The article proposes a hybrid cascaded neuro-fuzzy system for classifying defects in renewable energy facilities based on multisensor data. The system architecture implements a sequence of computational levels (from sensor acquisition and preprocessing to result integration) and combines a compact feature representation with adaptive decision-making logic. The formation of a compact feature space is carried out using a modified convolutional neural network, after which the initial classification is performed using the hypersector FLVQ method. In cases of increased uncertainty, the expert modules Fuzzy BSB and the modified Wang–Mendel method are activated, which ensures robustness and explainability of the results. Experimental studies have shown that the proposed system provides consistently high Accuracy values (over 94%), balanced Precision and Recall indicators for all defect classes. Analysis of the discrepancy matrix showed that the main errors occur between the “erosion” and “corrosion” classes, which is explained by the similarity of their textural characteristics. The results obtained confirm the effectiveness of the cascade architecture and its feasibility for practical application in automated monitoring systems for renewable energy facilities.
This study presents a method for detecting ship coating quality using a portable Vis/NIR spectroscopy system combined with machine learning. To improve accuracy, we integrated spectral transformation (Nippy), feature selection methods (PSO and GA), and ensemble learning models. The experiments involved four coating quality levels, producing 148 spectral samples. Results show that the proposed approach consistently outperforms single baseline models and traditional feature selection methods such as PCA and IFS. The best performance was achieved by combining Nippy with PSO, where the LDA algorithm reached 99.33% accuracy, while GA also showed strong results with both single and ensemble models. We also examined ensemble results at different stages of preprocessing and feature selection, showing that the ensemble maintained stable performance throughout the process. These findings demonstrate that the integration of spectral transformation and metaheuristic feature selection can enhance model robustness, providing more reliable and accurate coating quality detection for maritime applications.
The algorithm used in the feature selection process is Hybrid Binary Grey Wolf Optimizer-Binary Sine Cosine Algorithm (HBGWO-BSCA). The dataset used for the HBGWO-BSCA is Parkinson's. This dataset is used because Parkinson's disease is one of the most frequently discussed diseases throughout the world. The aim of this paper is to find out what features are used in the process of predicting Parkinson's disease. The HBGWO-BSC feature selection algorithm was proven to be able to increase the accuracy value of the KNN classification algorithm by 92%, while the HGWO-SCA only obtains an accuracy of 88%. The value of the HBGWO-BSCA is higher than that of HGWO-SCA because the relationship between features selected with the HBGWO BSCA is more accurate than that of the Hybrid Grey Wolf Optimizer-Sine Cosine Algorithm (HGWO-SCA). This proves that the HBGWO-BSC feature selection algorithm obtains the highest accuracy, precision, recall and F1-Score values compared to the HGWO-SCA. The HBGWO-BSCA in feature selection uses parameters N and alpha with parameter value ranges N= 9-11 and alpha=2-8. HBGWO-BSCA is an algorithm used for the feature selection process in datasets that already have labels or classes.
The goal of the paper is to create a mathematical model for optimizing the composition of an IT project team, which would take into account the need to maximize the technical skills of candidates, their ability to perform the necessary roles, considering a set of assessments of these factors, as well as the agreeableness and conscientiousness of its members. The model should take into account the limitations on team members' competencies, their working hours, and labor costs. An approach to solving the problem is proposed, which involves presenting the requirements for candidates' competencies and assessing their competencies and abilities to perform specific roles in the project using fuzzy sets. At the same time, assessments of competencies and abilities to perform certain roles are represented by trapezoidal fuzzy intervals. A method is proposed for defuzzifying the problem by calculating the value of the membership function of the requirement at a point equal to the lower modal value of the fuzzy interval describing the properties of the candidates. The proposed task may not have a solution if the candidates' competencies do not meet the constraints. The model allows finding the optimal extension of the set of feasible solutions by training candidates. An example of solving the task of forming an IT team is considered.
In today's competitive market, effective customer segmentation is essential for businesses to refine marketing strategies, optimize resources, and enhance customer satisfaction. This study introduces the Weather-Responsive Segmentation (WRS) framework by integrating autonomous K-means clustering with Principal Component Analysis (PCA) while incorporating real-time meteorological data. Our approach achieves 27.8% higher segmentation accuracy (Silhouette Score: 0.524 vs 0.410) and 23.4% improved campaign conversion rates compared to traditional demographic-only methods. Traditional segmentation relies primarily on demographic and behavioral data, but our method adds a new dimension by considering weather-related factors that significantly impact consumer behavior. Using K-means, we identified four distinct customer segments and applied PCA for data visualization and dimensionality reduction. The inclusion of meteorological variables enhances segmentation accuracy and relevance, providing businesses with more actionable insights. By incorporating external environmental factors, this approach offers a deeper understanding of customer groups and enables more precise, data-driven marketing strategies. This research contributes to the field by demonstrating the added value of weather-based segmentation in consumer analysis, offering businesses a novel perspective to optimize marketing efforts and improve decision-making.
This article proposes and empirically tests an integrated approach to monitoring university digital transformation in a single-institution setting. The approach combines the measurement of students’ digital competence (Digital Competence Index, DCI) and teachers’ digital pedagogical competence (Digital Competence in Education, DCEdu) with selected indicators of course process quality and digital governance practices. The study is based on two standardised online surveys administered in the LMS Moodle environment (students: n = 386; teachers: n = 89), with index normalisation to a 0–100 scale and domain decomposition. The results revealed substantial internal heterogeneity in students’ digital competence and marked inter-faculty differentiation in DCI, primarily associated with the size of the lower segment of the distribution. The largest differences in DCI were associated with accessibility barriers, internet stability, and regularity of interaction with the LMS. For teachers, the aggregated DCEdu index appeared relatively insensitive to basic infrastructure conditions but showed strong variation across pedagogical governance variables, especially the formalisation of rules for students’ use of generative artificial intelligence and cyber hygiene training. The study formulates a set of institutional recommendations focused on minimum course standards, authentic rubric-based assessment, formal regulations for the use of generative AI, and mandatory training in cyber hygiene and data management.
This paper critically examines how the integration of Edge Artificial Intelligence (Edge AI) and Cloud-Native Artificial Intelligence (Cloud-Native AI) can enhance agility, scalability, and governance in modern project management. Drawing on literature published between 2019 and 2025, the study applies the Technology-Organization-Environment (TOE) and Diffusion of Innovations (DOI) frameworks to explore the technical, organizational, and environmental factors enabling hybrid Edge-Cloud adoption. Using a structured narrative review, the research synthesizes evidence on how distributed intelligence architectures reshape project agility, responsiveness, and lifecycle automation. Beyond synthesis, the paper introduces a novel Dual-Loop Edge-Cloud Governance Framework. This domain-specific conceptual model formalizes two complementary governance cycles: an Operational Loop at the edge, supporting real-time, autonomous project execution and local decision-making; and a Strategic Loop in the cloud, driving organizational learning, policy refinement, and global consistency. This framework provides essential theoretical and practical guidance for project leaders seeking to manage the inherent complexity and conflicting demands of hybrid AI systems, ensuring both rapid project responsiveness and long-term organizational alignment and accountability.
This study is devoted to the problem of generating text corpora for their use during the development and testing of natural language processing information systems. The CorDeGen and CorDeGen+ methods are among the approaches that address this problem. However, as shown in this paper, the application of these methods to the development and testing of information systems for processing texts in “regional” languages (less widely spoken than English) has not yet been considered, despite its challenges. In this study, the languages of Northern Europe are considered as such “regional” languages, and the issue of removing part of the terms (if they coincide with the stop words of these languages) from the generated corpora during preprocessing is solved. To address this issue, the paper proposes seven new language variations of the CorDeGen+ method, specifically for Lithuanian, Danish, Swedish, Norwegian, Northern Sami, Lule Sami, and Icelandic languages. Latvian, Estonian, Finnish, and Southern Sami languages are also considered in this study, and the results show that the use of the CorDeGen+(0-9) variation, already described in the literature, is sufficient for them. The experimental verification of the effect of removing part of the terms showed that the use of the proposed language variations and CorDeGen+(0-9) variation prevents the removal of 20–43% of all terms from the generated corpus during preprocessing.
This research is devoted to the problem of developing a specialized novel method for identifying the role of HCI subjects based on their multifactor portraits of perception subjectivization of the object of this interaction, which provides the possibility of increasing the level of HCI automation and intellectualization by additional evaluation of the factor of perception subjectivization of the interaction object by the subjects of this interaction. The proposed method is based on the developed polycomponent model, which can contain any (and necessary) number of impact factors as its components, and also is represented in two variations, namely strongly structured (when the placement order of dominant impact factors has significant and fundamental importance, and must be met), or weakly structured (when the placement order of dominant impact factors is not required and obligatory). Another fundamental component of the proposed method is a specialized algorithm developed for identifying the role of HCI subjects based on their multifactor portrait, which provides the possibility of algorithmization of the researched processes, as well as the opportunity for further software implementation and computer modeling of the developed method. As a practical approbation of the developed method, the relevant applied task of identifying a potential candidate(s) from among all available into the highly specialized support team of the given software product based on their compliance with the declared role pattern, has been solved. The obtained results confirm the effectiveness of the developed method, as well as its perspective in the context of further research in the field of automation and intellectualization of HCI and its components.
Mobile Ad hoc Network (MANET) attacks can be classified into active and passive attacks. Among active attacks, the rushing attack is one of the basic network layer attacks. In MANET, it early exploits the duplicate suppression mechanism of Ad hoc on-demand distance vector (AODV) protocol by quickly forwarding the RREQ packet to neighboring nodes without processing it, to influence a source node to include the rushed node in its route, which leads to data loss when transmitting the data packet to the correct destination node. This Early Rushing Attack Detection and Prevention in AODV MANETs (E-RADP) paper is proposed to fill this gap. To advance the security of our proposed algorithm, threshold value, ratio, and intermediate delay are introduced in rushing attack detection and prevention processes. For the performance analysis, the network simulator NS2.35 is used. The proposed protocol is compared with AODV, Prevention of Multiple Rushing Attacks Using AODV Routing Protocol (PMRA), and Rushing Attack Prevention with modified AODV (MAODV) based on true positive, true negative, false positive, and false negative values of normal and malicious nodes, and throughput, packet delivery ratio and end-to-end delay. E-RADP improves the rushing attack detection rate (DR), throughput, and packet loss rate (PLR) of existing protocols. It also improves the end-to-end delay (E2ED) of existing protocols when a rushing node is present in a MANET. Thus, the performance analysis shows that E-RADP is highly secure and faster than existing algorithms.
The proliferation of unstructured financial video content presents a significant challenge for traditional investment analysis. Natural Language Processing (NLP) offers a promising solution for extracting value from this data. This empirical study investigates whether NLP technologies can automatically extract, structure, and validate actionable investment insights from financial videos. We propose an automated pipeline using video transcription services and Large Language Models (LLMs). The methodology was tested on 22 YouTube financial analysis videos focusing on Amazon (AMZN) and Google (GOOG). The ChatGPT-4 model processed transcripts to extract stock tickers, risk levels, and price forecasts into a structured JSON format. The system achieved 100% accuracy in company recognition and filtering irrelevant content. Empirical validation against actual market data revealed an overall forecast accuracy of 85% (90% for AMZN, 70% for GOOG). This NLP approach also outperformed traditional ARIMA time-series models. The findings confirm that NLP can feasibly automate the analysis of financial video, transforming unstructured media into validated, structured data to support investment decisions.
Bandpass filters with a rectangular amplitude-frequency response (AFR) are relatively easy to implement in the frequency domain by zeroing out spectral components outside the desired passband. An advantage of this approach is the ease of constructing high-order non-recursive filters with a linear phase response. However, as demonstrated in this paper, this benefit comes at the cost of significant bias in reverberation time (RT) estimates, particularly in cases of narrow bandwidth and short reverberation time. It can be assumed that this drawback may be largely eliminated by using filters with a non-rectangular AFR, which is simple to implement in practice. However, the validity of this assumption remained untested until recently. In this paper, the influence of the filter’s AFR shape and bandwidth on the bias in RT estimation is analyzed. It is shown that the bias of RT estimates based on T20 and T30 ranges from 60% to 100% when one-third-octave filters with a rectangular AFR are used in the frequency range of 25–200 Hz. When a Tukey window is used as the filter AFR, the bias can be reduced to 4%. Similar results were obtained for Early Decay Time and T10 estimates of RT.
The paper considers the problem of recognizing objects (anti-tank mines, anti-personnel mines, M14, booby traps) located in soils with different structures (dry sandy, dry calcareous, dry humus, moist sandy, humid and humus, moist calcareous) using MLP with symmetric synaptic connections between neurons of different layers (SMLP) based on data from FLC100 magnetic field sensors with a sensitivity of 10-10 – 10-4 Tesla. To train the SMLP network, a dataset of fluxgate magnetometer measurements in six soil types was used, divided into training (80%) and test (20%) subsets. The input data is encoded as a three-dimensional vector (voltage, height of the sensor above the ground with a mine, one-hot encoding of 6 soil types). The main difference between SMLP and MLP is the imposition of symmetric constraints on the weight matrix of the second hidden layer, which almost halves the number of synaptic connections between neurons in the hidden layers without significantly degrading recognition quality. The proposed symmetrization mechanism mitigates over-parameterization and overfitting in small-scale magnetometric datasets by introducing structural inductive bias and implicit regularization through symmetry constraints in parameter space. In contrast to convolutional architectures developed for high-dimensional spatial representations, the proposed approach is tailored to structured low-dimensional magnetometric measurements. The MLP and SMLP models were trained for 100 epochs with Adam, RMSprop, and SGD optimizers with a learning rate reduction using ReduceLROnPlateau callback. Experimental results show that SMLP achieves an average accuracy of 99.20% and AUC = 0.9996, which is only 0.09% lower than the traditional MLP (99.29%, AUC = 0.9997), but reduces the training time by 20–30% and behaves more stably under different optimizers. Thus, the SMLP model is productive for use on embedded devices with limited resources.
One of the key challenges of modern e-learning is the timely identification of students at risk of expulsion in conditions of uncertainty and incompleteness of input information. The purpose of this work is to develop a mathematical model of e-learning system based on fuzzy logic that can predict the probability of successful completion of the course. The proposed model considers key factors of academic success, such as activity, time spent in the system, average score, attendance, participation in discussions, and test scores. To process input data, triangular belonging functions and a fuzzy rule base are used to formalize uncertainty in educational data. The centroid method is used as a defuzzification method. The paper formulates and solves the problem of predicting the expulsion of students using an adaptive neuron-fuzzy system. A numerical experiment was conducted on the data of 1000 students, during which the classification accuracy of 81.7% and the value of AUC = 0.90 were achieved, which confirms the high efficiency of the model. The practical significance of the proposed approach lies in the possibility of its integration into existing e-learning systems for the early identification of students at high risk of academic failure and subsequent adaptive management of the educational process aimed at reducing expulsions and increasing the effectiveness of e-learning.
Osteoarthritis is a degenerative joint disease that affects millions of people worldwide. Early detection and diagnosis of osteoarthritis is critical for effective treatment and management of the disease. In recent years, X ray imaging has emerged as a promising non-invasive technique for detecting osteoarthritis. However, existing techniques for osteoarthritis detection in thermal images suffer from several limitations, such as low accuracy, limited generalizability, and lack of interpretability. To address these challenges, we propose a novel approach for osteoarthritis detection in x ray images using the SqueezeNet model deep learning architecture. The proposed approach involves pre-processing the X-ray images to enhance their features, followed by segmentation to extract the region of interest. The segmented region is then fed into the SqueezeNet model , which is trained to classify the thermal image as normal or abnormal based on the presence of osteoarthritis. The parameter of SqueezeNet is tuned using grey wolf optimizer to reach maximum accuracy. We evaluated the performance of the proposed approach on a dataset of thermal images collected from patients with osteoarthritis and healthy controls. Our results show that the proposed approach achieved an accuracy of 94%, sensitivity of 97%, specificity of 92%, and an AUC of 0.94, outperforming several state-of-the-art approaches. We also conducted extensive experiments to investigate the impact of different pre-processing techniques and hyperparameters on the performance of the SqueezeNet model. Moreover, we conducted a detailed analysis of the learned features and identified the regions of the thermal image that were most important for osteoarthritis detection. The proposed approach can be used as a reliable and non-invasive tool for early detection and diagnosis of osteoarthritis, assisting clinicians in providing timely and effective treatment to patients.
This research presents a method and system architecture for ontology-driven processing of scientific natural-language texts that integrates a linguistic processor, a domain OWL ontology, a reasoning engine, and a knowledge graph. The system constructs a consistent knowledge graph suitable for the automatic interpretation of structurally complex user queries with subsequent transformation into SPARQL queries. The proposed pipeline includes formal mappings of linguistic analysis results to ontology classes, properties, and assertions; consistency checking; fact materialization; and an explanation mechanism that derives minimal sets of axioms and facts to justify the reasoning engine’s conclusions or to identify the causes of inconsistency. The output is a set of consistency-compliant subject–predicate–object triples. An ontology-consistency evaluation metric is introduced as the proportion of triples in the knowledge graph that do not violate ontological constraints, and the impact of logical processing on the harmonic mean of precision and recall for triple extraction is evaluated on a representative corpus of scientific texts in the field of knowledge engineering.
This research presents the development of a universal genetic optimizer (UGO) aimed at solving optimization problems across a wide range of test functions, focusing on achieving reliable global extrema for both theoretical and practical applications. The study utilizes a genetic algorithm (GA) enhanced with neural network-based approaches, implementing key genetic operators such as crossover, mutation, and elitism. The algorithm was tested on benchmark functions including Rosenbrock, De Jong, and Griewank, among others, with statistical analysis identifying optimal parameter settings. The results demonstrate superior performance compared to traditional tools like Excel Solver and GAMS, particularly when elitism is included. The proposed framework highlights the adaptability of evolutionary computation when combined with machine learning techniques. Moreover, it opens perspectives for applying UGO in real-world optimization challenges such as logistics, energy systems, and engineering design.
This paper presents a real-time augmented reality (AR) system for trying on outfits virtually to visualize apparel on the human body. An Augmented Reality-based visualization system is developed that can display chosen clothing styles and designs on a human body while in front of a camera. The objectives are to gather the requirements of the AR visualization system, design the system based on the requirements obtained, and implement, test and evaluate the performance of the system. Operational, non-functional and hardware requirements of the system were elicited via an extensive literature review. The AR system’s architecture was outlined and designed using UML tools like flowchart, use-case and activity diagrams. The apparel visualization system implementation leverages computer vision techniques, real-time image processing and augmented reality visualization. Python 3, Mediapipe v0.8.6.2, OpenCV and Visual Studio Code were utilized in the development of the system. The system was evaluated via beta testing using a survey method with a sample size of twenty-one (21), and a questionnaire tool. The evaluation was based on performance metrics including user experience, responsiveness, accuracy and usability, realism and performance in different lighting conditions. 66.7% had a “good” user experience, 86.7% of the respondents indicated that the system responded adequately, and 76.2% of users responded that the garments aligned correctly with points on their bodies. An average success rate of 48.42% and 8.9 FPS was recorded in bright lighting conditions while a 40.23% success rate and 8.23 FPS when it was darker.
This research explores the integration of neural networks into software protection mechanisms, focusing on enhancing cryptographic robustness against unlicensed copying and unauthorized access. The primary purpose is to develop a robust method that combines obfuscation and encryption to protect software by binding its activation and operation to specific hardware and user data, thereby preventing unlicensed replication and disassembly. The approach involves generating a unique hash of hardware identifiers, training a neural network on a remote server to produce a bytecode sequence for a virtual machine, and using the network’s weights as an activation code. Key results show that optimal neural network configurations, particularly those with two dense layers and one LSTM layer, achieve 100% accuracy in mapping predefined inputs to specific bytecodes within an average training time of 11 seconds, while generating pseudorandom outputs for all other inputs. Statistical analysis of output distributions reveals high entropy, demonstrating resilience against statistical attacks. The proposed method offers a layered defense against common attack vectors and ensures persistent security throughout the software’s lifecycle.
Zero-knowledge succinct non-interactive arguments of knowledge (zk-SNARKs) have emerged as a critical technology for privacy-preserving computation and blockchain applications. However, systematic performance analysis of practical implementations remains limited, hindering informed technology adoption decisions. This study presents a comprehensive benchmarking analysis of the Groth16 protocol implementation using the widely-adopted Circom-snarkjs framework. We developed an automated benchmarking platform that systematically measures performance across seven representative circuit types with varying computational complexity (1-11 R1CS constraints). Our methodology ensures reproducible measurements through controlled experimental design with statistical validation. The platform captures detailed metrics for all three phases of the Groth16 protocol: witness generation, proof creation, and verification. Results from 35 independent measurements reveal several important findings. Witness generation demonstrates consistent performance across circuit types, averaging 57.6±12.1 milliseconds. Proof generation times range from 832 to 1,147 milliseconds, showing non-linear scaling with circuit complexity. Verification times remain relatively stable (741-884 milliseconds), confirming Groth16's theoretical constant-time verification advantage. All measurements achieved 100% success rate with complete proof validation. Notably, circuit structure significantly impacts performance beyond simple constraint counting. Comparison-based circuits achieve 13.22 constraints per second efficiency, substantially outperforming arithmetic circuits (1.02-4.36 constraints/second). This finding provides actionable guidance for circuit design optimization. The study contributes an open-source benchmarking framework for reproducible zk-SNARK research and provides empirical performance data for technology adoption decisions. Our findings support the practical deployment of Groth16 for applications requiring efficient zero-knowledge proofs while highlighting optimization opportunities for circuit designers.
