strengths and potential improvements.

Keywords: modeling, load balancing, Docker, Apache Kafka, microservice architecture, distributed systems, query optimization

The article is devoted to the study of the problem of estimating unknown parameters of linear regression models using the least absolute deviations method. Two well-known approaches to identifying regression models are considered: the first is based on solving a linear programming problem; the second, known as the iterative least-squares method, allows one to obtain an approximate solution to the problem. To test this method, a special program was developed using the Gretl software package. A dataset of house prices and factors influencing them, consisting of 20640 observations, was used for computational experiments. The best results were obtained using the quantreg function built into Gretl, which implements the Frisch-Newton algorithm; the second result was obtained using an iterative method; and the third result was achieved by solving a linear program using the LPSolve software package.

Keywords: regression analysis, least absolute deviations method, linear programming, iterative least squares method, variational weighted quadratic approximation method

The use of machine learning when working with text documents significantly increases the efficiency of work and expands the range of tasks to be solved. The paper provides an analysis of the main methods of presenting data in a digital format and machine learning algorithms, and a conclusion is made about the optimal solution for generative and discriminative tasks.

Keywords: machine learning, natural language processing, transformer architecture models, gradient boosting, large language models

This article presents a comprehensive analysis of Russian-language texts utilizing neural network models based on the Bidirectional Encoder Representations from Transformers (BERT) architecture. The study employs specialized models for the Russian language: RuBERT-tiny, RuBERT-tiny2, and RuBERT-base-cased. The proposed methodology encompasses morphological, syntactic, and semantic levels of analysis, integrating lemmatization, part-of-speech tagging, morphological feature identification, syntactic dependency parsing, semantic role labeling, and relation extraction. The application of BERT-family models achieves accuracy rates exceeding 98% for lemmatization, 97% for part-of-speech tagging and morphological feature identification, 96% for syntactic parsing, and 94% for semantic analysis. The method is suitable for tasks requiring deep text comprehension and can be optimized for processing large corpora.

Keywords: BERT, Russian-language texts, morphological analysis, syntactic analysis, semantic analysis, lemmatization, RuBERT, natural language processing, NLP

The article presents an analysis of modern methods of image generation: variational autoencoders (VAE), generative adversarial networks (GAN) and diffusion models. The main attention is paid to a comparative analysis of their performance, generation quality and computational requirements. The Frechet Inception Distance (FID) metric is used to assess the image quality. Diffusion models showed the best results (FID 20.8), outperforming VAE (FID 59.75) and GAN (FID 38.9), but require significant resources. VAEs are stable, but generate blurry images. GANs provide high quality, but suffer from training instability and mode collapse. Diffusion models, due to step-by-step noise decoding, combine detail and structure, which makes them the most promising. Also considered are methods of image-to-image generation used for image modification. The results of the study are useful for specialists in the field of machine learning and computer vision, contributing to the improvement of algorithms and expansion of the areas of application of generative models.

Keywords: deepfake, deep learning, artificial intelligence, GAN, VAE, diffusion model