Background. The rapid growth of Internet of Things (IoT) systems has increased the demand for scalable and low-latency data processing architectures. Traditional cloud-centric approaches often suffer from high communication delays and bandwidth limitations. Edge–Fog–Cloud computing introduces a multi-tier model that distributes computational tasks closer to data sources. However, evaluating computational methods in such heterogeneous environments requires systematic performance analysis and architectural optimization. In this context, integrating mathematically stable and computationally efficient methods, such as harmonic potential field–based approaches, is essential to ensure reliable real-time operation, scalability, and system resilience across distributed layers. Methods. This study evaluates the Laplace artificial potential field method implemented within a multi-tier Edge–Fog–Cloud architecture. The experimental framework includes distributed simulation, real-time processing scenarios, and comparative benchmarking. Performance metrics such as latency, computational load, and system stability were analyzed. The proposed approach was tested under variable workload conditions to assess scalability and efficiency across architectural layers. Results and Discussion. Experimental results demonstrate reduced end-to-end latency and improved task distribution across edge and fog layers. Compared to centralized processing, the proposed architecture maintains stability under increased workload. The Laplace-based computational model ensures efficient obstacle handling and balanced resource utilization. These findings confirm that multi-tier orchestration enhances system responsiveness while preserving acceptable computational overhead in dynamic IoT environments. Conclusion. Integrating the Laplace artificial potential field method within an Edge–Fog–Cloud architecture significantly improves distributed system performance. The proposed framework increases scalability, reliability, and computational efficiency in real-time IoT applications, providing a solid foundation for further optimization of resource management and intelligent task allocation in heterogeneous distributed environments.
Background. Monocular visual odometry is an important component of visual navigation systems. However, its accuracy depends on the quality of local features and inter-frame correspondences. In the VO task, not only is geometric consistency important, but also motion observability, the physical validity of the recovered configuration, and the spatial-structural properties of local features. This study aims to provide a comprehensive evaluation of keypoint detection and description methods for monocular visual odometry. Materials and Methods. The study was conducted on the EuRoC MAV dataset. The ORB, BRISK, AKAZE, KAZE, SIFT, SURF, and SuperPoint methods were analyzed for the number of keypoints, ranging from 200 to 1000. Motion estimation was performed using the essential matrix, the USAC_FAST filter, the recoverPose method, a minimum parallax check, and spatially guided keypoint selection. The accuracy of the recovered trajectory was evaluated using the APE and RPE metrics. To analyze the quality of local features and correspondences, the geometric component, the parallax indicator, the correct cheirality ratio, and metrics of keypoint coverage uniformity, local redundancy, and structural consistency were used. An integral quality indicator was applied to summarize the results. Results and Discussion. The geometric metrics most often highlight AKAZE and SURF, whereas SuperPoint shows strong performance in terms of spatial characteristics. In terms of the structural consistency of correspondences, SURF consistently demonstrates the best results. As the number of keypoints increases, most methods show an initial improvement followed by saturation, and in some cases, a deterioration of individual characteristics. SURF was found to be the most balanced method across the set of criteria, whereas ORB showed the weakest results in most cases. The correlation analysis showed that the informativeness of the metrics varies by sequence type. Conclusion. The proposed approach confirmed the relevance of multicriteria evaluation of local features in monocular visual odometry. It was shown that no single metric is universal across all scene types. In contrast, the integral indicator enables the summary of different aspects of quality and a more well-grounded ranking of the methods.
Background. Graphene field-effect transistors (FETs) have high potential for application in sensor electronics as detectors of electromagnetic radiation in a wide spectral range due to the high sensitivity of graphene ambipolar conductivity to local changes in the electric field. The use of a reduced graphene oxide (RGO) film provides cost reduction of photodetectors based on graphene FETs. On the other hand, an additional porous silicon light-absorbing layer can increase their sensitivity due to an increase in surface area. Materials and methods. Graphene field-effect photodetectors were created by drying a film-forming RGO suspension deposited on the surface of the porous silicon on a silicon substrate, which served as the gate of the FET. Electrical source and drain contacts were formed on the surface of the obtained RGO film. To improve the insulating properties of porous silicon, it was electrochemically oxidized, and an additional layer of Al2O3 was deposited. The electrical and photoelectric properties of the created field-effect photodetectors were investigated in DC and AC modes using a white LED and standard optical equipment. Results. An increase in the conductivity and capacitance of the RGO channel of the FETs was detected under the influence of white light irradiation. Based on the analysis of the drain current dependencies on the gate voltage, it has been established that the efficiency and photosensitivity of the FETs based on the porous silicon and RGO film are increased by the deposition of an additional Al2O3 layer on the surface of electrochemically oxidized porous silicon. The maximum sensitivity of the created photodetectors is in the spectral range of 800–900 nm. The response time to white light pulses is about 0.5 ms. Passivation of the porous silicon surface with the oxide film and the Al2O3 layer causes an increase in the photosignal relaxation time. Conclusions. The features of using FETs based on porous silicon structures and RGO film as visible radiation detectors have been investigated. The electrical, spectral, and time characteristics of the created field-effect photodetectors were determined.
Background. Fall detection is a critical challenge in healthcare and elderly care, as delayed response often leads to severe injuries. With ageing populations, fall-related admissions continue to rise, increasing demands on automated monitoring. Approaches based on wearable devices or conventional classifiers produce frequent false alarms and show limited adaptability. Video-based systems offer broader coverage but still require models that capture posture and motion changes without handcrafted features. Vision Transformers, originally developed for image recognition, provide a promising alternative by leveraging self-attention to model complex dependencies across spatial and temporal dimensions. Materials and Methods. A Vision Transformer framework was applied to model spatial and temporal patterns in human motion. Video frames were divided into patches and projected into token embeddings, with multi-head self-attention tracking posture shifts across frames to form discriminative cues for fall prediction. Training was conducted on multiple public datasets with diverse backgrounds and subject body types. The model was compared with logistic regression and CNN baselines trained on identical data splits. Results and Discussion. The Vision Transformer achieved 99.1% accuracy on the primary dataset and 97.9% on the UR Fall Detection Dataset, surpassing logistic regression, CNN, and LSTM baselines. It maintained higher precision and recall in indoor and outdoor scenes and reduced false alarm rates. Stable performance under rapid movement and variable lighting demonstrated robustness gains. Cross-dataset evaluation confirmed effective transfer of learned spatial-temporal representations to unseen environments. Conclusion. Vision Transformers offer an effective approach for real-time, non-invasive fall detection in clinical and home settings. Their capacity to capture spatial-temporal motion patterns through self-attention, without handcrafted features, supports broader deployment in intelligent surveillance systems. The proposed framework demonstrates strong generalization across datasets and recording conditions. Future work will target edge-device optimization and multi-modal data integration.
Background. β-Ga2O3 gallium oxideis a promising wide-bandgap semiconductor widely used in optoelectronic and sensing applications. The electrical conductivity of thin films strongly depends on their structural quality, defect states, and post-deposition treatment conditions. In polycrystalline films, grain boundaries and defect complexes significantly affect charge transport mechanisms. The objective of this study is to investigate the influence of annealing atmospheres on the structural, morphological, and electrical properties of (Y0.06Ga0.94)2О3 thin films. Materials and Methods. Thin films of (Y0.06Ga0.94)2О3 with thicknesses of 0.3–1.0 μm were deposited by RF ion-plasma sputtering onto fused quartz substrates. Post-deposition annealing was carried out in oxygen and argon atmospheres at 1000–1100 °C, and in hydrogen at 600–650 °C. Structural properties were analyzed using X-ray diffraction, while surface morphology was examined by atomic force microscopy. Electrical conductivity was measured in the temperature range of 300–450 K, and activation energies were determined from temperature-dependent conductivity data. Results and Discussion. X-ray analysis confirmed the formation of films in the monoclinic β-Ga2O3 phase, with enhanced crystallinity and preferred orientation after annealing in oxygen. It has been established that freshly deposited films have a high resistivity (ρ > 10¹¹ Ω·cm), which decreases with increasing temperature and after annealing. Oxygen annealing resulted in activation energy of ~0.87 eV, while argon annealing produced higher values (~1.38 eV in 300–400 K range), indicating deeper donor levels associated with oxygen vacancies. Hydrogen annealing significantly reduced resistivity (~10⁸ Ω·cm) and activation energy (~0.40 eV), attributed to shallow donor states. Conclusion. The electrical conductivity of (Y0.06Ga0.94)2О3 thin films is governed by defect-related donor levels formed during annealing. Oxygen and argon atmospheres promote deep donor states, while hydrogen enhances shallow donor formation, leading to improved electrical conductivity.
Background. Auto-guidance for unmanned aerial vehicles (UAVs) requires reliable real-time target tracking on resource-constrained onboard hardware. Modern state-of-the-art CNN-based and Transformer-based deep trackers provide strong accuracy but are often too slow and computationally expensive for continuous deployment on edge devices. In contrast, lightweight correlation-filter trackers run at high frame rates but can easily drift or lose the target because of occlusions or fast maneuvers. This robustness–efficiency trade-off (edge AI paradox) motivates adaptive strategies that balance accuracy, speed, and resource usage while preserving compute headroom for other onboard tasks. Materials and methods. We propose an entropy-guided tracker switching method that combines a lightweight kernelized correlation filter (KCF) tracker augmented with Kalman motion prediction and a more accurate Siamese deep tracker. A motion-entropy scheduler quantifies the unpredictability of target motion using a normalized Shannon entropy over recent orientation changes. To avoid reacting to transient spikes, the entropy is exponentially smoothed, and threshold rules (with hysteresis) determine when KCF is sufficient and when to activate the deep tracker. Results and Discussion. Experiments on UAV benchmarks (UAV123, OTB100) show that the hybrid tracker improves success AUC by ~10% over KCF and reaches about 70% of a Transformer tracker’s AUC while running 1.5–3× faster than always-on deep tracking. The switcher invokes the deep tracker only during difficult intervals, sustaining real-time operation (~100 FPS) and reducing average computation to ≈0.6 GFLOPs per frame versus ≈1–4 GFLOPs for purely deep tracking. Conclusion. The proposed motion-entropy scheduler enables an adaptive trade-off between efficiency, speed, and accuracy. It maintains high tracking precision during target maneuvers and occlusions by temporarily switching to a robust tracker yet saves computational load during steady-motion periods. This framework offers a practical solution for high-performance UAV tracking on the edge, while leaving resource headroom to apply other improvement techniques.
Background. This paper presents a deep reinforcement learning approach for intelligent thermal management in embedded electronics, targeting energy-efficient and safe operation under dynamic workloads. A custom hardware switching circuit based on an NPN transistor was designed to enable GPIO-driven fan actuation on a resource-constrained platform. Materials and Methods. A real-time dataset was collected from a Raspberry Pi Zero W, capturing CPU temperature, usage metrics, and fan states over a 12-hour controlled experiment. The thermal regulation task was modeled as a Markov Decision Process, and a Deep Q-Network (DQN) was trained to learn optimal fan activation policies. The trained model was deployed directly on-device, interfaced with a custom GPIO-controlled fan circuit. Inference was performed in less than one millisecond per decision step using a lightweight PyTorch runtime. Results and Discussion. Evaluation results show that the DQN policy reduced total fan activation time by 23.2% compared to the rule-based hysteresis baseline, while maintaining CPU temperature below 60°C for over 99% of the test duration. The trained agent activated the fan only 23.7% of the time, demonstrating a conservative and energy-aware cooling strategy. Confusion matrix analysis yielded a precision of 1.000, a recall of 1.000, and an F1-score of 1.000 across 3442 model-controlled evaluation steps. The model correctly identified all 22 fan activation events without any false positives or false negatives. Comparative analysis against nine recent AI-driven approaches showed that the proposed method achieved an 11.2°C temperature reduction and 36.5% energy savings, while operating entirely on-device without cloud dependence. Conclusion. The model exhibited stable reward convergence, accurate action prediction, and anticipatory control that minimized overheating events. Thermal traces confirmed smooth transitions and low variance, demonstrating the feasibility of deploying learning-based thermal policies in real-time edge environments. This work contributes a practical framework for energy-aware cooling and provides a pathway for adaptive thermal intelligence in low-resource embedded systems.
Background. Cloud-integrated spectrometric laboratories face communication challenges in achieving real-time data access and analysis. This study compares two wireless protocols, MQTT (Message Queuing Telemetry Transport) and ESP-NOW (Espressif NOW), for LED control in such environments. MQTT offers lightweight, bandwidth-efficient, publish-subscribe messaging [1–3], while ESP-NOW provides energy-efficient direct communication without a Wi-Fi router. The objective is to evaluate their performance and suitability. Materials and Methods. An experimental setup involved a StellarNet spectrometer, LED light sources, and ESP32 microcontrollers. Two architectures were tested: 1) direct MQTT Communication, where each ESP32 connected directly via Wi-Fi to an MQTT broker; and 2) an edge device with ESP-NOW relay, using an edge ESP32 for MQTT/Wi-Fi communication, then relaying commands via ESP-NOW to other ESP32s. Response times for LED control were measured over 100 cycles, and data were analyzed using descriptive statistics and an independent samples t-test. Results and Discussion. Direct MQTT Communication exhibited significantly lower latency (median ~60 ms) and tighter distribution compared to the Edge Device with ESP-NOW Relay (median ~170 ms). A t-test confirmed a statistically significant difference (t=-46.28), with MQTT demonstrating faster response times. However, the ESP-NOW relay system offers architectural advantages: reduced Wi-Fi dependency for individual nodes, enhanced deployment flexibility in areas with poor Wi-Fi coverage, improved scalability [4], and energy efficiency, making its higher latency a practical trade-off for large-scale laboratory integration. Conclusion. Direct MQTT Communication provides superior low-latency performance. However, the edge device with ESP-NOW relay, despite higher latency, is a highly acceptable solution due to its flexibility, scalability, and reduced Wi-Fi dependency for distributed spectrometric laboratories. This highlights a critical trade-off between absolute speed and architectural benefits for robust, smart, cloud-enabled analytical laboratories.
Background. Large language models (LLMs) are increasingly used in educational analytics, particularly for processing large volumes of accreditation-related documents. However, it remains unclear how reliably such models can assess the quality of self-evaluation reports for educational programs, and which textual characteristics influence how models form their assessments. Materials and Methods. In the study, ten self-evaluation reports of educational programs were analyzed: five identified by the expert assessment as the strongest within the higher education institution over the last three years, and five as the weakest over the same period. GPT-5 and Gemini-2.5 models independently evaluated each document using the official ten Ukrainian National Agency for Higher Education Quality Assurance (NAQA) criteria and eight textual metrics reflecting structural, semantic, argumentative, and factual properties of the text. All evaluation grades were generated directly by the models on a unified scale from 1 to 10. To analyze the relationships between NAQA and textual criteria, Pearson's and Spearman’s correlation coefficients were used. Results and Discussion. LLMs demonstrated limited alignment with the NAQA criteria, yielding weak correlations. In contrast, textual criteria, primarily factual density, argumentativeness, semantic coherence, and lexical diversity, consistently differentiated between stronger and weaker reports. GPT-5 exhibited lower variability and reduced sensitivity to stylistic noise, while Gemini-2.5 reacted more strongly to structural and stylistic deficiencies. Correlation matrices showed that textual criteria better capture the latent quality characteristics of documents than the direct application of NAQA criteria. Conclusion. The results show that LLMs currently do not accurately reproduce expert evaluations based on the formal NAQA criteria but effectively analyze the structural and content-related characteristics of reports using textual metrics. These metrics complement the NAQA criteria by accelerating expert workflows and enhancing document monitoring. Future research will focus on expanding the dataset, standardizing prompts, and comparing a broader range of models.
Background. The increasing role of web-oriented information systems in business, education, and public administration is accompanied by a growing number and complexity of cyber threats. Traditional security mechanisms do not always enable the identification of actual system weaknesses, which necessitates the application of practice-oriented methods for assessing the level of information security. In this context, Penetration Testing is considered an effective instrument for simulating the actions of a potential attacker in order to detect and validate exploitable vulnerabilities. Materials and Methods. The study employs a risk-oriented approach in accordance with international standards ISO/IEC 27001 and ISO/IEC 27005, as well as the recommendations of OWASP and NIST SP 800-115. Penetration Testing is implemented as a structured, multi-stage process that includes information gathering, attack surface analysis, threat modeling, execution of non-invasive validation scenarios, and risk assessment. The practical component was conducted in a controlled test environment using Nmap, Burp Suite, and Wireshark, supplemented by custom-developed Python modules for automated analysis of HTTP security headers, TLS certificates, and exposed services. Results and Discussion. The study identified several configuration-related weaknesses at the application level, including the absence of essential HTTP security headers and deficiencies in TLS certificate management. The obtained results were formalized in a structured findings register with quantitative risk evaluation based on the Likelihood × Impact model. The analysis demonstrated that even in the absence of critical exploitable vulnerabilities, configuration errors significantly increase the overall risk level and may create preconditions for more sophisticated attacks. Conclusion. The findings confirm the effectiveness of Penetration Testing as a comprehensive instrument for assessing the information security of web-oriented systems. The proposed approach facilitates the transition from technical testing results to substantiated managerial decisions aimed at risk reduction and enhancement of the overall protection level of information resources.
Background. The rapid expansion of data-driven applications has increased the importance of efficient query execution in relational database systems, where even minor inefficiencies can significantly affect overall performance. Although Object-Relational Mapping (ORM) frameworks simplify development and improve maintainability, their abstraction layer can introduce measurable overhead, and the impact of foreign key constraints on execution speed remains a practical concern, particularly in microservice architectures that follow the “Database per Service” principle. Materials and Methods. An experimental information system is developed using a relational database and the SQLAlchemy ORM framework, with a schema that includes one-to-one, one-to-many, and many-to-many relationships tested both with and without foreign key constraints. Three representative queries retrieving booking details, aggregating related records, and calculating total payments are executed using raw SQL and ORM approaches, while an intelligent algorithm analyzed performance, detected potential N+1 query risks, and recommended optimal strategies such as explicit JOINs. Results and Discussion. Raw SQL consistently demonstrated superior performance across all scenarios. The most significant disparity occurred in ORM implementations affected by the N+1 problem, where execution time exceeded that of equivalent SQL queries by more than an order of magnitude. Aggregation queries showed smaller yet consistent overhead. The presence or absence of foreign key constraints had a negligible influence on raw SQL performance, with differences remaining within experimental variance. Explicit JOIN usage in ORM substantially reduced overhead compared to implicit relationship navigation. The intelligent analysis accurately predicted high-risk queries and provided effective strategy recommendations, confirmed by empirical results. Conclusion. ORM frameworks improve productivity and maintainability but introduce measurable overhead. Raw SQL remains preferable for performance-critical tasks, while foreign key constraints do not significantly degrade execution speed. Intelligent performance analysis supports balanced decisions between efficiency and maintainability in complex relational systems.
Introduction. This paper considers a method based on a Long Short-Term Memory (LSTM) neural network for optimal resource allocation in distributed systems. The developed algorithm ensures high accuracy in predicting resource states and optimal spatial distribution with minimal processing time. The relevance of this research is determined by the growing need for intelligent automation of resource management processes in service infrastructure facilities. A locker management system in sports facilities is used as a practical demonstration of the method's effectiveness. Materials and Methods. To address the prediction and optimization tasks, an LSTM-based architecture with 32 hidden neurons and a sequence length of 10 time steps is proposed. The LSTM model processes sequential occupancy data to capture temporal dependencies and generate probability estimates for future resource states. A multi-factor scoring function is developed to transform predictions into optimal allocation decisions, considering spatial constraints and user preferences. The method is systematically compared with classical approaches: heuristic algorithms (Sequential, Round-Robin), statistical time series models (ARIMA, exponential smoothing), and machine learning methods (logistic regression, random forest, gradient boosting). All methods are evaluated on identical datasets using consistent metrics, including prediction accuracy, F1-score, spatial balance index, and zone variance. Results. Using LSTM neural networks for the prediction task achieves 85% accuracy, which is statistically significantly higher than Random Forest (79%, p=0.0023) and ARIMA (68%, p=0.0001). The spatial balance index improved by 8.5% compared to the best classical method (0.89 versus 0.82). Inference time remains acceptable for real-time applications (18.9 ms per prediction). Conclusions. The proposed LSTM-based method demonstrates satisfactory accuracy in predicting resource states and optimizing their allocation within minimal timeframes. The ability to model long-term temporal dependencies provides significant advantages over classical fixed-window methods. Therefore, the method can be effectively applied to enhance the functionality of distributed resource management systems.
Background. In modern network security systems, DNS (Domain Name System) traffic has become an increasingly attractive vector for covert data exfiltration and command-and-control communication. Existing machine learning methods frequently suffer from limited adaptability to novel attack patterns and an imbalance between detection accuracy and false positive rates. Materials and Methods. This study proposes TunnelEye, a multi-level detection method for malicious DNS queries that integrates statistical feature analysis, structural n-gram modeling, and anomaly detection. Statistical properties of domain names, including string length, entropy, and alphanumeric ratio, are used for initial discrimination between benign and suspicious queries. Structural analysis based on character n-grams enables the identification of local patterns associated with encoded data such as Base32 and Base64. An autoencoder trained exclusively on legitimate DNS queries is employed as an independent anomaly detector to identify previously unseen and zero-day attacks. The supervised TunnelEye classifier and the autoencoder operate in parallel, each using an independently optimized F1-score based threshold to determine anomalous DNS queries. Results and Discussion. Experimental evaluation using standard machine learning metrics (precision, recall, F1-score, ROC-AUC, PR-AUC, and false positive rate) demonstrates that TunnelEye consistently outperforms baseline statistical models and standalone autoencoders. The proposed method achieves high precision and recall while maintaining a minimal false positive rate. Experimental results show that TunnelEye achieves an average precision, recall, and F1-score of approximately 0.99, outperforming the baseline statistical model by more than 10% and significantly reducing the false positive rate. Conclusion. TunnelEye provides a comprehensive and adaptive solution for malicious DNS query detection by combining supervised and unsupervised learning with dynamic threshold optimization. Its ability to balance detection accuracy and false positive reduction makes it well-suited for deployment in modern enterprise cybersecurity systems for real-time DNS traffic monitoring.
Background. The development of modern mathematical computing systems requires the effective implementation of machine learning algorithms while maintaining a balance between prediction accuracy and computational resources. Particular attention should be given to the phased integration of neural networks of varying complexity with minimized risks for production systems and investigation of the saturation effect when increasing architectural depth. Materials and Methods. This article aims to develop a methodology for evolutionary integration of neural networks from simple perceptrons to ultra-deep architectures in mathematical computing systems, with detailed comparative analysis of four architectural types and mathematical modeling of the accuracy saturation effect. Results and Discussion. For this purpose, four neural network architectures were investigated: a single-layer perceptron, a four-layer network (128→64→32), a ten-layer network (128→96→64→48→32→24→16→12), and a twenty-layer architecture with gradual dimensionality reduction. Experiments were conducted on a dataset from mathematical modeling results containing 45,000 samples with 24 characteristics. A comprehensive system of metrics was used to evaluate accuracy, processing speed, resource consumption, and model stability. The experimental design included stratified data splitting and cross-validation to ensure statistical reliability of the obtained results across different architectural configurations. Conclusion. As a result, the single-layer perceptron demonstrated baseline accuracy of 78.3% with minimal resource consumption (45 MB RAM, 15 ms latency). The four-layer network achieved 94.1% accuracy with a moderate increase in resource costs. The ten-layer architecture showed 95.6% accuracy, demonstrating the beginning of the saturation effect. The twenty-layer network achieved only 96.8% accuracy with disproportionate growth in resource consumption (1024 MB RAM, 270 ms latency). Mathematical modeling confirmed the logistic nature of the relationship between accuracy and architectural complexity. The findings provide practical guidelines for selecting optimal neural network architectures in resource-constrained production environments, establishing clear thresholds beyond which increased complexity yields diminishing returns.
Background. Indoor positioning systems based on Bluetooth Low Energy (BLE) beacons widely rely on estimating distance using the received signal strength indicator (RSSI). However, RSSI measurements in indoor environments are significantly affected by multipath propagation, shadowing, interference, and absorption by obstacles, resulting in high variability of signal strength and substantial distance estimation errors. The nonlinear logarithmic relationship between RSSI and distance further complicates the application of conventional linear filtering techniques such as the classical Kalman Filter, which requires prior transformation of measurements and may lead to loss of optimality. Materials and Methods. This study proposes a distance estimation method based on the Extended Kalman Filter (EKF), which directly processes RSSI measurements using the nonlinear log-distance path loss model. The experiment was performed in an indoor office environment using two Silicon Labs EFR32BG22 BLE beacons and a Nordic nRF52840 receiver. The EKF parameters were selected based on prior calibration of the propagation model coefficients. Results and Discussion. The experimental results demonstrate that the EKF effectively smooths RSSI. For the beacon with lower RSSI dispersion, the root mean square error (RMSE) reached 0.14 m, for the second beacon, the RMSE was 0.53 m. The analysis confirms that estimation accuracy strongly depends on signal stability and calibration quality. Compared to direct RSSI-to-distance conversion and the classical Kalman Filter approach reported in related work, the EKF-based algorithm reduces the mean absolute distance estimation error by approximately 20–30%, validating the advantages of nonlinear filtering. Conclusion. The proposed EKF-based method improves the accuracy and robustness of RSSI-based distance estimation in BLE indoor positioning systems. When model parameters are properly calibrated, the achieved accuracy is sufficient for practical applications such as smart building navigation, asset tracking, and robotic localization. The algorithm can be implemented on resource-constrained embedded platforms and serves as a foundation for further development of multisensor indoor positioning systems.
Background. Today, software is a critically important component of any information system. Its development requires significant resources and complex technical solutions, and the development of technologies is so rapid that not all concepts and definitions in the field of software are clearly formalized. This is especially true for the software functional state (SFS) throughout the software development life cycle (SDLC), as predicting all possible states is virtually impossible due to the dynamic nature of environments, changing requirements, component interactions, and the behavior of project participants. This creates a challenge for formalizing, analyzing, forecasting, monitoring, and managing these states. Materials and Methods. The definition and formalization of SFSs encompass concepts from state theory in computer science, as well as quality models from international standards ISO/IEC 25010:2011 and the State Standard of Ukraine ISO/IEC 9126-1:2005. The defined concepts of SFS and SFS during SDLC are formalized mathematically, which allows building dynamic models of state evolution during SDLC based on the stochastic transition function. To build models, attributes such as functional compliance, reliability, vulnerability, testability, and others have been developed in combination with event-driven, finite-state machine, and state-driven models. Also presented are different types of SFS and their relationship with SDLC. Results and Discussion. The research results include the formalization of SFS, the development of evaluation metrics, and practical recommendations for SFS analytics at all stages of SDLC, which enable proactive control of the quality, reliability, security, and compliance of software systems. Conclusion. The formalization of the concept of SFSs, including their types, properties, and parameters, allowed for a reasonable connection to the SDLC phases. The proposed metrics and recommendations contribute to the development of SFS analytics, ensuring both the theoretical integrity of the approach and its practical applicability in the tasks of monitoring, analysis and predicting SFS. This methodology creates a new foundation for self-learning SDLC-oriented ecosystems in which SFSs are predicted, assessed and managed automatically in real-time.
Background. The rapid growth of electronic payments has intensified fraudulent activity, requiring adaptive anomaly detection methods. Traditional rule-based approaches lack flexibility and fail to generalize to previously unseen attacks. In contrast, unsupervised deep learning models, particularly autoencoders, can learn intrinsic data representations and detect anomalies without labeled attack samples. This study evaluates three unsupervised architectures – Autoencoder with Gaussian Mixture Model (AEGMM), Variational Autoencoder with Gaussian Mixture Model (VAEGMM), and a Deep Autoencoder – for network anomaly detection. Materials and Methods. Experiments were conducted using the KDD’99 (10%) benchmark dataset. Categorical features were transformed using one-hot encoding, while numerical features were standardized. All models were trained exclusively on normal traffic samples following a one-class learning paradigm. The experimental pipeline included preprocessing, model implementation in Python using TensorFlow and the Alibi Detect framework, percentile-based threshold calibration, and evaluation using accuracy, precision, recall, F1-score, and confusion matrices. Results. AEGMM achieved the highest performance with an F1-score of 0.9936 and an accuracy of 0.9908, demonstrating near-perfect separation between normal and malicious samples. VAEGMM reached an F1-score of 0.9751, showing stable convergence but slightly reduced accuracy due to the stochastic latent space. The Deep Autoencoder achieved approximately 97.5% accuracy, confirming the effectiveness of reconstruction-based methods without probabilistic density estimation. The optimal anomaly threshold, defined at the 99th percentile of reconstruction or density scores, ensured reliable discrimination between normal and attack states. Conclusion. Autoencoder-based unsupervised models are effective for anomaly detection in large, imbalanced tabular datasets. AEGMM outperformed alternative architectures due to its stable latent representation and deterministic optimization. The proposed approach is suitable for financial fraud detection, cybersecurity monitoring, and industrial anomaly detection. Future work will explore transformer-based models and Explainable AI to improve robustness and interpretability.
Background. The classification of in-game roles in team-based shooters, particularly Counter-Strike 2 (CS2), is an essential component of esports performance analytics. Existing approaches primarily rely on aggregate ratings or empirical assessments, which do not adequately capture the multidimensional structure of player behavior. As a result, there is a need to construct a behavioral feature space capable of reflecting role-specific differences and enabling reliable automated classification. Materials and Methods. To construct the feature space, publicly available statistics and, when necessary, .dem files containing detailed logs of in-game events were utilized. The foundation consists of seven HLTV behavioral attributes, supplemented with metrics specific to the Terrorist (T) and Counter-Terrorist (CT) sides, as well as map-dependent indicators. The data were pre-cleaned, normalized, and structured at the player–map level. For the analysis, Principal Component Analysis (PCA) was applied, along with Analysis of Variance (ANOVA) to identify map-dependent features, and correlation analysis to examine relationships among behavioral metrics. Results and Discussion. The results demonstrated that typical roles (entry-fragger, lurker, support, AWPer, anchor, and IGL) form distinct regions within the multidimensional feature space that cannot be reduced to a single numerical index. A set of features most influential for differentiating roles was identified, along with metrics that exhibit stable behavior regardless of map or side. The analysis based on grouping players revealed the absence of a universal player profile: strong performance in some metrics is accompanied by lower values in others, reflecting natural role specialization. Conclusion. The proposed approach provides an informative representation of behavioral features and enables automated identification of player roles in CS2 without relying on aggregate rating systems. The constructed feature space has practical value for scouting, roster optimization, and match analysis, and can also be adapted for detecting smurfing or other forms of anomalous activity. The methodology demonstrates interdisciplinary potential and is promising for broader applications in behavioral analytics within online services.
Background. Ensuring security for private areas and infrastructure hubs is a growing concern in the modern world. Traditional methods, such as human guards and mechanical barriers operated by physical tokens (keys or cards), are often slow, inefficient, and prone to security risks like unauthorized duplication or theft. Furthermore, legacy systems lack comprehensive auditing capabilities. This creates a critical need for modern, automated IoT-based systems that ensure reliable access management and real-time monitoring. Materials and Methods. The system uses several electronic components. The core is a low-cost microcontroller with a camera module. A radio-frequency identification (RFID) reader scans access cards. An ultrasonic distance sensor detects obstacles for safety, and a servo motor operates the physical barrier. The software backend was developed in Python, with a JavaScript (React) web control panel. The system combines two identification methods. First, a camera captures a vehicle's image, sending it to a server where an AI model reads the license plate. The server checks the number against an approved list. If not recognized, the driver scans an RFID card as a secondary method. A distance sensor continuously monitors the barrier area to prevent closing on an obstacle. A web interface allows an operator to monitor the camera, review logs, and manually open the barrier. Results. The developed system was tested successfully. The AI model achieved 75% accuracy in identifying license plates. The system proved fast, with an average response time from image capture to decision under one second. The safety sensor was validated, reliably detecting obstacles and preventing barrier movement, ensuring safe operation. Provided results of comparing video quality and system response time. The optimal balance between video quality and speed was found at 800x600 resolution. Conclusions. A reliable, cost-effective automated access control system was successfully designed, built, and tested. The combination of AI-based license plate reading with a backup RFID system provides a robust, flexible solution. This system is well-suited for improving security and efficiency in real-world applications like residential, office, and industrial zones.
