Vol 12, No 4 (2025)

Extracting structured information from text is a key task in natural language processing. Large language models for information extraction tasks achieve high accuracy thanks to pre-training on huge volumes of data. However, such models require significant computational resources and are unavailable for local use due to their dependence on cloud infrastructure. Therefore, compact, open-source large language models that can be retrained locally are increasingly being used to address this problem. This paper evaluates the effectiveness of retraining compact large language models for automated triplet information extraction from unstructured text. The Mistral model with seven billion parameters was used in the study. The model was fine-tuned on a custom dataset consisting of 650 examples, each containing an instruction, an input text and an expected output. The results confirm the effectiveness of retraining: the F1-score increased several-fold compared to the baseline model. The retrained version of the model demonstrates competitiveness with the large-scale DeepSeek language model with 685 billion parameters. The obtained results highlight the potential of compact open large language models for knowledge extraction tasks under resource constraints, such as knowledge graph construction.

In this work, the algorithm determines the identification of modal parameters of engineering structures and buildings described as a linear dynamic time-invariant system in spatial change. Modal parameters are estimated on the basis of recorded vibration response under the assumption of random nature of the disturbing forces. The paper describes the features of the algorithm and provides references to relevant sources that allow a deeper understanding of the algorithm details. The paper proposes an approach to determining the values of modal parameters when performing a number of consecutive identifications, which can be further applied to automate the process for real-time operation or when processing the results of multiple testing. The algorithm allows you to obtain a stable model. The invariance of the system in time is a key factor that ensures the synthesis of mathematical models for ensuring the information security of the objects under consideration.

Purpose of the study. To develop and theoretically substantiate a model of intelligent information processing technology designed to support management decision-making in small and medium-sized enterprises (SMEs) under conditions of uncertainty and data incompleteness, based on the application of a regularizing Bayesian approach (RBP).

Methods of research: systems analysis, decision theory, artificial intelligence and machine learning methods, in particular, Bayesian belief networks and neural networks, as well as probability theory and mathematical statistics. The core of the methodology is the regularizing Bayesian approach, which allows for the formalization and consideration of prior information to improve the robustness of models on small samples.

Results. Based on the conducted analysis, a structural and functional model of an intelligent SME management technology is proposed. The model integrates data collection and preprocessing modules and a Bayesian inference kernel implementing regularization procedures. It is shown that the use of BBP reduces the risk of model overfitting with the limited statistical data typical of SMEs and improves the quality of management forecasts and decisions. Recommendations for the application of the technology for demand forecasting, risk assessment, and personnel management are developed.

Scientific novelty: adaptation and development of the regularizing Bayesian approach methodology for solving semi structured management problems in small and medium-sized enterprises. Unlike standard machine learning methods, the proposed technology formally incorporates prior expert information and industry knowledge for decision regularization, which is critical in the highly volatile and data-poor environments typical of the SME sector.

The aim of the research is to develop a procedure for aggregating data on the required quality level of complex data processing systems through the integration of multicriteria analysis with machine learning methods. Existing approaches based on GOST R 59797–2021 and ISO/IEC 25010 demonstrate limited efficiency due to the absence of a unified aggregation procedure]. A three-stage hybrid procedure has been devised: collection and normalization of quality indicators using a modified z-transformation; calculation of adaptive weights via synthesis of the AHP method with a random forest algorithm; formation of an integrated criterion for the required quality level. Validation was carried out on two industrial systems with scales of 50–80 TB/day.

Results include an increase in forecast accuracy from 82.1 to 92.4%, a 3.4-fold reduction in decision-making time, and a decrease in critical incidents by 34–45%. The algorithmic complexity is O(n²m + n log nk), with execution time under 30 seconds. The procedure is applicable to CDPS with data volumes exceeding 10 TB/day and requires at least 500 historical observations. The findings are valuable for architects and specialists in quality management of critically important information systems.

This article presents a software and laboratory suite for studying the mathematical foundations and practical applications of elliptic curve cryptography (ECC). The suite is implemented in Python using the PyQt6 framework and the sympy library for cryptographic computations. The program provides an interactive interface for entering elliptic curve parameters, visualizing points on the curve, constructing Cayley tables for point addition, and checking group properties. Key features of the suite include the implementation of the Tonelli–Shanks algorithm for finding absolute square roots, the ability to work with curves over finite fields of large order, and a bilingual interface (Russian/English). The developed suite can be used in educational settings to teach the fundamentals of elliptic curve cryptography.

The article proposes an algorithm for alerting the population in case of fires using feedback via SMS and mapping services. The mechanism integrates distributed ledger technology and artificial intelligence to improve the accuracy, coverage and adaptability of the system. The article also develops a model that takes into account the distribution of population density during alerting in a given zone, as well as the optimal radius of the alert zone. The model is implemented in the form of software “Model of the effective radius of alerting in case of fire”. Results. The results obtained can be used to adjust the existing model of public notification in case of large fires, both natural and man-made. This work is intended for those who make managerial decisions and manage the forces and means of extinguishing fires.

Understanding the structural characteristics of amorphous alloys at the atomic scale is crucial for elucidating their unique mechanical, thermal, and magnetic properties. However, the absence of long-range order in these materials poses significant challenges for conventional structural analysis techniques. This work presents a GPU-accelerated software framework designed for high-throughput processing and quantitative analysis of High-Resolution Transmission Electron Microscopy (HRTEM) images to reveal hidden atomic orderliness in amorphous alloys. The proposed system integrates parallelized image preprocessing, processing, atom detection, radius-based clustering, and graph-theoretical and entropy-based metrics to quantify short- and medium-range order. A modular architecture enables efficient GPU computation using CUDA, CuPy, and optimized memory strategies, achieving speedups of up to 220× compared to CPU implementations. Validation was conducted on both simulated datasets (FeB, CoNiFeSiB) and real HRTEM images of amorphous alloys (CoP, NiW, Fe-based 71КНСР). Results demonstrate strong correlations between cluster size, bond angle distributions, and entropy metrics with macroscopic material properties such as hardness and thermal stability. Larger clusters and obtuse bond angles were found to indicate increased local structural order, while entropy measures provided sensitive discrimination of disorder.

This paper presents a prototype of a digital twin (DT) for an automated control system (ACS) of a smart power grid, developed for the analysis of cybersecurity threats. The proposed architecture replicates the behavior of the control layers of the energy system and incorporates modules for synthetic data generation, attack simulation, anomaly detection, and threat assessment. Experimental validation was carried out in a laboratory environment through the execution of typical cyberattack scenarios (DoS, malware injection, control signal compromise). A comparative evaluation of two configurations – one based solely on real data and the other incorporating synthetic data – demonstrated an increase in F1-score metrics when using extended datasets. The study discusses the limitations of the prototype, including simplified modeling of physical processes and the need for manual verification of generated data. The results suggest the applicability of the proposed approach for testing threat detection mechanisms in smart grid environments.

This article proposes a two-factor authentication system in WEB applications based on the fingerprint method. It provides an overview of known two-factor authentication methods in WEB applications and shows the disadvantages of such systems. The architecture of a two-factor authentication system is proposed, cryptographic algorithms for a digital signature are selected, and algorithms for interaction between system components during registration of new devices are developed. An algorithm for interaction between system components during authentication is developed and a software implementation in C# is proposed, and the choice of this programming language is justified. The program components are shown, component codes are presented, and the server and client parts of the system are presented during registration of a new user in the system. The quality of the developed system is assessed, and the coefficients of false rejections for five fingers of the right hand are shown.

The article is devoted to the development and substantiation of a multimodal methodology for non-destructive monitoring of the condition of the road surface based on acoustic data supplemented by visual observations. At the level of feature analysis, the study demonstrates that the spectrograms of signals recorded when driving over smooth and damaged surfaces contain stable differences in the time-frequency structure suitable for automatic classification and mapping of defects. The requirements for the type of microphone (sensitivity, bandwidth over 8 kHz, directivity) and placement conditions (height, distance to the source, shielding objects) are justified. A two-sensor architecture with time synchronization combining stereo video and acoustics is proposed; it is shown that the correlation of modalities increases the reliability of localization and the quality of defect classification in real traffic conditions. Taken together, the presented approach forms the basis for long-term, fault-tolerant monitoring systems for road infrastructure with early detection of risks, even with increased noise levels and reduced visibility.

This paper addresses the problem of decomposition of complex systems represented as Petri nets, with the aim of optimizing the synthesis procedure for new parallel systems. An improved approach is proposed, combining Johnson’s algorithm with the depth-first search method. Johnson’s algorithm is applied for the efficient detection of elementary cycles, which makes it possible to overcome the limitations of the traditional matrix-based method, characterized by high computational and spatial complexity, as well as the inability to guarantee the extraction of elementary cycles of a given length. An adapted depth-first search algorithm is used to identify linear fragments of acyclic Petri nets. An analytical and experimental comparison of the proposed algorithms with well-known counterparts has been carried out on both complete and sparse Petri nets. The experimental data indicates a significant advantage of Johnson’s algorithm when applied to sparse Petri net structures, manifested in increased execution speed and completeness in detecting elementary cycles. The proposed approach demonstrates a substantial reduction in time costs at the decomposition stage and contributes to the creation of conditions (a knowledge base) that restrict the set of synthesized structures and simplify subsequent stages of parallel system design. The obtained results confirm the practical efficiency and applicability of the proposed Petri net decomposition methods in the synthesis of complex system structures.

The aim of the study is to analyze the applicability of fuzzy cognitive maps (FCM) to analyze the reliability of complex human-machine systems (HMS), as well as to develop algorithms for evaluating factors influencing the reliability of the system, taking into account expert assessments. The paper highlights the limitations of classical probabilistic and regression methods, which are difficult to apply to HMS due to the interdependence of qualitative assessments, as well as the need to take into account the presence of a human factor. As an alternative solution, the use of fuzzy cognitive maps is considered, which provides the possibility of representing the dynamics of the system in the form of an oriented weighted graph, where the vertices are key concepts, and the arcs are cause–and-effect relationships evaluated by experts. Using the example of an analysis of the reliability of an intelligent video monitoring system of a protected object, the construction of a fuzzy cognitive map is demonstrated, an algorithm for calculating importance indices and coefficients of the combined influence of factors is given to determine the integral indicator of the reliability of the system. A computational algorithm has been developed, and the results of its software implementation are presented. The factors that have the greatest impact on the target variable are highlighted. The prospects of using graph knowledge bases for organizing the collection and storage of information forming cognitive fuzzy maps are noted. The advantages of the considered approach include the possibility of using the method when working with expert information, the integration of heterogeneous factors within a single model, its adaptability and scalability.