Computer systems and information technologies

INFORMATION AND LASER TECHNOLOGIES FOR ASSESSING THE LEVEL OF RISKS FROM HARMFUL EMISSIONS FROM MAN-MADE OBJECTS

Liubomyr SIKORA — 2025-03-27

The current stage of development of electricity, chemical, machine-building, and printing production is characterized by the use of a wide range of resource components - coal, oil, gas, paints, and polymers - that are environmentally aggressive. The intense production regimes which are dictated by the market lead to a sharp increase in resource consumption for energy-intensive production processes, causing, in turn, an increase in the concentration of dust and harmful gases and liquid emissions into the atmosphere and water environment, which leads to an increase in environmental pollution, the state of which cannot always be assessed in real time due to the complexity of data collection using standard methods.

The article substantiates the methods of creating sensors for measuring the concentration of dust and harmful substances emissions into the atmosphere and water environments using new physical effects that became the basis for the development of laser concentrators of air and water pollution, optogalvanic effects for creating integrated sensors that can be combined with measurement systems based on ion-selective sensors (OCS 5M), which makes it possible to increase the level of efficiency of environmental safety systems. A comprehensive solution to the problem is based on the creation of global environmental monitoring systems based on information and intellectual technologies and the development of new sensor models. The problem of environmental monitoring has been relevant for more than a century, because of the development of industrial technologies (railroads, weaving) has brought not only prosperity but also environ-mental pollution. The level of environmental pollution has increased especially with the development of thermal power plants and the petrochemical complex, which have become aggressive polluters. The military operations of the First and Second World Wars also contributed to this. Nuclear power and jet aviation have further polluted the global environment, and the war in Ukraine has created a specific environmental impact (explosions of shells and missiles), destruction of energy complexes and oil terminals.

HEATING OPTIMIZATION SYSTEM IN A SMART HOME BASED ON FUZZY LOGIC AND INTEGRATION WITH CLOUD SERVICES

Ihor LYTVINCHUK — 2025-03-27

This paper proposes a heating optimization system in a smart home based on fuzzy logic and integration with cloud services. According to the results obtained, the use of fuzzy logic significantly improves the stability of the temperature in the house, which is important for the comfort of residents. For the experiments, two models were compared - the basic heating model and the model based on fuzzy logic. The basic system, which does not take into account variable factors with such a level of flexibility, leads to large and sharp temperature fluctuations, which can create discomfort and increase energy consumption. In contrast, the model with fuzzy logic demonstrates a smoother and more stable temperature control, which allows you to significantly reduce energy costs.

MOBILE-ORIENTED CYBER-PHYSICAL SYSTEM FOR FOOD ALLERGEN DETECTION BASED ON MACHINE LEARNING AND IMAGE ANALYSIS

Valentyn TALAPCHUK — 2025-03-27

The prevalence of food allergies necessitates the development of effective methods for the timely detection of allergenic components in food products to prevent dangerous medical reactions. In this work, a mobile-oriented cyber-physical system is proposed, leveraging state-of-the-art machine learning techniques and image analysis for the automated detection of food allergens. The developed system integrates the capabilities of mobile devices equipped with high-quality cameras and efficient computational resources, enabling accurate processing and classification of food product images either locally or via cloud-based inference. This approach ensures flexibility in deployment while maintaining high detection accuracy across diverse environments.

This study examines both the theoretical and practical aspects of applying deep neural networks to object recognition tasks. Particular emphasis is placed on the EfficientDet model, which, due to its optimal balance between detection accuracy and computational cost, represents a promising solution for mobile applications. To enhance recognition performance, image pre-processing methods—including normalization, scaling, and data augmentation—are employed to increase the model’s resilience to variations in imaging conditions.

The methodology for data collection and image annotation is described in detail, including the pre-processing procedures that ensure improved model robustness under diverse external conditions. Experimental investigations conducted on a large annotated dataset demonstrate the high accuracy and effectiveness of the system in detecting the presence of food allergens, thereby enabling the prompt identification of potentially hazardous components.

The results of the work highlight the practical applicability of the proposed system in mobile applications for monitoring food quality and preventing allergic reactions. The conclusions outline prospects for further research, focusing on expanding the platform’s functional capabilities through the integration of additional sensor technologies and the refinement of data processing algorithms.

METHOD OF FPV DRONE STABILIZATION ON AN AUTOMATICALLY DETERMINED TARGET AND ITS FURTHER OBSERVATION

Oleksandr HALYTSKYI — 2025-03-27

The paper considers the problem of the lack of stabilization in FPV drones, which significantly limits their functionality for applications that require precise tracking of a specific target. Such drones, although characterized by high maneuverability and affordable cost, are inferior to commercial quadcopters such as the DJI Mavic, which are equipped with effective stabilization systems, but are significantly more expensive due to the use of proprietary technologies. The paper proposes a new approach to stabilizing FPV drones, based on the use of computer vision algorithms for automatic target detection and tracking. The main concept includes target detection based on image analysis from the drone camera, further determination of its trajectory and transmission of appropriate control commands to the flight controller using the MAVLink protocol. This approach allows to significantly increase the accuracy and stability of FPV drones when performing tasks that require focusing on an object, such as infrastructure inspection, search and rescue operations, or video shooting. The proposed solution is based on accessible and open technologies, which ensures its adaptability and low implementation cost. The paper describes in detail the developed system architecture, which includes a computer vision module for video stream analysis, algorithms for data processing and filtering, as well as integration mechanisms with existing flight controllers. A series of experiments were conducted to evaluate the effectiveness of the proposed approach. The results demonstrate that the proposed system is able to ensure stable drone tracking on a specified target even in difficult flight conditions.

PREDICTION OF ALZHEIMER'S DISEASE USING BAYESIAN NEURAL NETWORKS

Serhii HLADIHOLOV — 2025-03-27

This article presents a methodology for optimizing Bayesian neural networks and their application to complex prediction tasks, with a focus on diagnosing Alzheimer’s disease. Alzheimer’s is a neurodegenerative condition where early detection is vital for initiating timely interventions and improving patient outcomes. The proposed methodology includes determining the optimal structure of classical neural networks by performing grid search to identify the best combination of layers and neurons. The architecture identified through cross-validation forms the basis for constructing Bayesian neural networks, where weights derived from classical models are utilized as prior distributions. This integration improves prediction accuracy while preserving the Bayesian network’s capacity for quantifying uncertainty.

Bayesian models are trained using Markov Chain Monte Carlo methods, with experiments exploring the impact of prior distribution parameters, including variations in means and standard deviations. Results show that a mean value of zero and a standard deviation of 2.5 yield optimal outcomes, minimizing classification error while balancing uncertainty estimation. Increasing the standard deviation improved performance up to a threshold, beyond which further gains were statistically insignificant. The ability of Bayesian neural networks to incorporate uncertainty provides critical advantages for decision-making in medical contexts, particularly in scenarios involving incomplete or noisy data.

The findings demonstrate that Bayesian neural networks based on optimized classical architectures can effectively address prediction tasks in high-stakes domains like medicine. By leveraging prior knowledge, the proposed approach reduces training time and enhances model performance, offering a robust framework for diagnosing Alzheimer’s disease. Future research will explore automating structural optimization, assessing the impact of different prior distributions, and extending this methodology to other neurodegenerative disorders.

METHOD OF OPERATION OF THE CYBER-PHYSICAL WATER RESOURCES MONITORING SYSTEM

Yurii VOICHUR — 2025-03-27

The relevance of designing and developing a cyber-physical water monitoring system for Ukraine is driven by the need for effective water management in the face of climate change, water pollution, and growing water supply needs. Modern challenges, such as the lack of clean drinking water, irrational use of resources, emergency condition of water supply networks and environmental threats, require the introduction of innovative technologies. The use of sensor networks, artificial intelligence, and cloud computing allows us to quickly obtain information about water quality and quantity, predict changes, and prevent emergencies. The introduction of cyber-physical systems in the field of water resources monitoring will help to increase the efficiency of water management, reduce losses, improve the ecological condition of water bodies and provide the population with quality water. For Ukraine, where water security is a strategic issue, such solutions will be an important step towards sustainable development and environmental balance. The use of Internet of Things (IoT), Big Data, and artificial intelligence technologies can automate the processes of data collection, analysis, and forecasting, which will help optimize water use, prevent pollution, and increase the efficiency of water infrastructures. Thus, the task of designing and developing a cyber-physical water resources monitoring system is currently relevant for Ukraine.

The article develops a method for the operation of a cyber-physical water resources monitoring system that provides cyber-physical integration (a combination of physical (sensors, objects) and cybernetic (analytics, control) components), autonomy (the ability to function without constant human intervention), scalability (the ability to expand the geography of monitoring), and monitoring continuity (round-the-clock real-time monitoring).

DECISION-MAKING METHOD IN INTERDEPENDENT COMPUTING SYSTEMS

Dmytro KRYZHANYVSKYI — 2025-03-27

In this work, a decision-making method for interdependent computational systems has been developed. The proposed approach integrates Bayesian reputation updates, log-linear strategy selection, and reinforcement learning mechanisms to enable autonomous agents to make context-aware and reliable decisions. The method effectively balances strategic adaptability, system stability, and robustness to unreliable or malicious agents.

A distinctive feature of the method is its ability to dynamically adjust agent strategies based on reputation scores and prior interaction outcomes, thus facilitating convergence toward Bayesian-Nash equilibrium. The implementation includes a mechanism for iterative reputation correction and probabilistic strategy optimization, which ensures that the system achieves stable coordination in a decentralized environment.

Simulation results demonstrate that the proposed method significantly improves the convergence rate, reduces the impact of low-reputation agents, and enhances system-wide cooperation. Increasing the number of agents leads to moderate growth in system complexity, but the reputation-aware mechanism effectively mitigates instability and delays in strategy synchronization. Conversely, adding more interaction rounds improves reliability and accelerates equilibrium attainment.

Future research directions include adapting the model for various types of interdependent systems, such as edge computing environments, IoT infrastructures, and mobile multi-agent platforms. It will also be necessary to explore multi-objective optimization formulations that incorporate not only performance and stability, but also energy efficiency, communication overhead, and quality of service (QoS) constraints.

CHAT GPT FOR NETWORK ANALYSIS CRIMINAL CO-OFFENDERS

Olha KOVALCHUK — 2025-03-27

In the complex structure of global society, crime remains a persistent problem. It significantly threatens community foundations and hinders social and economic progress. Today, artificial intelligence (AI) technologies are actively being developed to predict possible offenses and detect and analyze criminal network structures through criminal data analysis. The article presents a new approach to studying social connections in criminal networks using GPT-4 tools. A methodology for visualizing criminal data in graphs has been developed to identify criminal group structures. Visual models of criminal co-offender networks were created using data from 2,113 criminal proceedings involving vehicle theft, robberies, and armed robberies committed in the Ternopil region between 2013 and 2024. Using the GPT-4 multimodal model, data processing was performed and graphs were constructed that reflect the structure of social connections between criminals. The analysis revealed significant differences in the structure of criminal interactions for different types of crimes: vehicle theft shows complex interconnected networks with a high degree of centralization and the presence of key coordinator figures; robberies are dominated by small stable groups of 2-3 people, which is explained by the specifics of executing these crimes; armed robberies are characterized by the formation of larger (4-6 people) and structured criminal groups with defined role distribution, due to the need for violence and ensuring control over victims. The proposed methodology effectively allows law enforcement agencies to counter organized crime in modern conditions. The obtained results have practical value for law enforcement agencies in making operational and strategic decisions, as they allow for the identification of key participants in criminal networks and the prediction of their potential criminal activities.

SCHEDULING: A FAST C++ THREAD POOL IMPLEMENTATION CAPABLE OF EXECUTING TASK GRAPHS

Dmytro PUYDA — 2025-03-27

In this paper, the author presents a simple and efficient C++ thread pool implementation capable of executing task graphs. The conducted experiments demonstrate that the proposed solution achieves CPU performance comparable to Taskflow, a highly optimized library for parallel and heterogeneous programming. The implementation is small and straightforward, consisting of less than one thousand lines of C++ code at the time of writing this paper.

ANALYSIS OF STUDENT PERFORMANCE BASED ON CANONICAL CORRELATION ANALYSIS

Kateryna BEREZKA — 2025-03-27

The article examines the application of Canonical Correlation Analysis (CCA) to investigate the relationships between student performance outcomes across different groups of disciplines. The disciplines were categorized into the following groups: mathematics, programming and algorithms, systems design, networks and distributed systems, applied software and technologies, and economic and managerial disciplines. The study aims to identify dependencies between these discipline groups that influence overall academic performance.

The analysis revealed that discrete mathematics plays a key role in shaping programming skills, with performance in mathematical disciplines significantly correlating with outcomes in other fields. Both strong and weak correlations were identified between specific discipline groups. The use of CCA provided deeper insights into the relationships between subjects, offering new opportunities for optimizing the educational process.

The findings of the article have both theoretical and practical significance, contributing to the improvement of educational approaches and methods for assessing academic performance.

METHODS OF ELECTROCARDIOGRAM CLASSIFICATION AND THEIR MATHEMATICAL MODEL IN THE FORM OF A CYCLIC DISCRETE RANDOM PROCESS

Lyubomyr MOSIY — 2025-03-27

This paper presents an advanced approach to modeling electrocardiogram signals by integrating amplitude-time characteristics to obtain novel and informative features for cardiac diagnostics. Based on a systematic analysis of 426 scientific publications from the Scopus database (2014-2024), we identified a significant transformation in methodological approaches from classical signal processing to the implementation of modern artificial intelligence technologies. A geographical analysis of publications revealed that India, the United States, and Germany led the research in this field, with 78, 64, and 37 publications, respectively. The thematic distribution of works encompasses computer science (23.3%), engineering (22.4%), medicine (13.8%), and related fields, highlighting the interdisciplinary nature of these studies. We identified key developmental directions in electrocardiogram signal processing methods, including the improvement of filtering algorithms and data preprocessing, the development of new methods for extracting informative features, and the creation of hybrid classification systems. Particular attention was paid to integrating machine learning methods with traditional approaches to electrocardiogram signal analysis. The research demonstrated that while convolutional neural networks exhibit high classification accuracy (>95%) for cardiac arrhythmias, there remains a need for mathematical models that account for both rhythmic and morphological features of ECS signals. We propose a model of cyclic discrete random process with a time rhythm function that incorporates amplitude values of characteristic ECS peaks (P, Q, R, S, T). This model effectively captures the inherent cyclicity of ECS signals while accounting for their stochastic variations and corresponding amplitude values of diagnostic waves. The model distinguishes between regular and irregular cardiac rhythms. Experimental validation using ECS signals from healthy individuals and patients with extrasystole demonstrates the model's sensitivity to changes in cardiovascular system states. The time rhythm function, considering amplitude, exhibits distinctive patterns that effectively differentiate between normal and pathological conditions. The proposed mathematical framework expands the analytical toolset for ECS signal processing and provides a foundation for developing new diagnostic algorithms with enhanced accuracy for cardiac rhythm disorders.

SCALABLE PARALLEL TRAINING OF DEEP NEURAL NETWORKS FOR IMAGE CLASSIFICATION USING TENSOR PROCESSING UNITS

Lesia MOCHURAD — 2025-03-27

Image classification using machine learning techniques is crucial in fields such as medicine, ecology, and agriculture, where large datasets of images need to be processed efficiently. However, traditional deep learning methods can be computationally expensive, particularly when handling massive amounts of data. This paper proposes a novel parallel training approach for deep neural networks using multiple Tensor Processing Units (TPUs) with the TensorFlow tf.distribute.Strategy API, aimed at solving the scalability issue in image including bird species classification tasks. The primary advantage of this approach is the ability to parallelize the training process without altering the model architecture, ensuring both flexibility and efficiency. By distributing the workload across multiple TPUs, the algorithm accelerates training significantly, enabling faster model convergence. Numerical experiments comparing the proposed parallel training method on 8 TPUs with a traditional sequential approach on a single Graphics Processing Unit (GPU) show that parallel training reduces training time by a factor of 4.6 while maintaining the classification accuracy achieved in sequential training. This demonstrates that the parallelized method not only speeds up the process but also retains model performance. The proposed algorithm has shown high scalability, making it suitable for processing large datasets. This scalability is particularly beneficial for tasks requiring rapid processing of large volumes of image data, such as real-time applications in environmental monitoring or wildlife research. In conclusion, parallel machine learning methods present a promising solution for improving the speed and efficiency of image classification tasks. Future research can focus on further optimizing the scalability of this approach and enhancing its performance for even larger datasets, as well as its application in time-sensitive real-world scenarios.

A NEW APPROACH FOR CREATING CHATBOTS BASED ON THE USE OF FINITE AUTOMATA THEORY

Yevhenii VOVK — 2025-03-27

Nowadays, the era of waiting in lines, writing official letters, and direct contact with employees of institutions, establishments, and companies is gradually becoming a thing of the past. Instead, the problem of creating tools that ensure the development, implementation, and implementation of chatbots and agents, their support, and expansion of functionality, and scalability, arises.

The main subject of this article is precisely the representation of a chatbot in the form of a state diagram. This technology, together with the technology of analysis and synthesis of formal chatbot models, constitute important components of the platform and information systems as a whole for institutions, establishments, and companies of various levels. An analysis of the possibilities of automata theory has shown the feasibility of using transitional systems and finite automata such as X-automata, Mealy and Moore automata as chatbot models.

The article describes a general approach to the effective representation of a chatbot in the form of a state diagram, implemented within the framework of a platform for the development, accumulation, and use of chatbots. As a formal model of a chatbot, it is proposed to use finite automata of Mealy and Moore, and the transformation of a regular expression, which is based on the input and output alphabets of the system, to a certain graphic configuration is proposed to be carried out according to the algorithm presented in [11]. In the case of a formal description of the business process, the corresponding transition system or automaton is formed on the basis of a decision tree containing pairs <initial state, final state>. If there is no description of the business process, then an algorithm for synthesizing the corresponding automaton based on a set of necessary lines of behavior/scenarios represented by regular expressions is proposed.

Based on the synthesis, a ready-made solution for a telegram bot was formed, on the basis of which a telegram bot was created using the existing messenger software interface and the execution time of a particular line of behavior/scenario for a specific task "Taxi Ordering". Taking into account the time and sequence of message and response creation, an approach was also proposed to calculate the chatbot operation time for different scenarios. It was determined that for standard scenarios T1= 204 (s), T2=324 (s), T3=467 (s), T4=80 (s) provided that the response from the data source (web service) on the available car types and the actual availability of the selected car is received in less than 20 seconds.

TASK OPTIMISATION IN MULTIPROCESSOR EMBEDDED SYSTEMS

Dmytro MARTINIUK — 2025-03-27

In this work, a task execution optimization method using replication in a multiprocessor system has been developed. This method effectively minimizes the overall execution time, ensures load balancing, and reduces communication delays. A key feature of the method is the implementation of task migration according to replication principles, utilizing an optimization objective function. The conducted experiment with the system demonstrated that the selected optimization method efficiently balances the load; however, additional objective functions are required for energy consumption optimization.

Simulation results show that increasing the number of processors reduces the maximum load and the number of migrations, while an increase in the number of tasks raises the system load and the number of migrations at the initial stages. The migration mechanism effectively balances the load, particularly in the early execution phases.

Future research directions include refining embedded device classifications and detailing their characteristics. For each class of embedded devices, it will be necessary to adapt task optimization algorithms and methods, as well as develop appropriate optimization objective functions.

METHOD AND CYBER-PHYSICAL SYSTEM FOR FORECASTING AND OPTIMIZING ELECTRICITY CONSUMPTION IN RESIDENTIAL DISTRICTS BASED ON MACHINE LEARNING ALGORITHMS

Volodymyr PYSMENIUK — 2025-03-27

Electricity is a key resource in the modern world, essential for industries, medicine, transportation, and daily life. With the increasing demand for electricity and the necessity of its efficient use, there is a growing need for advanced technologies for monitoring, forecasting, and optimizing electricity consumption. One promising solution in this field is the implementation of cyber-physical systems that integrate hardware and software for data collection, analysis, and energy resource management. The development of artificial intelligence and machine learning has led to an increasing number of solutions integrating these technologies into energy management. This study aims to develop a method and a cyber-physical system for forecasting and optimizing electricity consumption in residential districts using machine learning algorithms.

BAYESIAN OPTIMIZATION FOR TUNING HYPERPARAMETRS OF MACHINE LEARNING MODELS: A PERFORMANCE ANALYSIS IN XGBOOST

Mykola ZLOBIN — 2025-03-27

The performance of machine learning models depends on the selection and tuning of hyperparameters. As a widely used gradient boosting method, XGBoost relies on optimal hyperparameter configurations to balance model complexity, prevent overfitting, and improve generalization. Especially in high-dimensional hyperparameter spaces, traditional approaches including grid search and random search are computationally costly and ineffective. Recent findings in automated hyperparameter tuning, specifically Bayesian optimization with the tree-structured parzen estimator have shown promise in raising the accuracy and efficiency of model optimization. The aim of this paper is to analyze how effective Bayesian optimization is in tuning XGBoost hyperparameters for a real classification issue. Comparing Bayesian optimization with traditional search methods can help to assess its effects on model accuracy, convergence speed, and computing economy. As a case study in this research, a dataset of consumer spending behaviors was used. The classification task aimed to differentiate between two transaction categories: hotels, restaurants, and cafés against the retail sector. The performance of the model was evaluated using loss function minimization, convergence stability, and classification accuracy. This paper shows that Bayesian optimization improves XGBoost hyperparameter tuning, hence improving classification performance while lowering computational costs. The results offer empirical proof that Bayesian optimization outperforms traditional techniques in terms of accuracy, stability, and scalability.

USE OF SMART CONTRACTS ON THE TON BLOCKCHAIN FOR INNOVATIVE EDUCATIONAL SOLUTIONS DEVELOPMENT

Viacheslav ASKEROV — 2025-03-27

In the modern world, blockchain technologies are gaining popularity due to their ability to ensure security, transparency and decentralization of data. One of the most promising platforms is The Open Network (TON), which provides unique opportunities for the development of smart contracts. This article discusses the main features of the TON blockchain and its advantages in the context of educational process automation. Smart contracts implemented on the TON platform can serve as a tool for optimizing educational systems. They allow to automate processes related to knowledge validation, grade management, and even finance in educational institutions. For example, smart contracts can provide automatic scholarships based on students' grades, as well as control over the implementation of curricula. The paper also analyzes the benefits of using smart contracts in the educational process, such as reducing administrative costs, increasing transparency, and reducing fraud risks. In addition, blockchain technologies provide an opportunity to create decentralized platforms for storing and sharing knowledge, which makes learning more accessible and effective. Particular attention is paid to the mathematical aspects that ensure the functioning of TON, as well as sharing mechanisms that allow the platform to process thousands of transactions per second. These technologies can be used to create educational applications requiring high bandwidth and data processing speed. The paper contains formulas that illustrate the technical characteristics of the TON blockchain and provides a detailed analysis of its architecture. The study shows that smart contracts on the TON platform have the potential to revolutionize educational processes by providing new tools for data management and security.

ENHANCED TWO-STEP AUGMENTATION METHOD FOR ANALYZING SMALL DATASETS IN MEDICAL APPLICATIONS

Myroslav HAVRYLIUK — 2025-03-27

Despite the enormous possibilities for data collection, situations still often arise where data is scarce. Insufficient data can significantly complicate their effective analysis, since most known approaches require a sufficiently large training sample to obtain accurate predictions. In the field of medicine, the problems of lack of data are quite common for a number of reasons (confidentiality, fragmentation and natural rarity). Accordingly, the development of algorithms that can at least partially eliminate the scarcity of data and demonstrate satisfactory efficiency is relevant. Existing techniques for analyzing small data based on their augmentation can improve the efficiency of traditional methods. However, along with an increase in the number of instances in the sample, the number of features also increases significantly, which can negatively affect the performance of machine learning methods.

In this paper, an improved two-step method was proposed for the intelligent analysis of short high-dimensional data sets based on a generalized regression neural network. A peculiarity of this approach is the avoidance of a multiple increase in the number of features in the augmented sample. The method was used to solve two regression problems: predicting the value of a function and determining the compressive strength of the femur. Both data sets contained less than 100 instances. The optimal parameters were determined using the Dual Annealing optimization algorithm for five distance measures: Euclidean, Chebyshev, Manhattan, Canberra, and cosine. The proposed method showed a significant reduction in errors (such as MAE, RMSE) compared to the traditional GRNN model. The developed technique also surpassed the accuracy of the input doubling method for both solved problems. Along with increasing accuracy, the proposed model also increased the execution time. Therefore, the feasibility of its application depends on the priorities of the problem being solved.

CRISP-FUZZY RULE MANAGEMENT IN THE CLOUD: ENABLING SCALABLE DECISION-MAKING FOR CYBER-PHYSICAL SYSTEMS

Anatoliy MELNYK — 2025-03-27

Today, many cyber-physical systems (CPS) rely on local decision-making frameworks that often fail to address both precise thresholds and the ambiguity inherent in sensor data. There is a need to develop a scalable, cloud-based decision support system (DSS) that unifies crisp rule evaluation with fuzzy logic to improve decision accuracy and responsiveness across diverse applications. The aim of this paper is to design and implement a cloud-hosted crisp-fuzzy rule management system that supports centralized rule administration and asynchronous processing for multiple CPS domains.

Our approach employs a microservices architecture within a Microsoft Azure environment, comprising three core APIs: User Management, Knowledge Management, and Decision Support. The system integrates secure multi-tenant access using external identity providers and leverages a PostgreSQL database with a multi-tenant schema. Sensor data from various devices are transmitted via HTTP and queued through Azure Service Bus, thereby decoupling data ingestion from intensive rule evaluation. A background worker, known as the Decision Relay Consumer, processes each incoming message by applying direct threshold comparisons for crisp rules and linear interpolation for fuzzy membership functions, thus handling uncertain sensor readings effectively.

Experimental validation using a smart garden simulation demonstrates that the integration of crisp and fuzzy rule evaluations enhances the system’s ability to prioritize and trigger appropriate actions in real time. The results confirm that the proposed architecture not only improves decision-making reliability under ambiguous conditions but also reduces on-device computational burdens, facilitating centralized management and scalability.

The novelty of this work lies in its unified framework that seamlessly combines crisp thresholds with fuzzy logic in a cloud-based environment, enabling cross-domain applicability and adaptive rule management. The practical significance extends to various industries—including agriculture, manufacturing, and smart buildings—where timely and robust decision-making is essential.

METHOD FOR EXTENDING IMAGE CLASSIFICATOR VIA TEXT METADATA STATISTICAL ANALYSIS

Dmytro DASHENKOV — 2025-03-27

The goal of this article is to create a new method for obtaining a neural network for image classification that would work with classes for which examples are not available at the time of training. The task of image classification involves assigning one or more labels to an image based on the objects present in the image. The current state of the art method for creating such neural networks is to train models on the necessary data in a fine-tuning manner. The research methodology is to use existing machine learning models and expand the set of classes that the model operates on by manipulating the weight coefficients of the existing classifier model. The proposed method uses text metadata related to the images and descriptions of object classes to build assumptions about the relationship between different image classes. The method involves, using simple statistical calculations on text data, based on the existing weights of the neural network classifier, generating additional weights for recognizing new classes of objects in the image. The result of the research is the development of an algorithm for obtaining a classifier model that works with a class or classes that are not available during training. The model shows a classification accuracy result higher than the basic random one. At the same time, the classification accuracy for new classes, expressed in the F-score measure, is approximately 0.66, which is lower than the corresponding F-score measure for classes that were present during training, which is approximately 0.93. Also, the paper shows the limitations of the statistics-based approach to fine tuning, highlighting that it is not a full replacement for the classical model training. The scientific novelty lies in the development of methods for expanding image classifier models using statistical analysis of text metadata. The practical significance of the research lies in two aspects. The first aspect is obtaining a more stable base line of classification quality for classes that are added to the models after training using more sophisticated methods. The second aspect is obtaining a method for expanding the classifier for cases when extra data for additional training is not available and the training process itself is not possible due to a lack of computational resources.