Работая с сайтом, я даю свое согласие на использование файлов cookie. Это необходимо для нормального функционирования сайта, показа целевой рекламы и анализа трафика. Статистика использования сайта обрабатывается системой Яндекс.Метрика

Принять

Научный журнал Моделирование, оптимизация и информационные технологииThe scientific journal Modeling, Optimization and Information Technology

Online media
issn 2310-6018

submit
a manuscript

metadata of articles for the last 2 years

A method for automatic summarization of scientific texts based on diagnostically driven iterative generation

2026. T.14. № 9. id 2447

Gokarev V.N.

DOI: 10.26102/2310-6018/2026.60.9.006

This paper examines the problem of automatic abstract summarization of Russian-language scientific texts using local large-scale language models. It is shown that single-pass abstract generation by a language model without taking into account the characteristics of the source document leads to systematic errors: semantic incompleteness, mechanical copying, and information redundancy or insufficiency. To identify and eliminate these deficiencies, a hybrid software system is proposed that implements a closed-loop control system for the abstract generation process with feedback from a three-factor diagnostic quality assessment module. The system iteratively adjusts query parameters to the language model, such as temperature, the number of submitted keywords, the format of the system query instruction, and the minimum and maximum number of generated tokens, based on the classification of the generated abstract diagnostic profile. The diagnostic methodology proposed includes an asymmetric penalty transformation of z-normalized deviations, reflecting the substantive difference between redundancies and insufficiencies in the lexical, semantic, and compressional proximity of an abstract to the reference distribution. The system's modular architecture is described, including modules for iterative abstract generation, diagnostic evaluation, and an automatic iteration process stop block. The results of experimental testing on a corpus of Russian-language scientific articles are presented, demonstrating a statistically significant and sustainable increase in quality for the combined diagnostic metric relative to a single-pass baseline on a delayed test sample.

Keywords: automatic summarization, scientific texts, large language models, control loop, diagnostic quality evaluation, asymmetric penalty functions

Two-moment evaluation of joint execution of independent program sequences

2026. T.14. № 7. id 2433

Buevich E.A.

DOI: 10.26102/2310-6018/2026.58.7.010

Compiling high-level language source code into machine code is an NP-hard problem. One of the problems arising during compilation, which has a huge solution space, is the presence of data-independent sequences of operations in the code. This paper considers an algorithm for finding an optimal uniform mixing of arbitrary instruction sequences, which can be represented as an acyclic connected graph. The purpose of such mixing is to increase the instruction-level parallelism. To this end, a sequence is described as a vector of random variables, each of which characterizes its demand for a specific resource. This representation allows for an approximate estimation of the expected value of delays arising during the execution of their mixture due to competition among several sequences for resources, without actually compiling the code. The algorithm operates on scalar data describing sequences and their combinations, instead of long instruction vectors. The search for the optimal partitioning of sequences into groups is performed in O(nk) time, where n is the number of sequences, and k is some limiting value of the number of sequences in the mixture, depending on the code, instruction latency, and the number of registers. For computing cores of modern processors, in most practical problems k does not exceed 4.

Keywords: instruction-level parallelism, acyclic dependence graph, reiser-Lavenberg algorithm, closed queueing network, code mixing, compiler optimization

Development and software implementation of an intelligent decision support system for choosing mechanical ventilation modes

2026. T.14. № 7. id 2428

Frolov S.V. Sudakov D.E. Dolgov E.P.

DOI: 10.26102/2310-6018/2026.58.7.011

The relevance of this study is due to the high cognitive load on the intensivist when choosing a mechanical ventilation mode because of dozens of modes and more than 50 adjustable parameters. The lack of formalized criteria and transparent decision support systems leads to variability in therapy and an increased risk of ventilator-associated complications. The aim of this work is to develop a methodology, algorithm and software prototype of an intelligent clinical decision support system for choosing mechanical ventilation modes, providing explainability of recommendations. The methods used include formalization of clinical parameters, construction of a state classifier, creation of production rules, object-oriented programming in Python, and testing on literature-based clinical scenarios. Four groups of parameters (respiratory mechanics, gas exchange, patient activity, hemodynamics) are formalized. A state classifier (apnea, respiratory muscle weakness, obstructive and restrictive disorders) is built. A knowledge base of 30 "IF–THEN" production rules grouped into six clinical situations with priority selection is created. An algorithm with an explanation module that generates a textual justification for each recommendation is developed. The software prototype is implemented in Python (tkinter, experta). Testing on 20 literature‑based clinical scenarios showed complete agreement of recommendations with reference protocols in 85 % of cases, partial agreement in 10 %, and disagreement (a scenario requiring extracorporeal CO₂ removal) in 5 %. The average recommendation generation time was 0.42 seconds. The obtained results confirm the feasibility of the proposed methodology and allow the system to be used as a standalone physician’s workstation or as an embeddable module in ventilators.

Keywords: mechanical ventilation, decision support system, production rules, state classification, explanation module

An algorithm for determining the optimal route of the robot for monitoring objects, taking into account obstacles

2026. T.14. № 8. id 2427

Makarovskikh T.A. Sallam M.E.

DOI: 10.26102/2310-6018/2026.59.8.004

One of the urgent tasks is ground-based monitoring of the condition of various facilities. It may be necessary not only in industrial enterprises, but also in agriculture (inspection of orchards), medicine (monitoring the condition of patients in infectious diseases departments), and emergency situations. In these situations, it is not possible to place sensors and smart cameras in fixed locations. In addition, there are situations when monitoring goals are set immediately before it is carried out. Another urgent task is to build a robot's path, taking into account the presence of obstacles. This article discusses an algorithm for constructing the shortest route for monitoring at a finite set of points by a ground robot, taking into account obstacles. The task of determining the trajectory of a robot is formulated as a vehicle routing task. Data on the coordinates of the robot and the location of monitoring points is stored as a weighted planar graph using a list of edges. The article proposes the algorithm SearchWithObstacles, using function SPPA (Shortest Path with Precautionary Avoidance), which is based on the nearest neighbor method and solves the problem in polynomial time. The algorithm is compared with known approaches and the advantages and disadvantages of the proposed algorithm are discussed.

Keywords: vehicle routing, planar graph, heuristic algorithm, optimization, obstacle avoidance, traveling salesman problem

A method for transitioning to a fixed-length parameter vector of the objective function in optimization problems of overhead crane operations in a slab yard

2026. T.14. № 8. id 2423

Shchegolikhin I.S. Andreev S.M.

DOI: 10.26102/2310-6018/2026.59.8.001

A number of optimization problems that arise when rationalizing the warehouse logistics of a continuously cast slab yard of a metallurgical enterprise operate with a variable-length vector of parameters of the objective function. This circumstance, consequently, restricts the choice of methods applicable to such problems. Moreover, an increase in the parameter vector of the objective function in many cases raises the resource intensity of the optimization algorithm. Several existing methods that reduce the parameter vector of the objective function to a constant size and that find use in combinatorial optimization suffer from a pronounced limitation: they apply only to a highly specialized class of problems. The present study aims to reduce the parameter vector of the objective function that appears in problems of optimizing the work of bridge cranes in a slab yard to a constant size. Such a reduction will allow one to use, for the optimization problem under consideration, all known optimization methods from the appropriate class and, in addition, will lower the resource intensity of the optimization algorithm by decreasing the number of parameters of the objective function. In this work, the authors propose a method that generalizes L‑systems and cellular coding and makes it possible, for a given objective function, to move from a variable‑size parameter vector to a constant‑size parameter vector. The novelty of the work consists in extending ideas that are specific to highly specialized fields of scientific knowledge to a set of classes of optimization problems, in particular to scheduling problems for bridge cranes. The authors examine an objective function that arises when one jointly solves bridge crane scheduling problems and the problem of selecting slabs for a hot rolling schedule. They apply the proposed method to this objective function and describe the functions that change the search space in the form of simple algorithms. As a result, the dimensionality of the objective function vectors decreases at least by a factor of 10. The proposed method can be used in the design of a digital twin for a continuous casting billet storage yard, as well as for managing a hierarchy of commands: from the simplest (for controlling individual crane components) to the most complex (defining logistics operations involving slabs).

Keywords: parameter vector, objective function, slabs, cellular encoding, indirect encoding, direct encoding

Assessment of cancer prevalence based on analysis of statistics from the Municipal Indicator database created by the Rosstat

2026. T.14. № 8. id 2420

Stepanov V.S.

DOI: 10.26102/2310-6018/2026.59.8.006

Cancer diseases are a major health concern in Russia and many other countries. A national strategy for cancer control has been adopted in the Russian Federation, with primary prevention being one of its key components. However, studies in this area, including municipal-level modeling, are scarce. This work is useful for identifying the relationship between cancer prevalence in a territory and a set of contributing factors. The aim of the study is to develop a predictive regression model. The empirical basis for the model is a sample of 56 observations representing municipalities of the Altai Territory, most of which are rural areas. The results are practically applicable, as they rely on open official data, and the model's accuracy is sufficiently high. The dependent variable is the number of cancer patients registered at the end of year t. The observations include five factors introduced with different time lags; three of them underwent nonlinear transformation. Additionally, two qualitative factors characterizing territorial features were included, with their categories encoded by three dummy variables. A linear regression model was estimated using ordinary least squares. The resulting model achieved a mean absolute percentage error (MAPE) of less than 1%. In external validation on other municipalities, the MAPE ranged from 1.5 % to 2.4 %. The model enables one-year-ahead forecasts of cancer prevalence for a large number of municipalities in Russia. Forecasting examples are illustrated with graphs for seven cities in the Volga region and two territories in the Krasnoyarsk region over several years. Such forecasts are valuable for designing health, social, and environmental policies at the municipal level.

Keywords: variable-structure regression model, malignant neoplasms, technogenic pollution, per capita alcohol consumption, physician supply, drinking water quality

Spatial-temporal fusion of depth data for estimating three-dimensional distances

2026. T.14. № 8. id 2414

Yu X.

DOI: 10.26102/2310-6018/2026.59.8.013

Depth enhancement is commonly evaluated with pixel-wise errors, although robotic vision and three-dimensional metrology ultimately use geometric distances. This study develops and tests a protocol that evaluates enhancement methods by the error of Euclidean distances between back-projected point pairs. The primary experiments use the TUM Freiburg1 desk and xyz sequences. Additional validation includes the dynamic Freiburg3 walking_static sequence, controlled stress tests with moving depth boundaries, long range and elevated noise, three pseudo-reference variants, and a computational-cost benchmark. With temporal half-window N = 10, temporal mean fusion reduces distance RMSE by 24.07 % on desk and 23.66 % on xyz relative to raw depth. Direct metric comparison shows that temporal median has the lowest pixel MAE, whereas temporal mean has the lowest distance RMSE; Spearman rank correlations are 0.70 and 0.90. Temporal mean remains ranked first under all three pseudo-reference constructions. On the dynamic sequence, a temporal-stability gate retains 18.92 % of pixels and changes the method ranking, demonstrating that moving regions must be excluded or motion-compensated. For 640 × 480 frames at N = 10, the research CPU implementation requires 56.2 ms for temporal mean, 217.5 ms for temporal median, and 3155.6 ms for spatial-temporal median. The results delimit the practical use of the protocol and the systematic limitations of pseudo-reference evaluation.

Keywords: depth data, three-dimensional measurement, depth map, pseudo-reference, spatial-temporal fusion, geometric accuracy, depth enhancement

Learning analytics: predicting student performance based on student activity and performance data in e-learning course management systems

2026. T.14. № 9. id 2412

Danichev A.A. Liksonova D.I. Shestakov V.N.

DOI: 10.26102/2310-6018/2026.60.9.018

This study examines the problem of increasing the accuracy of predicting the learning outcomes of university students based on data on current academic performance in disciplines from electronic educational courses. The source data are numerical values – grades and activity indicators based on the results of each week of training – which may relate to different scale ranges and be highly sparse, making their subsequent use in forecasting algorithms difficult. Therefore, it is necessary to conduct high-quality preprocessing of the source information to construct a predictive mathematical model of learning outcomes. One proposed method for data processing is the use of percentiles, which effectively describe the distribution of features. The percentile method solves the issue of fairly comparing students enrolled in courses with differing difficulty levels by turning raw scores into relative rankings within a learner's peer group. This study presents a side-by-side comparison of two data processing techniques: the conventional centering-and-norming method versus the percentile-based strategy. A nonparametric estimator designed for the Nadaraya-Watson regression function is then employed to build the predictive model. Excerpts from computational experiments with varying numbers of groups are presented, yielding acceptable results from a practical perspective.

Keywords: educational data, learning analytics, student performance prediction, educational data mining, machine learning methods, data preprocessing, normalization, percentiles, electronic educational environment

Comparative modeling of computational load distribution heuristics in a hybrid cloud-fog architecture for IoT systems

2026. T.14. № 9. id 2405

Proshechkina N.V. Gasanov R.R. Zhityaikina V.D.

DOI: 10.26102/2310-6018/2026.60.9.007

This paper presents a comparative analysis of heuristics for workload distribution based on a hybrid cloud-fog architecture used for processing Internet of Things traffic. Such an architecture is optimal for applications requiring high performance and operating in real time, taking into account constraints on latency, power consumption and the load on computing nodes. Comparative modelling is carried out using a stochastic multi-criteria mathematical model of computational load distribution between fog and cloud resources. Constraints relating to capacity, latency, reliability and load balancing are applied to the model. Using iFogSim, a typical IoT scenario was modelled with variations in task flow intensity and a comparison of two distribution policies – a baseline and an optimised one. A dynamic analysis was carried out with the load varying according to the intensity of request arrivals. The results show that the optimised strategy reduces average latency, energy consumption, the proportion of SLA violations and the degree of fog node overload, ensuring more stable system operation as the load increases. With this approach, fog and cloud resources are utilised more efficiently, thereby improving the efficiency of IoT data processing. The practical significance of this work lies in the potential to apply the simulation results when developing intelligent load management algorithms for industrial IoT systems, smart cities, telemedicine and video surveillance systems.

Keywords: hybrid cloud-fog architecture, internet of Things, computational load distribution, multi criteria optimization, latency and energy consumption, iFogSim, SLA quality of service

An incremental method for updating a multidimensional cube based on a disordered stream of information system log events

2026. T.14. № 8. id 2404

Ufimtsev G.A. Zykov S.V.

DOI: 10.26102/2310-6018/2026.59.8.003

Information systems generate large volumes of event logs, which are used to analyze the operation of applications and services. In this case, events may arrive in the analytical circuit later than the moment of their actual occurrence and not in the original order. Such time inconsistency leads to errors in the construction of aggregated time-based indicators, while regular full recalculation of a multidimensional analytical cube requires significant computational costs. The purpose of the work is to develop an approach to updating a multidimensional cube based on the flow of log events, taking into account delays and violations of the order in which they arrive. The leading approach is to divide the final cube representation into a base cube and a compensation layer. Timely events update the base cube, while delayed events within a given compensation horizon add corrections to the compensation layer. Unique event identifiers are checked to exclude repeated event accounting. The paper presents a formal event model, describes rules for assigning events to cube cells by the occurrence time, proposes an incremental update algorithm, and provides a computational experiment on a software prototype. The experimental results show that the proposed method reduces aggregate errors compared with the window-based method and requires a smaller amount of repeated data processing compared with batch recalculation. The materials of the paper are of practical value for developing analytical pipelines for monitoring, audit, security event analysis, and user activity analysis.

Keywords: incremental updating, OLAP cube, event log, event flow, disordered data arrival, delayed events, compensation layer, stream data processing

Application of hybrid recommendation system model based on graph neural networks and feature description of users and communities

2026. T.14. № 7. id 2402

Азарнова Т.В. Mazur I.Y.

DOI: 10.26102/2310-6018/2026.58.7.008

In the conditions of growing competition in the social media market, there is an increasing need for personalized content recommendation tools that can increase user engagement. Traditional approaches to building recommendation systems based on collaborative filtering or content filtering have limitations in community recommendation tasks, where both the structure of social connections and the semantic description of groups are important. This article is aimed at studying a hybrid model of a recommendation system combining graph neural networks and feature description of users and communities. The model integrates information about user interactions with groups in the form of a graph, text embeddings of community descriptions obtained using transformer models, as well as numerical and categorical features of groups and users. The results of the study demonstrate a significant advantage of the proposed approach compared to baseline methods. A comparative analysis with popularity baseline and random baseline shows an improvement in the MAP@3 metric by 3.89 times relative to recommending popular groups and by 4045.9 times relative to random recommendation. The study is based on data from the OK social network, including information about 159,685 users and 54,425 communities.

Keywords: recommendation systems, graph neural networks, machine learning, personalization, social networks, collaborative filtering, embeddings

Geolocation determination from dashcam footage using ResNet

2026. T.14. № 7. id 2399

Baskhanov A.R.

DOI: 10.26102/2310-6018/2026.58.7.016

The relevance of this study is driven by the growing volume of street‑level imagery and the need for automatic georeferencing without relying on GPS metadata. Accordingly, this paper aims to assess the feasibility of determining photograph coordinates within a single city using deep neural networks. The leading research method is the application of convolutional neural networks ResNet18 and ResNet50 for latitude and longitude regression from 512×512-pixel frames. Data were collected using the open Mapillary platform, and a custom program with a 10,000‑cell grid was developed to ensure uniform coverage of Ufa. A total of 88,798 geotagged images were collected. The paper presents quantitative results of model training. The ResNet50‑based model achieves a median localization error of approximately 3.1 km after 13 epochs, with 92.6% of predictions falling within a 10 km radius of the ground truth. Smooth Grad‑CAM visualizations reveal that the model focuses on building facades, intersections, road markings, and other infrastructure elements – semantically meaningful landmarks for urban geolocation. The materials of the article are of practical value for automatic georeferencing systems for large collections of street‑level imagery and for the development of visual localization methods for autonomous vehicles. The work also provides a reproducible pipeline for city‑level image geolocation and analyses the behavioral attention of convolutional neural networks in this task.

Keywords: image geolocation, convolutional neural networks, resNet, urban scene analysis, deep regression, attention visualization, smooth Grad‑CAM, mapillary

Development of machine learning models for predicting gastrointestinal diseases

2026. T.14. № 9. id 2398

Almusawi M. Lyapuntsova E.V.

DOI: 10.26102/2310-6018/2026.60.9.014

The article addresses the problem of predicting pathological conditions of the gastrointestinal tract from endoscopic images using machine learning models. The relevance of the study is determined by the need to improve the reproducibility of preliminary assessment of endoscopic data and to reduce the risk of missing clinically significant images when processing large volumes of visual information. The aim of the study is to develop and experimentally evaluate a resource-efficient classification scheme for endoscopic images, taking into account not only the average recognition quality, but also the reliability of automatic decisions for pathological classes. The open Kvasir-v2 dataset containing 8000 images from eight classes was used as the experimental basis. A simple convolutional neural network, MobileNetV2, and EfficientNetB0 with transfer learning were compared. In addition, a selective scheme was proposed, combining the predictions of the two strongest models and sending uncertain images for manual review. The experiment showed that EfficientNetB0 achieved a macro-F-score of 0.8497 on the test set, while the combined prediction increased this value to 0.8681. The proposed selective scheme provided a macro-F-score of 0.9581 among automatically accepted decisions and reduced the number of dangerous automatic errors from 64 to 10. The obtained results demonstrate the potential of a clinically oriented approach in which automatic classification is supplemented by confidence control and assessment of the risk of missing a pathological condition.

Keywords: machine learning, endoscopic images, gastrointestinal tract, disease prediction, convolutional neural networks, transfer learning, selective classification, pathological conditions, medical decision support

Integrating information security and systems engineering into a multilingual sign language learning game

2026. T.14. № 8. id 2396

Ashrafi A. Philippovich Y.S. Mokhnachev V.S. Makhmud A.A.

DOI: 10.26102/2310-6018/2026.59.8.014

The development of accessible and secure digital tools is critical for inclusive education. While our prior work demonstrated the pedagogical efficacy of a multilingual sign language learning game-evidenced by a 35 % improvement in sign recognition and a System Usability Scale (SUS) score of 90/100-the transition to a production-ready, internet-connected platform introduces significant security and operational risks. To address this, this paper presents a unified framework that integrates four core disciplines-web application security analysis, software lifecycle automation (CI/CD), secure digital document workflows, and security information and event management (SIEM)-to systematically harden the platform and ensure its long-term resilience. We implement an integrated approach where continuous security testing, protected data workflows, and threat-informed development work in concert to ensure platform integrity. Results demonstrate the elimination of critical vulnerabilities, an 87.5 % reduction in high-severity issues within three months, and a 65.4 % improvement in mean time to remediate (MTTR). The secure certificate system achieved a 99.6 % issuance success rate and 100 % forgery detection, with public verification in under one second. This lab-validated methodology provides a practical blueprint for transforming innovative educational tools into resilient, production-ready systems.

Keywords: sign language, educational game, web application security, CI/CD automation, secure document flow, SIEM, data privacy, inclusive educational technologies

Methodology of system analysis of remote and hard-to-reach territories as complex systems

2026. T.14. № 8. id 2395

Serdechnyy D.V.

DOI: 10.26102/2310-6018/2026.59.8.005

This article examines the development management of remote and hard-to-reach territories of the Russian Federation, characterized by low population density, harsh climatic conditions, and high levels of infrastructure vulnerability. It substantiates the need to shift from sectoral and reactive management to proactive management based on systems analysis and predictive modeling. A formal model of a remote territory and its representation as a weighted directed graph have been developed, enabling a quantitative description of material, energy, and information flows. Integral indices of vulnerability and adaptive potential are proposed, as well as a matrix for classifying territories into nine resilience zones. A six-stage systems analysis algorithm has been developed that integrates remote sensing data collection, graph modeling, index calculation, scenario modeling, and the generation of management recommendations with feedback. The proposed methodology provides a foundation for building intelligent decision support systems for managing remote and hard-to-reach territories. Future research prospects lie in the development of hybrid modeling methods for creating full-scale digital twins of Arctic municipalities. Testing the methodology on pilot projects will refine the index weightings and create industry-specific databases for scaling up the proposed approach.

Keywords: systems analysis, predictive modeling, remote areas, infrastructure vulnerability, adaptive capacity, decision support system

Numerical modeling of one-dimensional wave processes in a bimodular medium: software implementation and convergence analysis

2026. T.14. № 7. id 2394

Polesia V.A. Dudko O.V. Lapteva A.A.

DOI: 10.26102/2310-6018/2026.58.7.015

Modern mechanics of deformable solids is actively advancing towards modeling bimodular materials – media with different resistance to tension and compression. This class includes both natural and artificial materials widely used in engineering practice. The difference in elastic moduli under tension and compression leads to a strong nonlinearity of the mechanical response, which under dynamic loading generates qualitatively complex wave patterns with the formation and interaction of strain discontinuities. As a result, the analytical solution of the corresponding boundary value problems becomes quite challenging due to numerous wavefront interactions. This paper presents the development, software implementation, and verification of a specialized system for the automated solution of boundary value problems of one-dimensional dynamics of bimodular elastic media under arbitrary boundary loading regimes. The system is based on a formalized iterative method for constructing the wave pattern, which uses piecewise-linear approximation of both the boundary conditions and the constitutive relations of the Myasnikov-Oleynikov bimodular medium. The software fully automates the construction of the wave pattern: it tracks the evolution of the wave system, classifies events, automatically generates and solves systems of nonlinear equations to determine the parameters of new local domains between wavefronts, and controls the evolutionarity of strain discontinuities and introduction of rigid layers. Verification was performed on test problems involving the formation of shock waves and rigid layers. A convergence analysis of the results against known analytical solutions confirms the high accuracy of the implemented method with an increasing number of boundary condition approximation nodes. Thus, the developed software tool is an effective instrument for automating the study of wave processes, providing physically correct modeling of the one-dimensional dynamic deformation of bimodular materials. The results obtained are of practical value to researchers in solid mechanics, specialists in numerical modeling, and engineers involved in the design of structures made of bimodular materials under dynamic loads.

Keywords: bimodular elastic medium, one-dimensional strain waves, boundary value problem, piecewise-linear approximation, evolutionarity of discontinuities, shock waves, rigid layers, numerical modeling, automation of calculations, convergence analysis

Mivar expert system for an access control system based on palm vein biometric identification

2026. T.14. № 6. id 2393

Grigorenko K.D. Gorenkov A.A. Belyaev I.A. Varlamov O.O.

DOI: 10.26102/2310-6018/2026.57.6.014

The paper addresses the task of designing an access control system (ACS) capable of making informed decisions not only on the basis of a biometric template but also taking into account the context: employee access level, zone of access, time-of-day policy and authentication history. An architecture of a complex intelligent system combining neural-network and logical levels of artificial intelligence is proposed. Biometric identification is implemented using a ResNet18 convolutional neural network adapted for grayscale palm-vein images and trained on a dataset of 834 subjects (8,340 images) using triplet metric learning with a classification head; Top-1 accuracy of 87.47 %, Top-5 of 96.58 %, Top-10 of 98.14 %, ROC-AUC of 0.9985 and EER of 1.64 % are achieved with an average confidence of correct matches equal to 0.908. The neural-network confidence together with five contextual parameters is passed to a mivar expert system (MES) implemented in the KESMI Wi!Mi Razumator environment. The MES contains three independent relations with shared inputs that produce an access decision, an alert level and a biometric reliability estimate. Shared inputs induce cross-edges in the bipartite solution graph, reflecting the multi-aspect nature of decision making in complex AI. A decision algorithm of 32 rules grouped into five priority tiers is developed. Testing on three representative scenarios demonstrates three distinct topologies of the solution graph – from a degenerate case to a full bipartite one. The results confirm the correctness of mivar logical inference and the scalability of the knowledge base without any change to the neural-network module.

Keywords: mivar expert system, ACS, biometric identification, palm veins, resNet18, triplet learning, complex artificial intelligence, KESMI, wi!Mi, neurosymbolic artificial intelligence

Method of blockchain-consensus distributed image classification

2026. T.14. № 7. id 2392

Konnikov E.A. Sandu N.R. Faizullin R.V.

DOI: 10.26102/2310-6018/2026.58.7.003

The study focuses on inter-node aggregation methods for model updates and local visual representations in distributed image classification systems. The paper aims to develop and experimentally validate a blockchain-consensus distributed image classification method that improves the robustness of the global model under heterogeneous, noisy, and partially malicious data sources. The proposed approach combines local class atlases, audit-based utility estimation of client updates, robust anomaly scoring of local trajectories, and blockchain-backed consensus weighting of node contributions. The blockchain layer is implemented as a reproducible software ledger for update provenance and audit logging. The experimental evaluation relies on the open CIFAR-10 and Olivetti Faces datasets in four scenarios: iid_clean, noniid_clean, noniid_noisy, and noniid_adversarial. The method is compared with Distributed Mean, Distributed Proximal, Distributed Trimmed Mean, and Atlas Consensus using Accuracy, Macro-F1, Balanced Accuracy, convergence dynamics, consensus weights, and statistical testing. The results show that the proposed method is not universally superior on clean distributions; however, in the target adversarial scenario it achieves the best performance among distributed schemes. The findings confirm the practical value of blockchain-consensus filtering of node updates for distributed visual data processing.

Keywords: blockchain, distributed image classification, inter-node aggregation, visual data, heterogeneous distributions, malicious nodes, robust aggregation, consensus, class atlas, information processes

A mathematical model for optimizing security measures for facilities based on matrix games

2026. T.14. № 9. id 2391

Sergeev N.P. Stepanov L.V. Zarubin V.S.

DOI: 10.26102/2310-6018/2026.60.9.013

The relevance of this study stems from the need to formalize the antagonistic confrontation between facility security measures and the operational vulnerabilities of systems while simultaneously accounting for uncertainty and resource limitations. Therefore, this article aims to develop a mathematical model that enables the rational selection of the optimal composition of protective measures in conflict situations. The leading approach to studying this problem is the framework of matrix games, which allows for a comprehensive consideration of the interaction of protective strategies and destructive factors based on uniform expert criteria. The article presents three options for forming a payoff matrix (homogeneous and heterogeneous strategies, direct expert assessment of mutual influence), reveals mechanisms for finding a solution in pure strategies through a saddle point, identifies conditions under which a solution exists only in mixed strategies, and substantiates a method for calculating the proportions of combining heterogeneous protective measures to achieve a guaranteed result. The article's materials are of practical value to specialists designing physical and information security systems, as well as to decision makers who allocate limited resources to protected facilities under deliberate countermeasures.

Keywords: mathematical modeling, game theory, matrix game, zero-sum game, facility security, security measures, vulnerability, expert assessment, saddle point, mixed strategies

Mivar expert system for identifying challenging observation conditions and selecting image correction methods for a single-camera vision system of an autonomous delivery robot

2026. T.14. № 7. id 2390

Milevich A.A. Ovchinnikov D.A. Varlamov O.O.

DOI: 10.26102/2310-6018/2026.58.7.002

The paper addresses the problem of intelligent decision support for the vision subsystem of a mobile autonomous delivery robot equipped with a single camera and operating under limited onboard computational resources without relying on cloud computing. The relevance of the study is determined by the degradation of image perception under insufficient illumination, overexposure, precipitation, haze, glare, digital noise, blur, partial occlusion, and lens contamination. A mivar expert system is proposed to formalize the domain knowledge in terms of parameters, relations, rules, and constraints. The user specifies 13 input parameters: ten normalized image features and three contextual attributes – location, time of day, and season. The system produces two semantically interpretable outputs: the current challenging observation condition and the recommended image enhancement action. A key feature of the model is its atomic relation structure: each relation contains no more than one conditional operator, while complex logic is represented as a chain of simple rules. Additional contextual service features are introduced to account for dense urban environment, nighttime city illumination, and seasonal effects when selecting the inference branch. A pretrained YOLO detector is used as the object recognition module, while the MES serves as an explainable and computationally efficient layer for interpreting challenging observation conditions. This combination makes the proposed approach suitable for local onboard deployment in an autonomous delivery robot. The obtained results confirm the feasibility of an explainable, modular, and extensible expert system for supporting a single-camera autonomous delivery robot under adverse observation conditions.

Keywords: mivar expert system, autonomous delivery robot, vision system, single-camera observation, challenging observation conditions, limited computational resources, explainable artificial intelligence, production rules, KESMI

Using systems analysis in designing intelligent protection systems against malicious software

2026. T.14. № 8. id 2389

Abedlhussain A.A. Lyapuntsova E.V.

DOI: 10.26102/2310-6018/2026.59.8.011

The paper addresses the problem of designing an intelligent protection system against malicious software under changing input flows, limited computational resources, and the need for stable operation on new data. The aim of the study was to show how systems analysis methods make it possible to move from a simple comparison of classification models to a justified choice of architecture, evaluation strategy, and update scenario. The experimental basis included 134435 objects, of which 57293 were malicious. After removing records without a recognized date and excluding the early archival fragment, an experimental subset of 123704 objects covering the period from August 2019 to September 2020 was formed. The study compares randomized and chronological evaluation schemes, baseline classification models, several feature reduction variants, and both single-stage and two-stage protection architectures. The results show that using 256 features in the first stage provides almost the same recognition quality as the full 2381-feature representation while requiring substantially lower computational costs. It is also shown that the two-stage architecture preserves the quality of the full model while sending only 0.25 percent of objects to the second stage and almost not increasing the average processing time per object. In addition, first-stage updating improves temporal robustness and reduces the risk of quality degradation on later monthly data segments.

Keywords: systems analysis, malicious software, intelligent protection system, chronological evaluation, two-stage architecture, feature reduction, adaptive updating, machine learning

Analysis of modern approaches and formalization of parameters for the management of mechanical ventilation modes

2026. T.14. № 5. id 2388

Frolov S.V. Sudakov D.E. Dolgov E.P.

DOI: 10.26102/2310-6018/2026.56.5.015

The relevance of this study is determined by the high need for respiratory support in intensive care unit patients (up to 50 % of patients) and the significant risk of ventilator-associated lung injury (VALI) due to suboptimal ventilator settings. Modern mechanical ventilators offer dozens of modes and over 50 adjustable parameters, creating a high cognitive load on the physician and increasing the likelihood of errors. The aim of this work is to systematize knowledge about modern mechanical ventilation modes and formalize key parameters of respiratory support for the subsequent development of intelligent clinical decision support systems (CDSS). The study employs methods of analytical review, classification, mathematical modeling of respiratory mechanics, and formalization of clinical criteria. An analysis of factors justifying the need for CDSS was performed: the complexity of interpreting respiratory mechanics (compliance, resistance, driving pressure), the high incidence of complications due to incorrect settings (barotrauma in 10–15 % of patients with plateau pressure >30 cm H₂O), time constraints for ICU physicians, and non-standardized nomenclature of modes across different manufacturers. A classification of ventilation modes by level of intelligence (from mandatory to fully automated) is provided, and key ventilation parameters (tidal volume, rate, pressures, flow, PEEP) are described in detail. Four groups of parameters for mode selection are formalized: lung mechanics (static compliance, resistance, plateau pressure, P0.1, driving pressure), gas exchange (PaO₂/FiO₂, PaCO₂, SpO₂), patient activity (respiratory rate, asynchrony signs), and hemodynamics (blood pressure, central venous pressure). Specific criteria for each parameter are proposed. A logical algorithm for mode selection based on formalized parameters is developed. The obtained results provide a foundation for building a production rule base for CDSS, enabling physicians in time-critical situations to receive justified recommendations. Further research should focus on clinical validation of the proposed criteria and the development of explainable artificial intelligence algorithms for personalizing respiratory support.

Keywords: mechanical ventilation, respiratory support, ventilator-associated lung injury, ventilation modes, clinical decision support, parameter formalization, respiratory mechanics, intelligent algorithms

Reducing of the microscope error for the control of through holes by optimizing the relative threshold offset coefficient with adaptive image binarization

2026. T.14. № 10. id 2386

Kovalev A.V. Bushuev S.V. Shulepov A.V.

DOI: 10.26102/2310-6018/2026.61.10.001

Video measuring microscopes (VMM) are widely used in scientific research and in production to control the geometry of objects of complex shape, however, the measurement results on them strongly depend on lighting conditions and image processing parameters. In this paper, the influence of the backlight level on the choice of the optimal value of the relative threshold offset coefficient relative to the local background for adaptive binarization of images when measuring through holes is investigated. This parameter has been optimized according to the criterion of minimizing the deviation of the measured diameter from the reference data obtained on the coordinate measuring machine (CMM, Carl Zeiss). After processing the images obtained from the microscope and calculating the parameters of the obtained circle profiles, it was shown that the optimal value of the relative threshold offset coefficient demonstrates a close to linear dependence on the level of the lower illumination (30–85 in conventional VMM units). Based on the data obtained, a method was proposed for automating the selection of the relative threshold offset coefficient, which reduces the operator's influence on the measurement results. The materials of the article can be used to increase reproducibility and reduce measurement error in industrial systems of automated optical geometry control of finished products.

Keywords: video measuring microscope, measurement optimization, optical measurement errors, projection measurements, pattern recognition

Multi-criteria model of rational management of information flows in a digital organizational system

2026. T.14. № 7. id 2380

Mikheev V.S.

DOI: 10.26102/2310-6018/2026.58.7.006

The article proposes a formalized model for the rational management of information flows within a digital organizational system. The aim of the study is to develop a procedure for selecting a control action in which the information flow is treated as an independent object of management, characterized by a measurable state, defined constraints, and a set of admissible alternatives. The methodological basis of the study includes a systemic description of the digital organizational system, a six-parameter model of the flow state, normalization of criteria with different units of measurement, weighted aggregation of indicators, and verification of threshold constraints. The state of the information flow is described through a set of parameters: intensity, information transmission delay, data quality, availability, maintenance cost, and integral risk. As a result, a multicriteria model is developed that makes it possible to compare organizational, software-related, and infrastructural control actions using a unified integral rationality index. The study demonstrates that a rational decision is determined not by maximizing a single indicator, but by achieving a coordinated balance between throughput, timeliness, quality, availability, costs, and risk level. To assess the applicability of the model, the article presents a numerical example involving four management alternatives. The results demonstrate the possibility of selecting a compromise control action under conditions of multiple interrelated criteria. The scientific novelty of the proposed approach lies in integrating the information flow state model and the control action selection procedure within a single computational framework. The proposed approach can be applied in the design of digital management loops, corporate monitoring systems, service platforms, and decision support systems.

Keywords: information flow, digital organizational system, rational management, multi-criteria choice, data quality, information availability, integrated risk, control action, rationality index

A mathematical model of the process of automated construction of review texts based on the use of quotation-aware summarization

2026. T.14. № 6. id 2375

Kuznetsov I.I.

DOI: 10.26102/2310-6018/2026.57.6.020

This paper examines the automated generation of scientific review texts based on the analysis of a corpus of scientific publications. A mathematical model has been developed that describes an approach based on the use of citation-aware summarization with preliminary selection of citation fragments and formalizes the full data processing cycle, from publication selection and extraction of citing fragments to the generation of individual summaries and the final review text. The model defines a unified formal description of the sequence of data transformation operators and includes a system of quality criteria that ensure control of the plausibility, coverage, and factual consistency of results at all stages of processing. A software system in the form of a modular pipeline is implemented based on the developed model. Experimental studies using the developed model were conducted on the SurGE dataset, which includes 114 topics, over 7,000 cited and over 73,000 citing publications. It is shown that the use of a citation-aware approach with preliminary fragment selection improves the quality of summary generation compared to alternative methods. The following quality criteria were achieved for the resulting review texts: credibility – 0.8744, coverage – 0.9356, factual reliability – 0.9713, and LLM score – 0.9232, which outperforms the results of generation based on the full text of sources (by 7.67 % for credibility, 4.63 % for coverage, 9.22 % for factual consistency, and 4.42 % for LLM score). The obtained results confirm the effectiveness of the proposed approach and the developed model and their applicability for the automated generation of reliable scientific reviews.

Keywords: citation-aware summarization, scientific reviews, large language models, text information analysis, scientific publications, citation, citation extraction

Control system for a vibrotactile vest and weapon controller of an XR trainer based on an inertial measurement unit

2026. T.14. № 10. id 2370

Kravets A.G. Korobkin P.D.

DOI: 10.26102/2310-6018/2026.61.10.007

A method for event-driven coordination of two interconnected XR-trainer subsystems – a three-channel vibrotactile vest and a weapon controller with an MPU6050 inertial module – is proposed. The method is formalized by the rule F: (s, q, c) → u, which maps event type, controller orientation vector, and scenario-state context to a control-action vector for the XR scene, weapon controller, and vest; exchange is implemented via JSON/HTTP messages. The vest provides targeted vibrostimulation across body zones, while the controller simulates recoil and smoke and transmits orientation parameters to the virtual environment. For the vest power stage, a computational model of current and energy regimes is developed, relating the number of simultaneously active channels to instantaneous and average current and battery-life estimates; the model's consistency is verified by LTspice circuit simulation. For the MPU6050, a processing algorithm is formed, comprising static bias correction, moving-average smoothing, and complementary filtering with α = 0.98. During preliminary quasi-static verification, the observed absolute pitch and roll deviations ranged from 1 to 2 degrees. Because the source data do not provide the number of measurements, repetitions, or statistical characteristics of the series, this range is not interpreted as a statistically validated accuracy estimate. Dynamic modes, yaw, and end-to-end system integration require separate validation.

Keywords: vibrotactile vest, XR trainer, IMU, MPU6050, data fusion, complementary filter, LTspice, MOSFET, raspberry Pi

Fake news detection in low-resource languages with LLMs

2026. T.14. № 6. id 2368

Kabir A. Khan S. Kharlamov A.A. Voronkov I.M.

DOI: 10.26102/2310-6018/2026.57.6.009

The proliferation of fake news is a global challenge to tackle in the digital era of information availability. The resourceful languages are tackling this issue through enormous research works whereas the low-resource languages are left behind to address the issue adequately. Bangla is one of the low-resource languages in computation despite being in the top ten most spoken languages in the world. To contribute in the field and address the issue of fake news, this research work focuses on the fake news detection in Bangla language leveraging large recent advancement of language models using cross-lingual prompting techniques for better response from the large language models. We leverage the open source models for resource accessibility and utilize DeepSeek-R1, Llama 3.2 and Qwen 2.5 large language models in our experiments and extensively analyze the fake news detection capacity of each model in Bangla language. We find that Qwen 2.5 outperforms the other models in this specific task achieving a maximum accuracy of 97.5 while it also reports no inconclusive response.

Keywords: fake news, bangla, large language models, low-resource language, cross-lingual prompting

Evaluation of the effectiveness of the hybrid RLS algorithm with α-β pre-filtering in the suppression of diverse interference in antenna arrays

2026. T.14. № 7. id 2366

Glushankov E.I. Ageychik N.O.

DOI: 10.26102/2310-6018/2026.58.7.004

The relevance of this study stems from the widespread use of antenna arrays in radar, navigation, and communication systems, where interference suppression efficiency critically depends on adaptive algorithms. The classical RLS algorithm loses stability in the face of pulsed and non-stationary interference, which poses challenges for operation in real-world conditions (moving objects, the Doppler effect, and intentional jamming). The objective of this study is a comparative analysis of the classical RLS and a hybrid RLS algorithm with adaptive α-β pre-filtering under the influence of 11 types of heterogeneous interference. The study was based on simulation modeling implemented in the MATLAB environment. For each of the 11 types of interference, the following parameters were varied: the interference-to-signal power ratio (ΔP, which ranged from 1 to 11), the angular misalignment between the signal and interference arrival directions (range from 0° to 9°), and the presence of the Doppler effect (carrier frequency shift of 0 % and 10 %). A total of 24,200 computational experiments were conducted, processing over 193 million data samples. The classical RLS algorithm demonstrates acceptable performance only under stationary conditions. When exposed to pulsed, semi-periodic, and combined interference, its output signal-to-noise ratio decreases to negative values. In contrast, the hybrid RLS+αβ algorithm provides a stable output signal-to-noise ratio level above 20 dB in all considered scenarios, and in the worst-case scenarios, it outperforms classical RLS by 20–25 dB. The Doppler effect has virtually no effect on the performance of the hybrid algorithm, whereas for classical RLS, it leads to divergence of weighting coefficients and loss of performance. The proposed hybrid algorithm is a universal and robust solution for antenna arrays operating in complex, non-stationary interference environments. The results can be used in the design of radar systems, unmanned aerial vehicles navigation, and electronic warfare systems.

Keywords: adaptive antenna arrays, hybrid RLS algorithm, α-β filter, spatial filtering, noise immunity, impulse noise, combined noise, signal-to-noise-plus-interference ratio, doppler effect, non-stationary noise

Optimizing the operation of a digitized organizational system through proactive management of information environment parameters

2026. T.14. № 6. id 2362

Gotishan A.A. Lvovich Y.Y.

DOI: 10.26102/2310-6018/2026.57.6.013

The paper considers the problem of optimizing the functioning of a digitalized organizational system in a dynamically changing information environment. It is shown that traditional reactive control methods do not provide the required level of reliability and efficiency under high load variability. The necessity of transition to proactive management based on anticipatory changes in the parameters of the information environment is substantiated. A formalization of the optimization problem is proposed, taking into account the total costs of system operation and probabilistic reliability requirements. A method for selecting optimal control parameters is developed, based on the introduction of a generalized functional that combines the efficiency criterion and penalty constraints, as well as an iterative procedure for parameter correction using pseudo-random initialization, an adaptive step, and a mechanism for updating penalty coefficients. A feature of the method is taking into account predicted changes in the state of the system, which makes it possible to implement a proactive control mechanism. It is shown that the proposed approach provides cost reduction, increased operational stability and prevention of critical system states. The limitations of the method (sensitivity to the choice of hyperparameters, dependence on the quality of predictive models) and directions for its further development are identified. The practical significance of the work lies in the possibility of applying the proposed approach in resource management systems for cloud platforms, container orchestrators and other digitalized organizational systems with dynamically changing loads.

Keywords: digitalized organizational system, proactive management, parameter optimization, information environment, reliability, costs, iterative algorithm, pseudo-random initialization, adaptive step, time series forecasting

Automated monitoring and control system for the parameters of an ionized steam environment

2026. T.14. № 6. id 2351

Stepulev A.I. Dulskiy E.Y. Ivanov P.Y. Kovshin A.S. Medvedev V.P.

DOI: 10.26102/2310-6018/2026.57.6.019

This paper addresses the development of an automated system for monitoring and controlling the parameters of an ionized vapor medium. The relevance of the work stems from the need to ensure stable physicochemical characteristics of such media in industrial and laboratory applications. It is shown that the formation of an ionized vapor medium should be considered a controllable process, whose state is determined by a set of measurable physicochemical parameters. A structural and functional scheme of the system is proposed, comprising a control plant, a measurement circuit, a data processing unit, and an actuating circuit that implements mode correction via closed-loop feedback. The system operates by continuously monitoring key parameters and automatically adjusting vapor generation modes to maintain specified characteristics. To formalize the state of the medium, a vector of controlled parameters is introduced, including oxidation-reduction potential (ORP), temperature, pH, and electrical conductivity, along with an integral criterion for deviation from the target regime. As a first research stage, initial validation of the ORP measurement channel was performed. Experiments were conducted on a laboratory bench simulating real vapor generation conditions. It was found that during intensive vapor generation, negative ORP values are recorded at all control points, and this parameter demonstrates spatiotemporal sensitivity to the vapor formation regime. The obtained results confirm the feasibility of using ORP as a basic feedback parameter for the subsequent construction of a multichannel automated system for monitoring and controlling ionized vapor medium parameters. Future research will focus on integrating all measurement channels and developing multiparameter control algorithms.

Keywords: ionized steam environment, automated monitoring, automated control, oxidation-reduction potential, feedback, environment state vector, structural and functional diagram, control algorithm