Menu
Publications
2026
2025
2024
2023
2022
2021
2020
2019
2018
2017
2016
2015
2014
2013
2012
2011
2010
2009
2008
2007
2006
2005
2004
2003
2002
2001
Editor-in-Chief
Nikiforov
Vladimir O.
D.Sc., Prof.
Partners
Summaries of the Issue
OPTICAL ENGINEERING
Dependence of the recognition probability of an aerial reconnaissance object with linear resolution on the ground and pixel information content of the image
Andrey S. Molchanov891
The paper considers a quantitative relationship between the probability of recognizing an aerial reconnaissance object, the actual linear resolution on the ground, and the pixel information content of a digital image. The objective is to obtain a computational model for assessing whether an object can be recognized at the levels of species, class, subclass, and type, taking into account the minimum number of informative pixels and the stability of recognition features. The digital portrait of an object is represented as a set of pixel configurations containing geometric, brightness, contrast, and structural features. The model includes the effective projection of the object, viewing angle, contrast, image smear, representativeness of features, and local resolution of characteristic details. A pixel-resolution indicator of recognition sufficiency is proposed. It expresses the ratio of the actual effective pixel information content of an image to the information content required for a specified classification level. Direct relationships are obtained for calculating recognition probability from a given linear resolution on the ground, and inverse relationships are obtained for determining the required resolution and the minimum number of pixels from a specified recognition probability. The model can be applied in planning aerial imaging, testing optoelectronic systems, and setting requirements for reference digital-portrait databases of observed objects.
The influence of approaches for measuring the spectral characteristics of fiber-optic elements of quantum communication systems on the key generation rate
Oleg A. Solomatin, Polina A. Morozova, Vladimir V. Chistiakov, Boris A. Nasedkin 900
A method for estimating insertion losses introduced by spectrally selective elements is proposed, which enables an increase in the dynamic range of measurements and subsequent calculations of the performance characteristics for a Continuous Variable Quantum Key Distribution system. In contrast to common approaches that employ supercontinuum sources, the proposed method provides more accurate assessment of the system’s resilience against the Trojanhorse attack. The essence of the method consists in using a tunable laser source with additional filtering incorporated into the optical setup for insertionloss measurements. The paper presents a comparative analysis of loss measurement results obtained with a supercontinuum source and a tunable laser source. For the tunable source, the addition of extra filtering in the optical scheme is proposed. It is demonstrated that the maximum loss values measured with the tunable source exceed those obtained with the supercontinuum source by 20 dB. The experiments reveal that employing the tunable source together with additional filtering enhances the dynamic range of insertionloss measurements for Coarse Wavelength Division Multiplexing filters. Using the Trojanhorse attack as a case study, it is shown that additional filtering in the tunablesource configuration yields more accurate transmittance values. This, in turn, affects the formulation of requirements for the implementation of countermeasures, aiming to maintain the quantum key generation rate as close as possible to the case without the attack.
Plasma plume radiation spectrograms for assessing weld quality during laser welding of steel
Gresko Vladislav R., Dolgopolov Arthur D. , Mikhail S. Sokolov 909
Steel grade 09G2S is widely employed in the manufacture of critical structures and components operating under high mechanical loads and in aggressive service environments. Laser beam welding is widely applied for joining this material due to its high precision and productivity. However, variations in process conditions may result in weld defects that affect joint quality and structural integrity. Optical emission spectroscopy (OES) of laser-induced plasma, combined with machine learning techniques, has emerged as a promising approach for real-time weld quality monitoring. Nevertheless, its industrial implementation requires a thorough understanding of the relationships between plasma spectral characteristics and weld defects, as well as the identification of the most informative spectral regions and extracted features. The aim of this study was to investigate the correlation between the spectral characteristics of laser-induced plasma generated during laser welding of 09G2S steel and the geometric dimensions of weld defects. Laser welding experiments were performed using a continuous-wave fibre laser under varying laser power and welding speed conditions. Plasma emission spectra were collected in three wavelength ranges using a fixed fibre-optic spectrometer. Several features, including the mean, standard deviation, and kurtosis, were extracted from the spectra and analysed using correlation analysis to determine their relationship with weld defect dimensions. The strongest correlations between the extracted features and weld defect geometry were observed for discontinuous undercut and sagging defects within the spectral range of 460–710 nm. The findings demonstrate the potential of optical emission spectroscopy for in-process quality monitoring during laser welding of 09G2S steel. The identified spectral features provide a foundation for the development of machine learning models capable of reliable and real-time weld defect detection in industrial laser welding applications.
MATERIAL SCIENCE AND NANOTECHNOLOGIES
Luminescence mechanisms in Tb- and Dy-doped magnesium tetraborate samples
Zakharchuk Ivan A. 919
The luminescent properties of thermoluminescent detectors based on magnesium tetraborate doped with terbium or dysprosium and sodium were studied. Analysis of the obtained results allows verifying the charge state of the dopant ions. Magnesium tetraborate was prepared by a solid-state reaction between boric acid and magnesium carbonate or magnesium oxide. After grinding, the reaction mixture was dried at 333–353 K, dehydrated at 473 K, and annealed in two stages at 1023 K and 1143 K, respectively. Dysprosium, terbium, and sodium impurities were introduced before the final annealing step. The photoluminescence signal was recorded using a spectrometer based on a double monochromator under monochromatically selected excitation from a xenon lamp. Cathodoluminescence was recorded with a CCD spectrometer under electron-beam excitation. Photoluminescence decay kinetics were measured with a photomultiplier tube under laser excitation, whereas cathodoluminescence decay kinetics were recorded using a photomultiplier tube and an oscilloscope. Photoluminescence, photoluminescence excitation, and cathodoluminescence spectra, as well as photo- and cathodoluminescence decay curves, were obtained for magnesium tetraborate samples doped with terbium or dysprosium and sodium. Dysprosium photoluminescence was absent in the non-irradiated sample and appeared only after electron-beam excitation. This indicates a change in the charge state of the dysprosium dopant. Thermally stimulated luminescence of the samples is associated with radiative transitions of terbium and dysprosium in the 3+ charge state. The dopant charge state determines the excitation pathway of the luminescence center during heating. In the studied samples, dysprosium predominantly occurs in the 2+ charge state, whereas terbium is initially present in the 3+ charge state. The obtained results can be used when selecting rare-earth dopants for the development of magnesium-tetraborate-based thermoluminescent materials. The established relationship between the dopant charge state and the thermoluminescence mechanism provides a basis for targeted control of the practical properties of these materials.
Modeling of electronic states and angular absorption anisotropy in cadmium telluride nanocrystals with different geometric dimension ratios
Daibagya Daniil S.926
Control over the electronic and optical properties of semiconductor nanocrystals is often considered primarily in terms of changes in their size. However, particle shape and the ratio of longitudinal to transverse dimensions may also influence their optical properties. In this work, numerical modeling of electronic states and optical density spectra was performed, and the angular anisotropy of absorption was investigated for cadmium telluride nanocrystals of different shapes by varying the ratio of their longitudinal to transverse dimensions. The calculations were carried out for spheroidal, cylindrical, and conical cadmium telluride nanocrystals with a fixed volume of 422 nm3. Electronic states and optical density spectra were calculated within the effective-mass approximation at a temperature of 300 K. The angular dependence of the optical density spectra was analyzed for incident angles of 0°, 30°, 45°, 60°, and 90°. It was shown that variation in the ratio of longitudinal to transverse dimensions changes the ground-state energy E0 and the first excited-state energy E1. The highest values, E0 = 1.7618 eV and E1 = 1.9528 eV, were obtained for conical nanocrystals, which may be associated with stronger spatial confinement of the electronic state due to the nonuniform transverse cross-section of the cone and the presence of an apex. It was also found that changing the ratio of longitudinal to transverse dimensions at a fixed volume leads to a restructuring of the optical density spectra. A pronounced angular anisotropy of absorption was revealed for cadmium telluride nanocrystals. The maximum positive value of the angular anisotropy coefficient was 16.12 % for spheroidal nanocrystals, whereas the most negative value was −10.59 % for cylindrical nanocrystals. In contrast to the conventional approach based on changing the overall size of nanocrystals, the proposed approach enables tuning of their electronic and optical properties by varying the ratio of longitudinal to transverse dimensions at a fixed volume. These results may be useful for tailoring the energy levels, absorption spectra, and angular sensitivity of cadmium telluride nanocrystals for the development of optoelectronic and photonic devices based on this material.
AUTOMATIC CONTROL AND ROBOTICS
Robust adaptive compensation of unknown multi-harmonic disturbances in multivariable linear systems (in English)
Cong Vinh Tu, Dudarenko Natalya Alexandrovna935
This paper proposes a new direct adaptive disturbance compensation method for multivariable linear systems, which is robust with respect to external disturbances. The external disturbances are assumed to be multi-harmonic signals (regular component) with additional random noise (irregular component), where the bias, phase, amplitude, and frequency are all unknown. The objective is to design a state-feedback control law capable of regulating the system output to zero in the absence of random noise. The robustness of the closed-loop system with respect to additive unmodeled disturbances is reflected in the fact that the proposed control law guarantees boundedness of all closed-loop signals. To solve this problem, first, the system inputs and outputs are decoupled using Falb–Wolovich approach. Then, the direct disturbance compensation method based on the internal model principle is applied to the resulting decoupled system. As a result, the adaptive algorithm can be constructed robustly and independently for each control channel. The adaptive algorithm with memory regressor extension is employed to improve the convergence rate of signals in the closed-loop system.
COMPUTER SCIENCE
Metabolic potentials method for adding loopless constraints in steady-state metabolic flux modeling using mixed-integer linear programming
Ivan A. Safonov, Sergushichev Alexey A.945
In modeling of steady-state metabolic fluxes in metabolic networks, Flux Balance Analysis (FBA) is widely used. However, the standard FBA formulation can admit biochemically infeasible solutions due to thermodynamically infeasible (futile) cycles. Enforcing thermodynamic feasibility with loopless constraints yields an NP-hard problem, which motivates the development of efficient Mixed-Integer Linear Programming (MILP) formulations of loopless FBA. In this work, we propose a new efficient metabolic potentials method for FBA with loopless constraints. This method is based on a novel MILP formulation of loopless constraints. Our analysis of the prior Fast-SNP method, which achieved an order-of-magnitude speedup, allowed us to identify and quantify the contributions of two key optimization ideas: using sparse constraints and reducing the number of integer variables. The proposed metabolic potentials method adopts these ideas and introduces an equivalent MILP formulation of loopless constraints. We compare the performance of the metabolic potentials method using metabolic models from the BiGG database. The proposed method shows performance comparable to Fast-SNP on small and medium scale metabolic models and achieves an approximately 30 % reduction in runtime on a large genome-scale human cell metabolic model RECON1. We also demonstrate that reducing the number of integer variables is the dominant factor in the overall speedup. The proposed metabolic potentials method retains the benefits of the optimization ideas introduced in Fast-SNP and achieves the same or better efficiency. Therefore, it can be used as a drop-in replacement for Fast-SNP and other standard methods for FBA with loopless constraints.
Hardware-Software system for crowdsourced monitoring and mapping of the geomagnetic field
Vorobeva Gulnara R. , Vorobev Andrei V. 954
This paper addresses the problem of conducting high-resolution geomagnetic surveys in urban environments using a crowdsourced approach based on mobile devices. Existing solutions, such as CrowdMag, are primarily focused on global data collection and do not provide capabilities for operational mapping of local technogenic anomalies. The question of the reliability of anomaly detection based on mass measurements performed by various non-professional devices under real-world conditions remains open. To solve these problems, a hardware-software system called Ground Monitoring has been developed. Its architecture is built on a three-tier model and includes a client mobile application for Android, a server part based on PostgreSQL/PostGIS, and a web-GIS with visualization and interpolation modules. The application records all three components of the magnetic induction vector at a frequency of up to 5 Hz, with a resolution of 0.25 μT and an error not exceeding ±5 %. Experimental validation was conducted in last August in the territory of Ufa with the participation of several hundred student volunteers. Over 50,000 spatially referenced measurements were collected, achieving a coverage density of up to 1500 points/km2. It was found that, despite the scatter of individual measurements, the average for the dataset differs from the reference observatory measurement by less than 0.15 %. Furthermore, it was revealed that 87 % of anomalies are localized near man-made objects. Statistical analysis also confirmed the high reliability of anomaly detection, and the results of spatial interpolation confirmed its robustness to noisy data. The obtained results confirmed the feasibility of using a crowdsourcing approach to solve geophysical mapping problems with high spatial detail and measurements consistent with reference observatory measurements. The developed system can serve as an experimental foundation for urban environment monitoring systems. Future work prospects involve integrating this approach into citizen science projects.
Diagnostics of inter-layer connectivity decay and minimal correction by residual connections in deep sequential networks
Bolshim Maksim A., Kugaevskikh Alexander V. 963
Deep sequential network design requires choosing the number and placement of residual connections, while architecture-search methods are resource-intensive and do not localize structural defects. The objective is to diagnose pre-training inter-layer coupling decay and use it for minimal residual-topology correction. The procedure analyzes off-diagonal Hessian blocks and introduces a global collapse index and a maximum local drop. A greedy algorithm places residual connections based on the profile without training or gradient-based architecture search. The method is evaluated on multilayer perceptrons with an artificial bottleneck on CIFAR-10. Sixteen configurations with 8, 12, 16, or 20 blocks and base width 64 or 128 were evaluated with three initializations each (48 runs), along with a 16-block architecture of base width 256, bottleneck width 16, and bottleneck position 8. The Spearman correlation between the maximum local drop and final accuracy is –0.692; the p-value is less than 0.001. After architectural adjustment, no significant association with accuracy is detected. With a fixed one-edge budget, guided placement yields higher accuracy than random placement; the p-value is 0.019 (exploratory analysis). Soft Jaccard coefficients range from 0.9583 to 0.9722. For adjacent target blocks, the 90 % confidence interval for the paired accuracy difference lies within the equivalence bounds; the larger p-value of the two one-sided tests is 0.014. The Hessian profile localizes a coupling drop and guides residual intervention rather than serving as an independent predictor of training quality. Conclusions obtained are limited to networks with an artificial bottleneck.
Methods for organising the data storage and transmission subsystem in a hardware accelerator for convolutional neural networks
Sergei M. Tabunshchik, Sergei V. Bykovskii, Ivan V. Lukashov 971
Modern Edge AI systems use specialized hardware neural network accelerators to process large amounts of data directly on user devices. The main challenges in organizing the data storage and transfer subsystem in hardware accelerators are limited bus bandwidth, limited memory capacity, and high requirements for processing speed and energy consumption. The growing demands on the performance and energy efficiency of neural network processors in the context of limited resources in embedded systems make it relevant to study the methods of organizing the data storage and transfer subsystem. The article proposes options for the architectural organization of the data storage and transfer subsystem: with full control from the central processor, with caching in BRAM-type memory, with direct access to memory through the common AXI_GP bus, and with a dedicated high-speed AXI_HP bus. To evaluate the efficiency, the article uses the method of simulating the inference of the convolutional layer of the YOLOv5s neural network on a neural network processor model at the register transfer level, calculating the bandwidth and performance for each option. As a result of the study of the dependence of the characteristics of the neural network processor on the method of organizing the data storage and transmission subsystem, it was found that the developed neural network processor achieves a maximum performance of 240 GOPS. The advantage of using a direct access interface to RAM with a bandwidth of 1.2 GB/s, which is dedicated for data transmission for calculations, has been confirmed. The comparison showed that the use of direct memory access with a dedicated high-speed interface can increase the performance of inference of neural networks by an average of 20 times. The described methods for organizing the data storage and transfer subsystem achieve a performance that is 1.19 times higher than that of analogues on convolutional neural network inference tasks. The results demonstrate the potential for using these methods in the design of hardware accelerators for Edge AI systems.
A technique and dataset of HTTP traffic for web application classification in business logic vulnerability detection
Aleksei A. Kosmachev, Kotenko Igor V. 981
This article explores creating a specialized HyperText Transfer Protocol (HTTP) traffic dataset for multi-class web application classification by topic (domain) to identify business logic vulnerabilities. Each record contains the full text of the HTTP request and response, including headers, body, path, and parameters, which distinguishes the dataset from existing analogs limited to either a single category or a unique resource address on the Internet (Uniform Resource Locator) without any content, and is the first publicly available HTTP traffic dataset with multi-domain markup by categories relevant to the tasks of detecting business logic vulnerabilities. A technique for automated web application scanning is proposed, comprising five sequential stages: defining the composition of subject categories, selecting auxiliary sources and a collection method, preparing an initial list of web resources, multi-level web crawling with preservation of the full content of requests and responses and simulating user actions, as well as post-processing and filtering of the collected data. A comparative study of five classification models was conducted on the collected data: three models based on Term Frequency — Inverse Document Frequency (logistic regression, linear support vector machine, and naive Bayes classifier) and two neural network architectures (the lightweight and fast FastText model and a convolutional neural network adapted for the texts based on Kim’s TextCNN architecture). Web application crawling resulted in a dataset of 53,074 records in Parquet format, covering four categories: e-commerce, games, finance, and social networking and messaging. The best result of model comparison was achieved by the TextCNN model. The dataset and technique are published openly. The limitations of the proposed approach to creating a dataset of HTTP traffic for the task of multi-class classification of web applications by topic are the limitation of the coverage of the dataset to four topic domains, the minimal thematically significant content of publicly accessible pages of applications in the category of social networks and instant messengers, and the introduction of translation artifacts into the feature space of about 30 % of the records.
Search for promising points of neural networks initialization
Tolstykh Victor N. , Vladimir N. Yershov 993
The paper examines the problem of initialization, which consists in determining the starting point in the space of the neural network weights for a qualitative gradient descent. It is known that the learning speed and learning result of a neural network mostly depends on its initialization. The problem is that in the process of execution, the algorithm can get into the anomalous zone of the local minimum and get stuck there for a long time, sometimes at the very beginning of training. In such cases, it is recommended to start the descent with a new initialization. The study proposes to start training from a point at the bottom of the local minimum, from which, as research shows, the descent in a new direction begins. The proposed method is based on the assumption that most of these anomalies are critical points that have a direction to descent. The neural network is initialized every time the learning algorithm encounters an anomaly in which the value of the error function increases. At the same time, a new initialization is determined at the minimum point of the error function in the segment connecting the point before the anomaly and the following one. Successive initializations result in the descent occurring along an optimized trajectory without significant oscillations. The effectiveness of the method is demonstrated by the example of a perceptron with a single hidden layer and one anomaly. It is proved that the descent from the new initialization point is faster; this demonstrates the possibility of gradient correction without calculating the Hessian; a conclusion is made about the expediency of multiple initializations of the neural network when it enters the zone of the next anomaly. It is proposed to use the method of correcting the gradient descent algorithm to reach the bottom of the anomaly in cases where the error begins to grow. The developed method gives a high probability of hitting a region with a low value of the error function than in the case of a random drawing of the initialization point, as well as a benefit in comparison with a Restricted Boltzmann’s machine in terms of execution time.
Training visual language models for MRI-based diagnostics
Gusarova Natalya Fedorovna, Dobrenko Natalia V. , Zubanenko Aleksei A. , Ivan V. Tomilov, Pavel G. Ulyanov 1000
In modern radiology, diagnostic Artificial Intelligence assistants for classifying Magnetic Resonance Imaging (MRI) brain images are typically based on Convolutional Neural Networks (CNNs), which work exclusively with the visual signal and are unable to use structured diagnostic descriptions of pathologies. Vision Language Models (VLMs) open up new possibilities: joint processing of visual and textual data allows for the consideration of the semantic context of pathologies, which is inaccessible to CNN classifiers. The key obstacle to their application remains the high resource intensity of retraining, which is incomparable with the capabilities of a typical clinic. This paper proposes a method for constructing a resource-efficient diagnostic AI assistant based on VLM Gemma3 (4b-it) using low-rank adaptation (LoRA). Retraining was performed on a dataset of 91,777 Fluid-Attenuated Inversion Recovery MRI brain images, distributed across eight pathology classes. Only low-rank adapters were trained — 12.5 % of the model parameters, which allowed training on a single NVIDIA RTX3090 Graphics Processing Unit with 24 GB of Video Random Access Memory without masked data annotation. The radiologist’s expert knowledge was integrated via a system prompt, which assigns the model the role of an experienced neuroradiologist with explicit analysis criteria. On the validation set (9178 examples), the model achieved: Accuracy 0.930, F1-score 0.929, Specificity 0.989, and Cohen’s Kappa 0.915. Comparison with the baseline ConvNeXt-Base classifier on an identical dataset showed the superiority of VLM+LoRA across all metrics: Accuracy increased by 7.4 percentage points, and Cohen’s Kappa by 0.090. The obtained results confirm the practical applicability of the proposed approach for clinics with limited computing resources.
Integral quality assessment of machine learning models for cyberattack detection in network traffic
Aleksey A. Lukoyanov, Evgeniy A. Rogozin, Ruslan E. Rogozin, Ilya I. Livshitz 1009
Machine learning models are widely used to detect cyberattacks through the analysis of network traffic. Traditional performance metrics do not account for the temporal characteristics of the data and may mask performance degradation within individual time intervals or under varying traffic intensity. Approaches based on key performance indicators, weighted geometric aggregation, and Bayesian methods allow individual performance characteristics to be combined; however, by themselves, they do not account for changes in model performance across predefined operating states. The developed integrated assessment methodology takes into account the frequency and certainty of erroneous decisions, as well as the degradation of model quality in specific operating states. A partial model quality indicator is calculated for each combination of time interval, target recall level, and traffic intensity mode. The indicator is formed by multiplicatively combining the classification score, the average calibrated confidence in correct decisions, the loss from erroneous decisions, and a time multiplier. The integrated indicator is obtained by averaging the partial indicators proportionally to the number of observations and subsequently adjusting for deviations below the weighted average. Experimental samples were generated using a single chronological data partition, eliminating the use of test records during model training, probabilistic calibration, and selection of classification thresholds. The methodology was evaluated on GothamDataset2025 using a test set of 1,777,237 network traffic records. Logistic regression, random forest, and histogram-based gradient boosting were compared across 18 operating states. At thresholds selected to provide validation recall of at least 0.50, random forest achieved the highest balanced accuracy of 0.693933 and the Matthews correlation coefficient of 0.466685. Histogram-based gradient boosting obtained the highest integral quality score of 0.264163. The methodology identified states with reduced or zero state-specific indicators that were not revealed by balanced accuracy and the Matthews correlation coefficient calculated over the entire test set, and changed the final model ranking. Operating states and aggregation parameters must be defined before assessment. The methodology supports the selection of cyberattack detection models under varying network traffic conditions.
Multi-objective optimization of software development processes using large language models as assistants
Petr V. Gorshkov, Bessmertny Igor Alexandrovich 1018
This article examines the problem of managing the use of large language models based on transformer architectures in software development processes. The relevance of the study is that the implementation of Large Language Models (LLMs) in software development is often accompanied by inflated expectations: the expected reduction in development time is not always achieved, while the risks of reduced quality, increased costs, vulnerabilities, and the need for additional verification of the result remain. The subject of this study is the use of LLMs in software development as a managed mode of project work. The scientific novelty of the proposed approach lies in presenting the use of LLMs not as a standalone code generation tool, but as a multi-objective model for selecting a rational development mode. The aim of this study is to develop a multi-objective approach to managing the use of LLMs in software development, taking into account time, cost, result quality, and risks. The methodological basis includes the analysis and systematization of scientific publications, the formalization of factors influencing the effectiveness of LLM use, multi-objective modeling, and the Pareto optimality principle. The study identifies the following controllable parameters of LLM application scenarios: the level of task decomposition, the degree of elaboration of project documentation, the quality of the problem statement and prompt, and the selected language model or a tool based on it. In addition, contextual parameters are identified: developer or team qualification, programming language, and technology stack. The identified parameters are shown to affect the time required for documentation preparation, result generation, code review and correction, as well as the cost, quality, and risks of the project. The experimental part uses scenario modeling of alternative LLM application modes, which allows comparing options according to time, cost, quality, and risk criteria. The proposed model enables the selection of rational LLM application scenarios on the Pareto set, taking into account the constraints of a specific IT project. Unlike existing approaches mainly focusing on evaluating individual AI tools, generation quality, or developer productivity, the proposed model considers LLMs as an element of a managed software development process. The practical application of the approach involves adapting IT project management tools: task visualization and routing tools, prompt-oriented documentation templates, and a registry of informal notes. This reduces the variability of generated results and improves the manageability of LLM use in software development.
Automatic formation of observation objects for system state assessment from time series
Lebedev Ilya S., Alexander A. Makeenko 1026
Improving classification quality metrics requires advances not only in data processing models but also in the related processes of dataset formation and training. The implementation of information sequence processing methods depends on data representation and requires an appropriate choice of model and tuning of its hyperparameters. The final result is significantly influenced by the way the original time series are transformed into observation objects, since this transformation determines the properties of the training dataset and the conditions for the subsequent application of classification algorithms. Therefore, the problem of observation object formation can be considered as an independent stage of data preparation affecting the quality of system state assessment. This paper considers the automatic formation of observation objects for system state assessment based on time series and information sequences. Unlike classical approaches, where the model is adapted to the data, the proposed method is based on solving the inverse problem, in which datasets of observation objects are formed from time series and information sequences so that a predefined model achieves the best value of a selected quality metric. Experiments were carried out using both model data and benchmark datasets. By varying the window width and shift, datasets of observation objects with different properties were generated from the time series. The models were trained on these datasets, and the characteristics of the observation objects and datasets yielding the best results were identified. Quality metrics were obtained and analysed. The results show that forming observation objects from sequences with specific window length and shift properties allows improving precision by 3–8 % compared with the average values obtained over the corresponding datasets. The proposed solution can be integrated into the development of new data processing methods and used to enhance existing models. The formation of datasets from observation objects with different properties extends the possibilities of model training and provides an additional mechanism for improving classification quality.
Automatic recognition method for communicative gestures of virtual interaction participants
Anastasia A. Dvoynikova, Karpov Alexey A 1039
Currently, interpersonal communication increasingly takes place in virtual environments through teleconferencing systems. Despite the advantages of this format, limited video window size, unstable Internet connections, and reduced transmission quality hinder the perception of nonverbal signals exchanged by participants. During interpersonal interaction, communicative gestures serve as an important means of conveying emotions and engagement, while their automatic analysis can improve the accuracy of systems designed to recognize behavioral and psychological states of participants in virtual communication. This paper proposes a hierarchical method for communicative gesture recognition in virtual communication environments, enabling the joint recognition of both static and dynamic gestures. The proposed method is based on the sequential application of models for gesture type recognition followed by the classification of static and dynamic gestures. At the first stage, gesture type is determined as static, dynamic, or no gesture. At the second stage, specialized models are employed to recognize three classes of static gestures and three classes of dynamic gestures. Various feature types are investigated for video analysis, including body keypoints, their geometric derivative features, optical flow, and features extracted from a pretrained Vision Transformer (ViT) model. Fully connected, convolutional-recurrent, and transformer-based neural networks are considered as classifiers. In addition, methods for improving classification accuracy under class imbalance and limited data conditions are investigated. Experimental studies were conducted using the multimodal ENERGI dataset containing recordings of group virtual communication annotated with communicative gestures. The best performance for static gesture recognition was achieved using ViT features and a convolutional-recurrent neural network, resulting in an average F1-score of 70.4 %. For dynamic gesture recognition, the most effective combination consisted of geometric derivative features and a fully connected neural network, achieving an average F1-score of 58.1 %. For gesture type recognition, the joint use of ViT features and optical flow yielded the highest average F1-score of 76.5 %. The overall average F1-score for recognizing seven communicative gesture classes reached 68.7 %, whereas the single-model approach achieved only 20.8 %. The results also demonstrate that applying class imbalance compensation techniques further improves classification performance. The results obtained confirm the effectiveness of the hierarchical approach compared to single-model classification due to the decomposition of the task into gesture type recognition and subsequent specialized classification. Different feature types were found to be most effective for different gesture categories: ViT features better capture static pose configurations, whereas geometric derivative features provide more informative representations of motion patterns. The proposed method can be applied in intelligent systems for behavioral analysis of participants in virtual communication, including engagement assessment, emotion recognition, and interaction effectiveness evaluation.
MODELING AND SIMULATION
Synthetic dataset generation for training object detectors for underwater environment (in English)
Bassel Alshawareb, Kolyubin Sergey A 1049
Adapting object detection models to underwater environment presents significant challenges due to underwater light attenuation and limited data availability. This work proposes a novel approach to address the challenge of task-oriented underwater datasets scarcity by generating prompt-guided, realistic synthetic images, and subsequently fine-tuning common object detectors on this data. The proposed pipeline begins by fine-tuning a text-to-image generation model on a small dataset of real underwater images using the Low-Rank Adaptation approach. The model generates a larger set of diverse synthetic underwater images aligned with given textual prompts. To ensure that the generated images are accurate, a verification module is introduced, which filters out misaligned samples based on cosine similarity between prompt and caption embeddings of the generated image. Segmentation masks are automatically extracted from the generated images using a salient instance segmentation model. Finally, we augment real underwater images with these synthetic image-mask pairs and fine-tune a detector model on an extended dataset. Experimental results on a subset of under 800 images from the USIS10K benchmark demonstrate a 2.2 % increase in mean Average Precision (mAP) with a moderate amount of synthetic data compared to the real-only baseline. Although synthetic images are generated for only the human divers and ruins classes, detection improves also for non-augmented classes. We further experimented with a scarcer data setting by reducing the training images of the two target classes from 345 to 180. In this setting, augmenting with our synthetic images showed a larger improvement of 3.3 % in mAP, underlining the usefulness of synthetic data when real training data are limited.
Operability of a low-temperature rectification scheme under variations in associated petroleum gas composition (in English)
Vadim O. Kholyavkin, Ekaterina V. Logvinenko 1057
The composition of Associated Petroleum Gas (APG) varies between fields and during field development, which affects phase behavior and component distribution in Low-Temperature Rectification (LTR) units. Because industrial pressure and temperature set-points are fixed or adjustable only within narrow limits, the relevant task is to determine whether an existing scheme remains operable without retuning when feed composition changes. A steady-state Aspen HYSYS model was developed for a mixture containing CH4, C2H6, C3H8, i-C4H10, n-C4H10, and a lumped C5+ fraction. Five reference APG compositions were simulated at identical set-points. For each component i, recovery to the bottoms, Ri, was calculated as the component mass flow in the bottoms divided by its total mass flow in the overhead and bottoms products. Recoveries for 100 realistic compositions were estimated by inverse-distance weighting in composition space. A case was classified as operable when methane recovery to the bottoms did not exceed 0.05, propane recovery was at least 0.905, and recoveries of i-butane, n-butane, and C5+ were at least 0.95. Under the selected fixed set-points, 85 of 100 compositions met all criteria, giving an operability index F = 0.85. The PASS/FAIL map showed that propane recovery is the limiting criterion near the operability boundary, whereas butanes and C5+ retain larger margins. The proposed method identifies the feed-composition domain in which an LTR scheme remains operable at fixed set-points while reducing the number of full process simulations. Its validity is limited to the composition domain covered by the reference cases; therefore, borderline and extrapolative cases require targeted simulations.
Vision transformer for reference-based super-resolution
Alexander V. Yakovenko, Aleksei K. Denisov, Dmitriy S. Vatolin, Pavel V. Kustarev 1065
Reference-based Super-Resolution (RefSR) aims to restore a low-resolution image by transferring real textures from a high-resolution reference of similar content, rather than hallucinating missing details as in single-image super-resolution. Most existing RefSR methods decompose the problem into separate matching, alignment, and fusion stages, implemented as sequentially trained modules. This decomposition causes errors to accumulate across stages and imposes inductive biases that limit what the model can learn from data. Instead, we propose to handle correspondence and texture transfer jointly through a single attention operation. We propose an end-to-end transformer that operates on the concatenated sequence of input and reference tokens, with self-attention implicitly performing both matching and texture transfer. The architecture introduces binocular 2D Rotary Positional Embedding, a shared coordinate frame for input and reference patches, and a resolution-adaptive attention scaling scheme in which the logits are rescaled by the length of the processed sequence, allowing the model to operate at resolutions beyond those seen at training time. On the standard CUFED5 and WR-SR benchmarks under bicubic ×4 downsampling, our method improves over the best previously published RefSR approach by 0.33 dB Peak Signal-to-Noise Ratio (PSNR) on CUFED5 and, for the first time, surpasses strong single-image super-resolution models on WR-SR by 0.29 dB PSNR. Unlike multi-stage pipelines, the proposed model is end-to-end trainable and requires no pretrained components, showing that a single attention operation over the joint token sequence is sufficient for RefSR. The approach is broadly applicable to image restoration settings where a high-resolution reference of similar content is available.
BRIEF PAPERS
Single-channel speech source separation for speech recognition and speaker verification
Anastasia A. Zorkina, Nikita A. Khmelev 1073
A method for single-channel separation of two speech sources based on conditional generative modeling is proposed. The task is characterized by uncertainty in the order of the two reconstructed speech signals: the same pair of separated sources can be represented in two permutations. To resolve this ambiguity, a frozen speaker encoder is used: it defines the source order during training, selects the best variant among N stochastic generations, and aligns the sources during chunk-wise processing of recordings of arbitrary length. On the Libri2Mix dataset, the proposed Transformer U-Net-based variant is competitive in terms of fundamental metrics for evaluating the quality and intelligibility of processed speech and achieves the best results when used in speech recognition and speaker verification tasks.
1077
We address voice-based speaker search in a large gallery of voices. Modern models pretrained with self-supervised learning, such as wav2vec 2.0, provide high search accuracy but require substantial computation for extracting speaker embeddings. To reduce this cost, we propose a cascaded method: the gallery is enrolled using a fast student model, dWhisper-ASV, while queries and candidate verification are performed by a teacher model, wav2vec 2.0 XLSR-1B. The alignment of embedding spaces enables cross-model cosine similarity. On ClovaCall, Speech-MASSIVE, and a proprietary Russian corpus, RusIVR_test, the cascade nearly matches teacher-only F1-score: 77.21 % vs. 77.24 %, 83.05 % vs. 83.16 %, and 92.26 % vs. 92.63 %. Only 1.5 %, 0.2 %, and 0.11 % of records reach the second stage and the enrollment speeds up eightfold.
Constructing the multimodal vector space of the accessibility graph for non-visual information access
Alexey V. Degtyar 1080
A multimodal vector space robust to the unavailability of individual perception channels — characteristic for blind and visually impaired users — is constructed. The proposed method represents nodes in a common vector space obtained by learnable adapter projections of four channel representations (text, image, program code, and optical character recognition), their weighted fusion with the availability mask, and enrichment via attention over the graph of nearest neighbours. Training uses a joint loss with classification, consistency and contrastive components. On 2,864 arXiv-CS papers labelled by 10 Computer Science Ontology branches and under text channel unavailability, the method reaches 0.790 accuracy against 0.498 for the baseline and 0.560 for the no-enrichment version.

