ABSTRACT. The increasing reliance on interconnected digital systems has intensified the need for authentication mechanisms that extend beyond traditional, static methods such as passwords or one-time biometric verification. These conventional approaches are vulnerable to credential compromise and are limited in their ability to detect unauthorized access after initial login. As a result, continuous authentication based on behavioral biometrics has gained attention as a complementary security mechanism. This paper, which is the initial part of a broader research, presents a focused review of behavioral biometrics for continuous authentication, with particular emphasis on keystroke dynamics. The review discusses the fundamental principles of behavioral authentication, the temporal features commonly extracted from typing behavior and the machine learning techniques typically employed to model user-specific patterns. To complement the overview, the paper also describes an illustrative JavaScript-based implementation for collecting keystroke dynamics data, detailing the capture of key press and release events, timing intervals and key hold durations. The structure and interpretation of the resulting dataset are examined to demonstrate how raw interaction events can be transformed into behavioral features suitable for authentication models.

ABSTRACT. Ordered relational queries do not require a model to retrieve a single stored edge. Instead, they require following relations in a specific order to return the final endpoint. A correct endpoint does not confirm that the model executed the traversal, since the same answer could be guessed, copied, or produced through a hidden shortcut.

We introduce a single-read diagnostic protocol. The model receives facts and a path query, builds a temporary associative field for that query, and the answer must be obtained in one read. Repeated reads, candidate search, and reader-side correction are not allowed. This setup distinguishes value selection, field writing, ordered traversal, and rejection under a wrong key.

Experiments show that the read path is not the main issue. When the correct value is known, exact-write and clean one-hop conditions yield the answer in one read. Instability shows up earlier. The model must choose the right endpoint from the complete set of distractors before writing it. This failure occurs in one-hop full-distractor queries, before multi-hop composition becomes necessary. Longer training helps some runs, but does not stabilize the condition across random seeds.

In this context, path length is not the main limitation as long as source-relation-target roles are available. An explicit step-by-step transition procedure solves chains up to length 32, including the supervised slot-grounded version. A separate boundary emerges after writing. A wrong key can retrieve the written endpoint through address overlap. Endpoint accuracy establishes endpoint reachability under the allowed read, but it does not identify the mechanism. Value selection before writing, ordered composition, canonical relation-label alignment, and wrong-key rejection remain separate checks.

ABSTRACT. The current work investigates the discovery of causal relationships in time series data through a deep learning approach. Traditionally, uncovering such dependencies requires extensive analytical work, including model construction, solving differential equations, hypothesis testing and relational analysis. These processes can be automated using transformer-based models with a large number of training parameters, such as CausalFormer. The main contribution of this work is the adaptation of this model to support incremental learning, allowing it to update with new data without full retraining on the entire dataset. This is particularly valuable in real-world applications where data arrives continuously and resources for retraining are limited. Using a functional magnetic resonance imaging (fMRI) dataset, a comparative analysis between full and incremental training was performed, which showed similar results between the two methods. This indicates that the combination of interpretable deep learning models and incremental learning provides an effective and adaptive alternative for causal analysis in real time and under continuously evolving data conditions. This opens new opportunities for applying such models in dynamic domains like environmental monitoring, healthcare, finance and beyond, where it is crucial not only to identify dependencies, but also to ensure that models are adaptive and interpretable by domain experts and end users.

ABSTRACT. The continuing development of large language models opens up significant opportunities for the analysis of online conversations. However, their usefulness is challenged when their responses are not grounded in the evidence of the social media data. Retrieval-Augmented Generation can help in this regard by providing text as evidence, but many questions related to social network analysis require relational information such as user interactions, communities, centrality measures, and diffusion patterns. In this paper, we propose an intent-driven GraphRAG framework that enables LLM-based analysis of online conversations and addresses the aforementioned limitation. The proposed framework transforms a social interaction graph into a post-centered knowledge graph by making users, posts, hashtags, URLs, domains, entities, communities, and sentiment first-class graph elements rather than simple attributes in a user-user interaction graph. Several analytical intents are derived from the literature, including topic overview, sentiment analysis, influential-user detection, community comparison, information diffusion, temporal trend analysis, and actionable-opportunity identification. User questions are mapped to these intents and predefined Cypher queries retrieve the appropriate information from Neo4j, which is then provided to the LLM as evidence. We evaluate the framework using a COVID-19 vaccine dataset constructed from X (formerly Twitter) by comparing three approaches: vanilla LLM prompting, text-based RAG, and the proposed GraphRAG method. The results of the comparative experiment show that GraphRAG achieves strong results primarily because of the structural nature of the questions. Text-RAG provides a strong baseline for textual grounding, but remains limited for questions requiring structural reasoning. Vanilla LLMs provide ungrounded general responses or refuse to respond due to lack of evidence. These findings suggest that knowledge-graph-based retrieval can improve the reliability, interpretability, and analytical usefulness of LLMs for social media analysis.

ABSTRACT. Most flight-delay prediction systems rely on real-time meteorological data, which are rarely available in large historical aviation archives. This paper shows that competitive performance is still achievable using only historical operational data. We train a cost-sensitive CatBoost classifier on a 240-million-record dataset of U.S. domestic flights, under strict chronological validation, and introduce a Temporal Undersampling strategy that addresses the severe class imbalance by removing the oldest On-Time records rather than random ones. The proposed approach outperforms four baseline imbalance-handling strategies and, on a balanced evaluation set, exceeds the best F1-Score reported by a recent IEEE DASC reference study that relied on weather data.

ABSTRACT. This paper presents a new approach to facial emotion recognition (FER), as a component of human- computer interaction (HCR), with applications in healthcare, psychology, security, marketing, e-learning, and humanoid robot design. The proposed model is based on using an ensemble of three independent Deep CNN classifiers followed by a decision fusion based on Random Forest (RF) method. We have considered two variants for choosing the three component CNN modules. First variant uses AlexNet, VGG16, and ResNet18 as CNN modules; second variant of decision fusion model uses three ResNet18 modules with identic architectures, trained independently with the same dataset, but using different initialization seeds. For experimental evaluation, we have chosen CK+ and FER2013 datasets. We have obtained an improvement of accuracy for RF fusion model by comparison to the case of using any of the standalone classifies, for both considered decision fusion model variants, as well as for both chosen datasets.

ABSTRACT. Large language models (LLMs) have transformed natural language processing through their scale and complexity, driven by the use of enormous corpora and resources. However, these data introduce noise and biases that limit performance in specialized domains. To overcome these constraints, domain-specific curated datasets are used to enable more efficient fine-tuning, although this poses challenges in terms of data availability, the balance between specialization, generalization, and computational cost. This paper presents a finetuned version of the BERT-base-multilingual-cased model on our proprietary dataset with different sample sizes to answer the following questions: How does the size of the training set affect the model's performance in the multiclass classification task with the available hardware? Furthermore, at what training epoch does the model begin to overfit for each of the evaluated dataset sizes? And how does the number of classes affect the performance of the model? The results establish a set of consistent patterns in the behavior of the model. As the amount of data increases, F1-score measures above 0.82 are obtained, followed by a reduction in validation loss. The model shows effective learning during the first five epochs, regardless of the size of the dataset. A progressive degradation in performance is also observed alongside an increase in computational cost as the number of classes increases.

ABSTRACT. Electoral accessibility is one of the most significant hurdles to democratic participation, particularly users who are visually impaired and encounter numerous challenges in navigating digital ballot interfaces. Today’s accessible voting systems are basic support for static audio and screen-reader interfaces, which do not consider user abilities, preferences, or interaction behavior. In this paper, we introduce personalized ballot navigation as driven by machine learning approach which dynamically modifies interface features and interaction flows through real-time user modeling. We predict that machine learning-driven personalization minimizes interaction errors and time of completion when compared to static interfaces. In order to test this under realistic conditions, we built a semi-synthetic dataset of visually impaired user interaction profiles. This was a needed first step to fill the void of available public datasets for this community. From the extant literature, we incorporated the controlled feature noise that simulated the real-world factors (e.g., user fatigue, tactile hesitation) to describe behavioral variation. We trained three supervised learning algorithms (Random Forest, Logistic Regression, and Decision Tree) on features (interaction speed and error pattern). Finally, all models performed reasonably well in noisy conditions while none achieved a perfect result. It was Random Forest that exhibited the best robustness. The simulated findings demonstrate that the adaptive ML-based interface decreases average ballot completion time by around 28% and interaction error rates 49.7%. To be sure, while these gains are huge, they may actually be quite different across different hardware. These results imply that machine learning-enhanced personalized voting can meaningfully enhance electoral inclusivity

ABSTRACT. The linear eddy viscosity hypothesis, fundamentally embedded in standard Reynolds-averaged Navier-Stokes models, frequently fails to capture the complex Reynolds stress anisotropy present in flows with separation, curvature, or strong pressure gradients. To address this limitation, this study introduces a physics-informed, data-driven framework to accurately predict the Reynolds stress anisotropy tensor using high-fidelity direct numerical simulation data. The methodology leverages the general effective-viscosity hypothesis proposed by Pope to map the local mean velocity gradients into a physically invariant feature space. Five objective scalar invariants are extracted and used to train a Random Forest regressor, which predicts the corresponding scalar coefficients for a ten-term orthogonal tensor basis. The target coefficients for training are obtained via a mathematically rigorous least-squares projection of the true high-fidelity anisotropy onto the baseline tensor bases. The framework's predictive performance is evaluated not only through standard statistical error metrics but also through stringent physical realizability constraints. Barycentric maps demonstrate that the model successfully confines its predictions within physically permissible turbulence states. Furthermore, near-wall profiling of the primary shear stress across the viscous sublayer, buffer layer, and log-law regions confirms the model's ability to correct baseline fluid assumptions. This framework offers a robust, computationally efficient, and physically consistent pathway to augment standard solvers with highly accurate predictions. Our Model 4 achieves R^2 > 0.99 for active anisotropy components with 44% reduction in spatial RMSE compared to coordinate-based baselines. Lumley tringle map confirms 100% physical realizability

ABSTRACT. This paper presents a pipeline for analyzing media bias and framing in online news. The pipeline groups articles into topics and events, adds named-entity and sentiment annotations, and compares news sources through people mentions, source-level tone, and event-level coverage patterns. We apply it to 8,358 Albanian news articles collected from GDELT and compare the resulting annotations with GDELT's automated annotations. The results show moderate agreement for sentiment and entity extraction, as well as additional person-entity pairs that can potentially support the bias analysis. We compare two annotation prompts and find that stricter sentiment-validation rules remove label-score inconsistencies but increase execution time and reduce annotation coverage. Based on these results, the simpler prompt is used for the rest of the analysis. We have provided sample analysis on source-level framing profiles, person-level tone differences across sources, and event-level gatekeeping and coverage indicators. These outputs show how the same news collection can be used to examine what sources cover, how they describe public figures, and where coverage is concentrated. The approach is particularly useful in settings where manually verified datasets or specialized language tools are limited.

ABSTRACT. This paper first describes the development of a small, remote-controlled, mobile robot designed primarily to find victims in areas affected by natural disasters. The robot is equipped with an integrated video camera that can capture and send a continuous stream of images at speeds up to 60 FPS in real time. The camera position can be adjusted remotely and independently of the robot’s movements. The most valuable feature of the mobile robot is its wireless connection, which facilitates communication with servers for statistical data and with the person controlling the robot. The paper also suggests the direct tuning of anti-windup Takagi-Sugeno-Kang (TSK) Proportional-Integral (PI) fuzzy controllers for angular position control of the two front wheels using Iterative Feedback Tuning (IFT). A back-calculation and tracking anti-windup scheme is included in order to mitigate integrator windup while compensating for the process dead‑zone and saturation effects. The direct data-driven tuning approach performs experiment-based parameter tuning of the TSK PI fuzzy controller and anti-windup tracking gain in a few iterations of IFT. The experimental results demonstrate the application of the data-driven fuzzy control approach and represent a step forward in designing low-cost rescue mobile robots.

ABSTRACT. As the population increases, parking lots are swarming with cars, leading to collisions and long lines of people waiting to park. To solve this issue, this paper proposes implementing a small autonomous electric car, which is simple yet effective. The paper presents few hardware and software aspects. It also suggests a data-driven tuning approach for low-cost (l.c.) Proportional-Integral (PI) fuzzy control of the angular position of the two front wheels using Iterative Feedback Tuning (IFT). The paper further presents a back-calculation and tracking–based anti-windup strategy designed to prevent integrator windup and to compensate for the process dead zone and saturation nonlinearities. PI fuzzy controller tuning is carried out in two stages: (i) tuning the parameters of PI fuzzy controller’s fuzzy logic subsystem and the anti-windup tracking gain in a model-based manner using the improved hybrid Particle Filter-Particle Swarm Optimization algorithm with an added information feedback model, and (ii) tuning the parameters of PI fuzzy controller’s linear subsystem in a data-driven manner using IFT and mapping them onto the remaining PI fuzzy controller parameters using the modal equivalence principle. Experimental results illustrate the application of the data-driven fuzzy control approach, representing a proof-of-concept validation of a data-driven control approach for l.c. self-parking cars i.e. autonomous vehicle prototypes.

ABSTRACT. In this paper we present a robotic control system based on hand gestures, inspired by the master–slave surgical teleoperation principle. The system is implemented on a Rasp-berry Pi 5 platform, equipped with a Hailo-8 neural accelerator, and enables real-time hand tracking as well as the extraction of relevant gesture features. For interpreting hand movements, we use a geometry-based method that extracts positional and angular features directly from hand landmarks detected in real time, associating them with continuous servo targets for each robot joint. These commands are transmitted via USB serial to a robotic arm (TinkerKit Braccio++), which reproduces the operator’s movements. The developed system demonstrates the feasibility of using edge-AI technologies for gesture control in surgical robotics and provides a low-cost, fully open-source platform, useful both for educational purposes and for research in the field of human–robot interaction.

ABSTRACT. Smart electrical protection devices design and development require rigorous strategies that integrate surveillance over voltage fluctuations, overload conditions, and short circuits with the main goal to avoid or prevent equipment damage, reduced lifetime, and safety hazards. Reliable protection mechanisms are therefore essential in industrial, but also consumer applications. In this context, this paper proposes a smart electrical protection device based on the STM32G030 microcontroller, capable of real-time monitoring of electrical parameters and fast response to fault conditions. The system continuously measures voltage and current and provides rapid protection by disconnecting the load through a MOSFET switch. For analog signal acquisition, an internal ADC with DMA support is used, while FreeRTOS ensures real-time operation through dedicated monitoring and communication tasks. A Bluetooth HC- 06 module enables wireless communication, allowing users to remotely monitor electrical parameters and control the protection mechanism. Experimental results demonstrate response times between 127 μs and 1065 μs, highlighting significantly faster protection compared to conventional fuses.

ABSTRACT. This This paper proposes a smart rotating-house concept designed to improve energy efficiency of residential buildings through continuous solar tracking position. The proposed structure is designed for buildings that rotate according to the sun's position, raising solar energy usage and thermal gain throughout the day. To minimize the overall structural and dynamical load, lightweight construction materials, including dome-shaped glass-based shells, are considered. The main goal is to reduce the total weight of the building. Another criterion used in this work is the integration of a meteorological algorithm based on artificial intelligence (AI). Processed inputs are obtained from the Metar service Brașov’s airport, with the ICAO-LRBV [11], which is situated at the base of the Tâmpa mountain. Using a control algorithm System - Alpha Fractional Order (FOPID), and one dynamic Positioning System with two axes, the Solar route could be followed, and the rotation could be adjusted depending on the weather forecast in real time. The walls of the building will be made from glass with the ability to clean itself and to collect water with some special tubes situated at the base of the building. This water can be used for household water and can be warmed by Sun with vacuum tubes. Simulation results demonstrate that the FOPID controller reduces peak overshoot by 98%, while the RMS tracking error by 31%. At the same time the control energy consumption is minimized by 18% in comparison with a conventional PID controller, while solar energy is raised by up to 56% through continuous solar tracking.

ABSTRACT. This paper presents an intelligent modular system for the unannounced metrological verification of taximeters under real vehicle operating conditions. The system verifies time and distance indications by combining a mobile roller-based mechanical module, data acquisition and processing software, robotic visual assistance and IoT communication. The correspondence between the indication displayed by the taximeter and the value calculated by the verification system is established by taking into account the vehicle characteristics, the actual profile of the drive wheel and the correction coefficient associated with the support conditions on the rollers. Verification data are processed locally in order to issue a verification bulletin and are then transmitted securely to a dedicated server, where georeferenced information can populate public maps of the areas in which taximeter checks were performed. Field testing on 48 vehicles demonstrated the feasibility of the approach, with two vehicles being referred for laboratory metrological verification.

ABSTRACT. This study introduces a high-performance motion controller specifically engineered for multi-axis articulated manipulators. By harnessing the heterogeneous computing synergy between a System-on-Chip (SoC) and an FPGA, we realized a synchronous 6-axis motion control framework. In this architecture, the SoC serves as a dedicated engine for real-time forward and inverse kinematics, while the FPGA functions as a hardware accelerator to execute parallel, closed-loop position and velocity control across all motor axes. This distributed control strategy effectively decouples real-time kinematic computations from the PC-based host, which is then redirected to focus on high-level spatial trajectory planning. These planned trajectories are subsequently discretized by the SoC and streamed to the FPGA for synchronized execution, ensuring the end-effector meets rigorous precision standards.

ABSTRACT. Ensuring the security of medical imaging has garnered increasing attention in recent years; this interest comes within the context of smart healthcare systems. On the one hand, different encryption methods can be taken into consideration. Nevertheless, most of them have been designed for execution on PCs that suffer from serious disadvantages such as portability limitations, high energy consumption, and high costs. In order to overcome such limitations, hardware approaches have become in focus, offering real-time processing and enhanced security along with energy-efficient performance using embedded platforms and parallel computing architectures. We hereby contribute a novel and optimized compression-encryption scheme specifically designed for execution on Raspberry Pi. The proposed approach integrates three complementary phases: (1) image compression using discrete orthogonal Hahn moments for compact and faithful representation; (2) the encryption is based on a 4-D hyper-chaotic memristor system, a modified logistic map, and DNA encoding to offer strong defense against a variety of threats and a high degree of security; and (3) in addition, an optimization algorithm, namely Chaotic Archimedes Optimization Algorithm (CAO), is used to dynamically minimize the correlation between encrypted and original data. The design of this architecture provides a very secure, real-time process and low energy cost; therefore, it is an applicable and scalable solution for secure medical image transmission and storage in embedded and mobile health systems.

ABSTRACT. This study compares several metaheuristic optimisation algorithms for planning 3D drone trajectories in mountainous environments, using cubic spline interpolation. The trajectory is parameterised by intermediate control points and optimised using a multi-criteria cost function that takes into account path length, safety distance from the terrain, altitude compliance and flight fluidity. A realistic, non-linear relief model was used to create a complex optimisation landscape. Six algorithms (POA, PSO, GGO, HHO, SMA and RUN) were evaluated under identical conditions, with a population of 50 agents and 500 iterations. The results show significant performance differences. The Slime Mould Algorithm (SMA) achieved the best results with a minimum cost of 275,822, ahead of PSO (351,546), HHO (489,276), POA (489,513), RUN (655,739) and GGO (793,585). Convergence analysis and 3D trajectory visualisation confirm that adaptive exploration-exploitation mechanisms significantly improve performance in constrained environments. Conclusion: advanced metaheuristics are particularly effective for continuous 3D spline trajectory planning in complex mountainous terrain.

ABSTRACT. This paper presents a three-level gallium nitride (GaN) inverter for a residential heat-pump compressor drive. The proposed inverter is based on the active neutral-point-clamped (ANPC) topology and is integrated into a complete power-conversion system comprising a grid-side SiC active rectifier, a photovoltaic (PV) interface through a SiC DC/DC converter, and a common DC bus. The main motivation is to combine the reduced voltage stress and smaller output-voltage steps of multilevel conversion with the high switching speed and power-density potential of GaN devices. A hardware prototype was built and experimentally tested. The measurements include switching-transient evaluation of the GaN devices, for which a turn-off time of approximately 25 ns was obtained, and an investigation of parasitic ringing and oscillations. The results show that unsuitable turn-on and turn-off settings, together with an unfavorable PCB layout, can lead to oscillations around 23 MHz and may degrade switching performance and device reliability. The inverter was also verified in motor-drive operation, first with an induction motor under V/f control and subsequently with a permanent-magnet synchronous machine (PMSM) under sensorless vector control. The experimental results demonstrate the feasibility of the prototype and identify DC-bus capacitor voltage imbalance and switching-transient sensitivity as the main issues requiring further optimization.

ABSTRACT. Early defect forecasting enables preventive interventions in extrusion additive manufacturing before visible failures become irreversible. This paper presents a multi-sensor acquisition and labeling pipeline that couples synchronized process telemetry with a vision-based defect detector to produce horizon-based early-warning labels, together with four calibrated predictive models that output defect risk at 5 Hz. A temporal convolutional network achieves ROC-AUC 0.93 with 1.2 ms inference latency on a Raspberry Pi 5, offering the strongest deployment tradeoff; a lightweight Transformer reaches 0.94 at a higher compute cost. Cross-printer evaluation shows that fine-tuning on approximately 10% of target-domain data markedly improves zero-shot transfer. A cost-sensitive threshold policy converts calibrated probabilities into actionable binary alerts.

ABSTRACT. In the first part of the paper, the areas of interference effects of pulse rectifiers on the power supply network are studied in the context of technical standards, and the importance of the LCL filter on the power supply network side is presented. In the second part, the paper presents computational procedures for the design of the parameters of the input LCL filter. The next part of the paper deals with the possibilities of modifying the calculated parameters of the LCL filter so that it is possible to reduce the dimensions, weight, and price of the filter while maintaining sufficient filtration ability. In the final part, the influence of some parameters and conditions of LCL filter operation on the input current spectrum is analyzed. The degree of filtering effects is assessed according to the frequency spectrum of the input current and in the context of technical legislation for the low-frequency range.

ABSTRACT. This paper presents the results of computational analyses of the output voltage parameters of several types of multiphase inverters. The focus is on the analysis of output voltages under rectangular control and under PWM with triangular reference comparison. Five-phase and nine-phase inverters are analyzed, and some comparisons with three-phase inverters are also included. The study primarily examines the possible values of the RMS output voltages and the total harmonic distortion (THD) of the inverter output voltages.

ABSTRACT. This paper presents the design, mathematical formulation, and real-time implementation of a linear Active Disturbance Rejection Controller (ADRC) for speed regulation of a Permanent Magnet Synchronous Motor (PMSM). The proposed ADRC employs a Linear Extended State Observer (LESO) to estimate and actively compensate for total disturbances, including parameter uncertainties, unmodelled dynamics, and external load torque variations, without requiring an accurate plant model. The controller is designed in MATLAB/Simulink and deployed on a Texas Instruments C2000 LaunchPad (TMS320F28379D) using the Embedded Coder support package for real-time execution. A comprehensive mathematical model of the PMSM in the d-q reference frame is presented, followed by the complete ADRC formulation including the LESO design and bandwidth parameterization. The proposed ADRC is experimentally compared against an optimally tuned PID controller under step reference tracking, load disturbance rejection, and speed reversal tests.

ABSTRACT. This study proposes a new approach based on the incremental conductance (INC) maximum power point tracking (MPPT) method. The proposed approach controls a single-stage solar water pumping system (SWPS) driven by a brushless DC (BLDC) motor. Its idea is to divide the power-voltage (P-V) curve of a photovoltaic (PV) system into an optimal number of bands or segments and then use INC-MPPT to identify local peaks and select the global peak. It searches within each band to capture the global maximum power point (GMPP) efficiently and accurately track it, rather than continuously scanning the full range. This approach avoids the complexity of a global search and achieves fast and accurate tracking under rapidly changing environmental conditions (e.g., cloud shading). An indirect MPPT control structure is used, employing a proportional-integral-derivative (PID) controller to regulate the motor speed smoothly and adjust the PV voltage level for each MPPT voltage setpoint. The system model was successfully tested under different weather conditions using MATLAB simulations. The BLDC motor operates efficiently, achieving an MPPT efficiency of more than 99%.

ABSTRACT. In this paper we present our implementation of the behavioral models for the SEPIC converter in MATLAB/Simulink® and we compare the response with the results in LTSpice® in order to check if this setup is suitable to design such a DC/DC converter. Additionally, we make use of MATLAB/Simulink® numerical capabilities to analyze the changes that take place in the poles and zeros positions due to operating conditions. Based on the state-space analysis, we derived the transfer functions of the SEPIC and simulated it in MATLAB/Simulink®. Transfer function parameter comparison vs behavioral model results confirm method validity.

ABSTRACT. With the increasing electricity demand experienced at global level, it has become mandatory for all the parties involved in its production and distribution to adopt efficient and responsible methods of energy management. In the past few decades, research in the field of electrical engineering focused on non-intrusive load monitoring (NILM) methods, aiming to simplify the characterization of local grids. The proposed paper approaches the load identification phenomena as a problem of optimization and uses genetic algorithms (GA) to correctly identify the combination of connected electrical consumers from an aggregated current signal.

ABSTRACT. In this study, a novel method for speed control of six-phase asymmetric interior permanent magnet synchronous motors (IPMSMs) is proposed using the fuzzy fast terminal sliding mode control (FFTSMC) technique. Moreover, field-oriented control (FOC) technique is applied to the speed control structure of six-phase asymmetric IPMSM. To evaluate the performance of FFTSMC, its results are compared with two other controllers, including conventional sliding mode control (SMC) and classical proportional-integral (PI) controllers. Simulations are performed in MATLAB™/Simulink® environment and under different speeds and suddenly load-torque conditions. The results show that the proposed FFTSMC has a lower finite-time convergence in speed tracking than other controllers while maintaining tracking accuracy under load conditions. Also, this method reduces total harmonic distortion (THD) under different load-torque and speed conditions.

ABSTRACT. Abstract— The integration of green energy sources and Storage solutions into the grid system has become highly Embraced in past several years. The nature of these distributed power technologies being plug-and-play in nature causes different types of disturbances and this needs to be prevented through high efficiency, strong effectiveness, and rapid responsiveness energy management (EM) techniques or systems (EMS). Different EMS have been investigated to date by number of research scientists. In this paper, we present a Finite State Machine (FSM)-based control strategy for power system management solution in a hybrid DC microgrid connected to photovoltaic (PV), wind integration, battery storage, and utility grid. This strategy relies on maintaining stability within the microgrid while maximizing the percent of renewable generation being utilized using a logical state transition topology based on actual energy balance and the battery's state of charge. The FSM transitions through operational modes, for example, connected to the grid import or exporting power and islanding mode with renewable only, charging and discharging states. The FSM provides operational dynamic governance in controlling the microgrid's power supply when loads change and environmental inputs change. A MATLAB/Simulink model was developed to validate the proposed control logic. The proposed control logic is a low-complexity, reliable, scalable, solution for microgrid control while allowing for future layers of supervisory control based on prediction and optimization.

ABSTRACT. This paper investigates a Photovoltaic Water Pumping System (PVWPS) driven by an induction motor (IM) controlled through Direct Torque Control (DTC). The study focuses on enhancing overall system performance by comparing two Maximum Power Point Tracking (MPPT) algorithms: the Incremental Conductance (INC) algorithm and the Kalman Filter-based MPPT (KF-MPPT) algorithm. The INC algorithm is widely used due to its simplicity, but it often suffers from steady-state oscillations and a slow response under rapid irradiance variations. In contrast, the KF-MPPT aims to enhance tracking accuracy and dynamic performance by estimating the optimal operating point more precisely. The DTC technique is employed to control the electromagnetic torque and flux of the IM, ensuring efficient water pumping operation. The complete PVWPS was modeled and simulated in MATLAB/Simulink under variable solar conditions. Comparative results demonstrate that the integration of KF-MPPT with DTC significantly reduces power oscillations, improves transient response, minimizes torque and flux ripples, and increases the pumped water flow compared to the INC -DTC combination.

ABSTRACT. Green hydrogen produced using renewable energy sources has attracted growing interest serving as a sustainable solution for reducing greenhouse gas emissions and supporting the global energy transition. Among renewable technologies, photovoltaic (PV) systems coupled with proton exchange membrane (PEM) electrolyzers stand out as an effective alternative for hydrogen generation from solar energy. This paper investigates the impact of PV module technology on the solar-to-hydrogen (STH) efficiency of a PV-powered PEM electrolyzer system. Three PV modules were analyzed: Canadian Solar CS6X-300P, Jinko JKM300M60, and Jinko Solar JKM400M-72HL-V representing the polycrystalline, monocrystalline and Mono-PERC technologies, respectively. The analysis was performed over a full year using real climate data, including solar irradiance and temperature profiles. The simulation results indicate that PV module technology significantly affects the STH performance of the system. The STH efficiency varies from 10.49% for the polycrystalline module to 12.42% for the monocrystalline module, reaching 13.54% with the Mono-PERC module. These results demonstrate that selecting high-efficiency PV technologies plays a key role in improving the overall performance of PV-powered hydrogen production systems.

ABSTRACT. Achieving precise and robust speed regulation in permanent magnet synchronous motor drives remains a significant challenge due to the inherent nonlinearity and strong coupling of the motor dynamics combined with unavoidable parametric variations and external load disturbances. This paper proposes an integral backstepping controller that addresses these challenges through a recursive Lyapunov-based design applied to the third-order PMSM model in the d-q reference frame. Integral action is embedded exclusively within the speed control loop to eliminate steady-state speed error under sustained disturbances, while the current control loops retain a compact and tractable structure. A single composite Lyapunov function is constructed to encompass all error variables including the integral state, and its time derivative is proven strictly negative definite, establishing the global asymptotic stability of the overall closed-loop system. MATLAB/Simulink simulation results obtained under speed tracking, load disturbance rejection, and variable-speed operation scenarios confirm overshoot-free transient behavior, negligible steady-state speed error, and stator current THD values of 0.23% at 6.25 Hz and 0.13% at 25 Hz, both compliant with the IEEE 519 standard, demonstrating the accuracy, robustness, and practical implementability of the proposed controller for high-performance PMSM drive applications.

ABSTRACT. Increased demand for clean and sustainable sources of energy is encouraging the use of hybrid energy systems with high technological potential. Photovoltaic/thermal systems are considered to be a highly efficient hybrid system as they produce electrical as well as thermal power using the same surface. To increase the efficiency of these systems, accurate predictive tools must be developed to analyze the effect of changing environmental conditions. In this context, the use of data-driven models is considered to be an efficient alternative to traditional models. In the present study, the authors have proposed a data-driven model to predict the thermal efficiency of an air-based photovoltaic/thermal system using the concept of ensemble machine learning models. The authors have considered the Random Forest, XGBoost, and CatBoost models to predict the thermal efficiency of the system. The results indicate that the Random Forest model is the best model to predict the thermal efficiency of the system with R² value 80.873% and MAE 2.751%. XGBoost and CatBoost models have also shown good results with R² value 80.173% and 80.3% with MAE 2.890 and 2.948 respectively. Moreover, the results indicate that the feature ‘global horizontal irradiance’ is dominant in determining the efficiency of the photovoltaic/thermal system.

ABSTRACT. This study proposes an intelligent Neural PI–based DPC (NPI-DPC) strategy for DFIG wind energy systems. In this approach, the classical linear PI controllers are replaced by neural network regulators, that are integrated into the DPC-PWM control structure in order to adapt the PI gains Kp and Ki automatically during operation to improve the dynamic performance of the DFIG system under variable wind conditions. The neural controller adjusts the control gains in real time, which helps to reduce active power ripples, decrease electromagnetic torque oscillations, and limit the THD of the stator current. The performance of the proposed NPI-DPC strategy was evaluated using MATLAB/Simulink simulation and compared with the conventional PI-DPC control method. The results show that the neural-based controller provides better tracking of active and reactive power under step wind speed variations, and it maintains good decoupling between active and reactive power, even in the presence of nonlinear aerodynamic disturbances. Also it reduces active power ripples and electromagnetic torque oscillations compared with the classical PI-DPC.

ABSTRACT. Industrial Surveillance in high-risk environments such as data centers often relies of manual inspection or static camera systems with limited coverage. Ensuring workers safety in hazardous areas such as battery rooms, generator zones, and cooling units, requires strict compliance with Personal Protective Equipment (PPE) regulations defined by standards including ISO 20471 and ISO 45001. This paper presents an autonomous vision-based robotics system for continuous PPE compliance monitoring in data center environments. A custom line-following patrol robot equipped with an RGB camera streams video through a ROS2 pipeline to a YOLOv8l detection model trained on a dataset of 5,760 images across all the classes. The model achieved a mAP@50 of 96% and mAP@50-95 of 78%. When violations are detected, the system sends notifications to a monitoring dashboard and emits and audio alert. The feasibility of employing mobile robotic surveillance in industrial safety monitoring was validated through a Gazebo-ROS2 simulated data center environment.

ABSTRACT. The rapid proliferation of Internet of Things (IoT) devices in smart agriculture has exposed critical vulnerabilities in irrigation monitoring infrastructure, particularly for time-series sensor streams transmitted over Low-Power Wide-Area Networks (LPWAN). This paper presents a comprehensive review of cybersecurity challenges and countermeasures for agricultural IoT systems, with emphasis on three interconnected domains: (i) TinyML-based anomaly detection for on-device security of irrigation time-series data, (ii) hierarchical edge-fog-cloud security architectures for deployment of intrusion detection and authentication mechanisms, and (iii) communication-layer hardening for Wireless Sensor Networks (WSN) and LPWAN protocols including LoRaWAN, NB-IoT, and ZigBee. We systematically survey attack taxonomies across all IoT layers, from false data injection and replay attacks on sensor streams to cloud-layer ransomware and supply chain threats. Drawing on a curated corpus of recent literature, we identify open gaps in standardized security benchmarks, lightweight cryptographic protocol design, and TinyML model validation under adversarial conditions. Our analysis highlights the emerging role of federated learning and blockchain-based authentication as scalable solutions for distributed agricultural deployments and identifies explainable AI (XAI) as an essential complement for operator trust and auditability of on-device anomaly detection decisions.

ABSTRACT. Recently, WECS has achieved more than 11% of the world's energy demand. With this growing demand, wind turbine systems require a high-efficiency technique to extract the maximum wind power. However, this new study remains a challenge and proposes a comparative performance analysis of MPPT controllers based on ANN and ANFIS with PI. These techniques aimed at enhancing the power extraction. All the MPPT approaches were developed through simulation in the MATLAB/SIMULINK environment under realistic wind speed data compared with PI. The intelligent strategies demonstrated significant improvement. The obtained results show that both intelligent techniques achieve fast convergence time by approximately 70%, and reduce the steady-state error by more than 90%. These findings confirm the effectiveness of the AI-based technique in improving the overall system performance.

ABSTRACT. For the long of its history, power grid relied on large electrical generation plants with stable and predictable output. The integration of renewable energy changed this. With wind and solar injecting intermittent power from thousands of distributed nodes increasing the variables that the grid could manage. Consequently, advanced algorithms such as model predictive control and neural network based control become necessary to maintain stability and performance under such conditions. In this context, this paper explores efficient computing solutions for implementing these control algorithms into Field Programmable Gate Arrays (FPGAs). FPGAs are reconfigurable integrated circuits capable of offering high-performance parallel processing at the lower cost of energy consumption. Three hardware architectures for MVM are compared, with focus placed on the trade-off between processing latency and silicon area. Asymmetric quantization is also studied to assess its effect on hardware resource usage and model accuracy. Afterward, the suitability of each architecture for renewable energy integration is discussed. The broader goal of this study is to promote the adoption of FPGA-based implementations as a practical path to low-latency, intelligent control at the grid edge, increasing the computational and energy efficiency associated with modern renewable energy.

ABSTRACT. Long Term Evolution (LTE) networks are the backbone of modern mobile communications and support a wide range of applications with significantly different Quality of Service (QoS) requirements, ranging from best-effort data traffic to latency-sensitive services such as voice and video streaming. In this context, evaluating the behaviour of LTE networks as a function of application type, transport protocol, and QoS class is essential for the design and optimization of mobile networks. This paper presents a comparative performance analysis of an LTE network simulated in the NetSim environment, with the objective of assessing the impact of TCP and UDP transport protocols, as well as the Best Effort (BE) and Extended Real-Time Polling Service (ERTPS) QoS classes, on the main performance indicators: throughput, delay, and jitter. The analysis considers several representative application types (CBR, HTTP, EMAIL, FTP, VIDEO, and VOICE), and the resulting data support the identification of optimal protocol - QoS combinations for different traffic categories. The conclusions highlight the advantages of UDP for real-time applications and the critical role of QoS configuration in LTE networks.

ABSTRACT. This paper presents an Edge of Things (EoT) architecture and implementation for a real-time multisensor home safety monitoring system. The system integrates sensors for combustible gas and volatile organic compound detection (MQ-2), infrared flame detection (LM393), and three-dimensional seismic vibration monitoring (ADXL335). Data acquisition and initial processing are performed locally on an ATmega328 microcontroller, with the objective of significantly reducing latency, increasing autonomy, and conserving communication bandwidth. IoT communication and integration are provided by an ESP32 node using MQTT and ESP Home protocols toward the Home Assistant platform, which offers a unified interface for visualization, notifications, and event correlation. An ESP32-CAM module ensures passive visual redundancy, enabling manual validation of detected events. Experimental results highlight high sensitivity to methane, robust flame detection within a ~1 m range, and coherent vibration response, confirming the feasibility of a scalable EoT architecture with local decision-making processing for critical residential safety applications. Unlike typical IoT solutions, this work adopts a strictly decoupled Edge-of-Things architecture for reliable local safety decisions.

ABSTRACT. The rapid expansion of Internet of Things (IoT) infrastructures has increased their exposure to Distributed Denial of Service (DDoS) attacks originating from Botnet infected devices. Due to limited computational resources and weak security mechanisms, IoT nodes are highly vulnerable to such attacks, which can severely degrade network performance and compromise service availability. This paper presents an experimental evaluation of the impact of Botnet driven DDoS attacks on an IoT network using the NetSim simulation environment. In this study, we analyse multiple attack scenarios that include one, two, and three compromised nodes, and we compare the behaviour of UDP and TCP transport protocols under different traffic‑generation rates. We quantify network performance using standard QoS metrics such as throughput, end‑to‑end delay, and jitter. The results reveal a strongly nonlinear degradation pattern under attack conditions and highlight significant limitations of TCP in congestion prone IoT environments. While UDP sustains a higher throughput, it experiences substantial packet loss, whereas TCP becomes ineffective under intense traffic and attack conditions. The study underscores the necessity for advanced traffic control and security mechanisms capable of ensuring reliable IoT operation under DDoS threats.

ABSTRACT. Currently, AI tools are used to learn the properties of noise generators. This is useful in some applications, such as audio noise generators (NG), but may hamper the use of NGs in jamming and cryptography. We investigate the statistic properties of a class of NGs with pseudo-random dynamics. Coincidences between two statistical processes are defined by approximate equality of their outputs. Typically, coincidences can be described by Poisson distribution, with the time series generated by coincidences having Poisson statistic. In this article, coincidences between simple, independent nonlinear processes are studied. From two nonlinear processes, a process with Poisson distribution is produced by coincidences and a related process (signal) with exponential distribution is generated for the mixing the two nonlinear processes. The usefulness of the resulted pseudo-random number generators, based on coincidences, against AI-based filtering is discussed in relation with applications for security, privacy, and jamming. While not intended for cryptography, with adaptations, the solution might be useful also in that domain.

ABSTRACT. The paper presents the use of the IBM watsonx.ai AutoAI platform for the automatic classification of benign and malicious IoT traffic. The experiment was conducted using the CIC IoT-DIAD 2024 dataset, which was preprocessed in Python and organized into a balanced subset of 10,000 records corresponding to five traffic classes: Benign, DDoS_HTTP_Flood, DNS_Spoofing, DoS-TCP_Flood, and MITM-ArpSpoofing. The optimal model selected by AutoAI was Pipeline 14, based on the Batched Tree Ensemble Classifier, achieving 99.3% accuracy on the holdout set. The results were evaluated using performance metrics, feature importance analysis, the confusion matrix, and ROC curves. The model was subsequently deployed through the Online Deployment service and tested on 500 new observations, confirming its ability to generate automated predictions with high confidence.

ABSTRACT. Vision-based parking slot occupancy detection is a key component of intelligent transportation systems, but its performance degrades under adverse weather conditions such as rain, fog, and snow. Public datasets such as PKLot and CNRPark-EXT contain limited variability in severe weather, making robust generalization difficult. This paper introduces a weather-aware data augmentation pipeline based on Qwen-Image-Edit, a diffusion-based vision-language image editing model. The method generates synthetic snow, fog, and rain variants from a subset of overcast images using natural-language instructions, while preserving parking geometry and slot-level annotations; however, careful prompt design is required to ensure structural consistency during image editing. A patch-level ResNet-50 classifier is retrained on the augmented dataset and evaluated in a cross-dataset setting from CNRPark-EXT to PKLot. Experimental results show improvements of up to 4.0 percentage points in accuracy and up to 10.6 percentage points in precision, with consistent gains across all weather subsets in F1-score and accuracy.

ABSTRACT. Spectrum monitoring systems enables identification of signal modulation without prior knowledge of the transmitter, being one of the hot topics in signal processing. Artificial Intelligence tools now enable deploying automatic modulation classification (AMC), being a critical component of modern cognitive radio and spectrum monitoring systems, enabling the final task to retrieve information about the transmitter type or class. This study presents a comparative analysis of well-known architectures and proposes RVGG, a novel hybrid neural network architecture designed to enhance AMC in challenging, low Signal-to-Noise Ratio (SNR) environments. By augmenting a modified Visual Geometry Group (VGG) network backbone with residual connections, Squeeze-and-Excitation (SE) attention mechanisms, and Global Average Pooling (GAP), the proposed model dynamically filters channel-wise noise and significantly mitigates overfitting. Experimental evaluations on the comprehensive RadioML 2018 dataset demonstrate that RVGG consistently outperforms the baseline VGG model, yielding a 1.5% absolute improvement in overall accuracy and achieving superior F1-Scores across critical transition zones. These results validate that the targeted architectural optimizations successfully extract robust feature representations while maintaining strict computational viability for real-time Software Defined Radio (SDR) deployments.

ABSTRACT. This study presents a comparative performance evaluation of ALOHA and Carrier Sense Multiple Access (CSMA) protocols within a heterogeneous Wireless Sensor Network (WSN) subjected to structural and biological interference. Utilizing a low-cost node architecture integrated with nRF905 transceivers, we analyze channel throughput under varying transmission power levels (Ptx) and spatial constraints. Experimental results demonstrate that while CSMA offers superior channel utilization—achieving up to 0.38 Erlang—it remains susceptible to cumulative attenuation from 24 cm thick Autoclaved Aerated Concrete (AAC/Ytong) barriers. Furthermore, the introduction of human density was found to reduce throughput by approximately 12% due to dielectric absorption and shadowing. These findings underscore the necessity of accounting for structural impedance and biological factors in the design of resilient medium access control (MAC) mechanisms for indoor IoT environments.

ABSTRACT. Smart city entertainment, including movies, concerts, and games in stadiums, is consumed by young people through kiosks and mobile flows amid changing lighting, noise, and crowds. To bridge this divergence between average acceptance and runtime behavior tuning, we present FedA11y B, a five-mode accessibility engine whose context-aware switching policy is realized via Federated Learning (FL) and tuned through a dropout-aware post-prediction layer with task completion predisposed into account. The design extends WCAG 2.2 and EN 301 549 as compared with prior designs by allowing for live, real-time policy control for public-facing ICT without compromising existing conformance objectives. We train the policy using FedAvg on a non-IID venue split and then discuss how it aligns with organizational privacy frameworks and differentially private training, as appropriate. In the controlled simulation for entertainment journeys, FedA11y B improves the time (s) from 43.000 to 40.870 and then to 37.430; reduces errors (count) from 2.560 to 2.140 and then to 1.700; and lowers dropout (probability) from 0.248 to 0.223 and then to 0.213 compared with non-adaptive and rule-based baselines. In addition, the worst group gaps are lower than before, enabling equitable access by youth.

ABSTRACT. The electromagnetic diaphragm pump system is developed to transfer liquid media at low pressures and flow rates using the electromagnetic forces acting between the permanent magnet positioned on rubber membrane and the electromagnets. The fluid flow rate and temperature are control and monitor by a digital cabinet. This study proposes an equivalent electrical circuit model of electromagnetic diaphragm pump system obtained starting from temperature and pumped fluid measured within the laboratory. The estimation of the transfer functions was calculated by means of orthonormal rational basis functions based on frequency response measurements. Finally, the circuit synthesis based on input temperature and output pumped liquid was performed using the Cauer’s method for the representation of analogue circuit combining branch of resistor, inductor and capacitor.

ABSTRACT. The proposed work is based on the implementation of a mobile application for motorcycle drivers. This application is implemented in the Swift programming language combined with various APIs such as Apple WeatherKit and interfaces for integrating artificial intelligence in the backend, while the SwiftUI programming language is used in the frontend. The application uses the MVVM (models, views, view models) architecture and has two LLM providers integrated into it: Apple Foundation Models - for processing local tasks and Google Gemini API - for processing tasks in the cloud and making LLM available on devices that are not compatible or do not have the necessary resources to run locally.

ABSTRACT. This paper proposes a conceptual design methodology for ferrite geopolymer composites intended to guide and concentrate magnetic flux in WPT systems. The proposed materials are based on a metakaolin geopolymer matrix modified with magnetic and functional oxides, including Fe₃O₄, MgO, ZnO, TiO₂, and CuO. The geopolymer matrix provides high electrical resistivity and good mechanical stability, while the embedded magnetic phase contributes to increasing the effective magnetic permeability and improving the magnetic coupling between coils. Key design parameters such as magnetic phase volume fraction, particle size and dispersion, and the geometry and thickness of the magnetic layer are analyzed in relation to their influence on electromagnetic performance. To evaluate the potential impact of the proposed materials on system performance, a parametric simulation approach is implemented in Matlab/Simulink. The effect of the geopolymer-ferrite layer is modeled through variations of the magnetic coupling coefficient as a function of material composition and layer thickness. The proposed methodology is initially validated for low-power WPT systems (10 W), with the possibility of extending the approach to higher-power applications. The results highlight the potential of ferrite geopolymer composites as structurally robust and functionally adaptable materials for improving magnetic flux management in wireless power transfer systems.

ABSTRACT. The paper presents the development of a hardware-software system for reproducible colorimetric detection and analysis using a TCS3200 optical sensor integrated into an embedded architecture based on Arduino. The proposed platform combines a linear positioning subsystem made with a NEMA17 stepper motor, A4988 driver and TR8×8 trapezoidal screw mechanism, with a software application developed in Python for motion control, data acquisition and display of results in real time. The system allows the controlled movement of the optical sensor relative to the analyzed surface and the performance of RGB measurements under reproducible geometric conditions. To reduce the influence of illumination and variation in the measurement distance, a calibration method using white, black, red, green and blue color standards is proposed. The procedure allows the determination of RGB correction coefficients depending on the position of the sensor and compensation for errors generated by optical and geometric variations of the system. Experimental data is saved and used for automatic correction of subsequent measurements. The developed system represents a compact and low-cost solution for colorimetric analysis applications, pigment monitoring and non-invasive investigation of materials and cultural heritage.

ABSTRACT. Proton exchange membrane (PEM) water electrolysis is a premier technology for green hydrogen production, but achieving high energy efficiency and stack durability requires careful optimization of operating conditions. This paper investigates the coupled effects of elevated operating temperatures and pressures on PEM water electrolyzers. By integrating theoretical thermodynamic models, electrochemical overpotential analyses, and multi-physics representations of gas crossover, we analyze the performance of both balanced and differential pressure configurations. Our analysis reveals that while increasing temperatures up to 80-85°C enhances reaction kinetics, reduces thermodynamic onset potentials, and minimizes ohmic losses, operating above 90°C causes rapid membrane degradation, passivation of titanium components, and localized dry-out. Furthermore, we evaluate how gas crossover (diffusion and convection) limits low-load operations and Faradaic efficiency. Finally, we propose a hybrid sequential compression strategy (40-80 bar electrochemical compression followed by non-mechanical cyclic adsorption-desorption) combined with dynamic temperature control to maximize system efficiency and operating lifetime.

ABSTRACT. This paper proposes the enhancement of Supervisory Control and Data Acquisition (SCADA) systems associated with hydropower plants located on the Argeș River through the integration of a continuous vibration monitoring system. The study focuses on Hydropower Unit No. 1 of CHE Noapteș (7.7 MW), where experimental data were collected and analysed under various operating conditions. The results reveal a clear dependency between the vibration regime and the unit load, highlighting the presence of operational instabilities. These instabilities were primarily attributed to mechanical factors, such as shaft misalignment and uneven tightening of couplings, which significantly influence the dynamic behaviour of the system. From an economic perspective, the implementation cost of the proposed monitoring solution represents only 5.25% of a level 3 maintenance operation. However, the indirect benefits—such as the prevention of major failures and the reduction of downtime—can exceed the initial investment by a factor of 5 to 10. Consequently, the system demonstrates a cost recovery period of less than one year, supporting its viability as an efficient predictive maintenance tool for hydropower applications.

ABSTRACT. Embedded systems are increasingly integral to cyber-physical infrastructures, ranging from resource-constrained microcontrollers in IoT (Internet of Things) nodes to high-performance edge AI (Artificial Intelligence) platforms supporting perception and autonomous decision-making. This review provides a platform-centric perspective on cybersecurity in embedded systems, emphasizing how architectural and operating constraints influence attack surfaces and feasible defenses. Key characteristics such as real-time requirements, limited compute/memory/energy resources, heterogeneous peripherals, and long deployment lifecycles are linked to prevalent threat vectors, including firmware tampering, insecure boot chains, communication-layer attacks, physical access threats, and supply-chain exposure. Security capabilities are then contrasted across representative platform classes, distinguishing microcontroller-based systems from embedded Linux single-board computers and edge AI devices, with attention to the availability and maturity of primitives such as secure boot, memory protection, cryptographic acceleration, secure key storage, integrity verification, and secure update mechanisms. Current trends and open challenges are summarized, including secure OTA (Over-the-Air) for constrained devices, hardware-assisted isolation, robustness of edge AI workloads under adversarial conditions, and the need for reproducible security benchmarking. The proposed taxonomy and discussion support platform selection and security mechanism prioritization under explicit threat models and system requirements.

ABSTRACT. The accelerated digitalization of port infrastructures and the emergence of the Smart Port paradigm have significantly increased the interconnectivity between operational systems, industrial equipment, and digital logistics platforms. In this context, securing data communications has become a critical requirement for ensuring operational continuity and protecting critical infrastructures against cyber threats. This study investigates the role of handshake protocols in securing communications within modern freight terminals, using a smart port architecture as a representative case study. The proposed framework combines TLS 1.3 security mechanisms with hybrid key-exchange approaches inspired by post-quantum cryptography based on supersingular elliptic curve isogenies. The research methodology employs a Digital Twin model of the port infrastructure and network simulations involving 100 interconnected nodes, MQTT/TLS traffic, and multiple cyberattack scenarios, including spoofing, MQTT hijacking, and Distributed Denial-of-Service (DDoS) attacks. System performance is evaluated through availability, latency, packet loss, recovery time, and a composite Cyber Resilience Index. The results demonstrate that hybrid architectures integrating TLS and post-quantum cryptographic mechanisms provide superior resilience and service continuity compared to conventional security solutions. The findings highlight the potential of supersingular elliptic curve-based cryptography as a strategic solution for enhancing Smart Port cybersecurity and preparing critical maritime infrastructures for future quantum computing threats.

ABSTRACT. The rapid growth of phishing campaigns and malicious web services has transformed URLs (Uniform Resource Locators) into a primary attack vector in modern cyber threats. Traditional blacklist-based filtering and simple pattern-matching techniques are often unable to keep pace with the evolving landscape of URL obfuscation strategies. This paper proposes a machine learning framework for malicious URL detection, complemented by a LabVIEW-based virtual instrument that visualizes and simulates phishing attacks. A focused taxonomy of phishing URL manipulation techniques is developed, including sound squatting, combo squatting, homograph attacks, typo squatting, subdomain squatting and TLD (Top-Level Domain) squatting, and representative examples for each category are illustrated within the virtual instrument. This interactive environment allows users to explore how small lexical and structural changes in URLs can bypass naive security filters. On the detection side, lexical and statistical features are extracted, such as URL length, character entropy, special symbol density, subdomain depth and brand-similarity indicators and several supervised machine learning models are evaluated for binary classification of benign versus malicious URLs. Experimental results show that tree-based ensemble methods outperform linear baselines, achieving high detection rates with low false-positive rates. Finally, potential avenues are outlined for integrating the proposed framework into real-world security workflows and for using it as an educational tool to demonstrate both the mechanics and the risks of phishing attacks.

ABSTRACT. Mobile marketplace applications, in conjunction with aspects of governance, food security, public policy and economic regulation on the one hand, and the impact on the power of economic actors within the supply chain on the other, establish a direct link between producers and consumers while eliminating intermediaries. The aim of this paper is to promote sustainable agriculture, transparency and responsible consumption by facilitating the sale of organic agri-food products directly from farm to consumer. The application was implemented using the .NET MAUI platform for the mobile interface and WCF services for server-side data management, supported by a SQL Server database. The system integrates IoT sensors for monitoring transport conditions, a fuzzy logic module implemented in MATLAB for data analysis, and a Dialogflow chatbot that provides assistance to users. As a result, a comprehensive digital solution is proposed that combines modern technologies, including cross-platform development, web services, artificial intelligence and the Internet of Things, thereby optimising the sales process for organic products and supporting local producers in connecting directly with consumers.

ABSTRACT. This paper addresses early acoustic forest fire detection as a methodological problem of designing a reproducible software chain rather than as a mere competition among classifiers. The emphasis is placed on defining an end-to-end pipeline that includes dataset standardization, segmentation, engineering-justified preprocessing, input representation, classification, and operational evaluation. The paper synthesizes the essential directions in the recent literature on fire detection, acoustic fire detection, environmental sound classification, bioacoustics, and deployment on edge platforms. On this basis, it delineates the current gap in the field and argues for the necessity of a leakage-free protocol, practically relevant metrics, and transparent reporting of experimental decisions.

ABSTRACT. This paper formulates a lightweight signal-processing front-end for acoustic forest fire detection as a reusable methodological layer within a broader audio-analysis pipeline. The study consolidates evidence from eight thematic literature sets covering feature fusion, autocorrelation, handcrafted and interpretable descriptors, Hilbert-envelope analysis, low-complexity audio descriptors, spectral representations, and time-frequency modeling, and relates this corpus to three prior fire/non-fire studies developed by the present research group. The main contributions are a structured synthesis of the relevant descriptor literature, a lightweight and classifier-agnostic front-end design, and a compact mathematical formulation suitable for downstream integration with PCA-based models, conventional machine-learning classifiers, CNNs, RNNs, Audio Spectrogram Transformers, and hybrid CNN–Transformer architectures. The resulting manuscript provides the theoretical and methodological foundation required to position the proposed front-end as a reusable component for acoustic fire monitoring.

ABSTRACT. This paper proposes an integrated framework for investment portfolio optimization that combines gradient boosting prediction models (LightGBM), volatility forecasting via the Heterogeneous Autoregressive Realized Volatility model (HAR-RV), market regime detection via Hidden Markov Models (HMM), and reinforcement learning agents. The system was implemented as an interactive web application (Streamlit) and evaluated on a diversified portfolio of 30 assets spanning six economic sectors, automatically selected from a universe of 530 tickers with a 40% sector concentration cap enforced at each rebalancing step to ensure structural diversification. Four reinforcement learning algorithms - PPO, SAC, TD3, and A2C - were evaluated under strictly identical experimental conditions to provide a comprehensive algorithmic comparison. The LightGBM return direction model achieved a walk-forward out-of-fold directional accuracy of 64.8%, while HAR-RV demonstrated consistent volatility forecasting superiority over LightGBM on 49 of 51 evaluated assets, motivating its adoption in the Black-Litterman-HRP covariance adjustment. Over the full 2015-2024 period, SAC achieved the highest Sharpe ratio (1.193) and Jensen's Alpha (+10.58% annual), followed by PPO (Sharpe 1.157). Out-of-sample (2021-2024), SAC maintained dominance (Sharpe 1.002, Alpha +6.33%), confirmed robust across eight distinct evaluation windows in a temporal sensitivity analysis. Deflated Sharpe Ratio analysis (p_DSR = 1.000 for all strategies) highlights the inherent statistical limitation of ~2.5-year evaluation horizons, motivating the presented framework as a validated software architecture for integrated ML+RL portfolio management rather than a definitively proven superior strategy.

ABSTRACT. This paper presents the development and validation of an experimental Digital Twin for a linear actuator with servomotor, implemented in the MATLAB/Simulink environment. The model reproduces the dynamic behaviour of the drive system based on real parameters and interface signals defined at the servomotor driver level, using a modular architecture that explicitly separates the motion profile, the control structure and the physical system model. Three control strategies are compared: the classical PID controller in a hierarchical structure, the Mamdani-type fuzzy control and the Sugeno-type fuzzy control, on systematic positioning and velocity tracking scenarios, with mechanical loads varying between 100 g and 10 kg. The quantitative evaluation uses standardized performance indicators: rise time, percentage overshoot, absolute stationary error and stabilization time, synthesized in comparative graphs and a normalized radar graph. The results demonstrate that the Sugeno fuzzy controller offers the fastest response with a trade-off in overshoot, the PID control provides superior stability and reduced stationary errors, and the Mamdani controller exhibits good robustness to load variations without overshooting. The proposed Digital Twin constitutes a reproducible and extensible platform for analysing, comparing, and optimizing control strategies in industrial linear drive systems.

ABSTRACT. Richards finite-state machines represent an efficient alternative to classical Mealy and Moore models in the implementation of complex sequential control systems. Their modular organization simplifies hardware synthesis and enables flexible modification of the control algorithm without complete redesign of the controller structure. This paper presents the synthesis methodology and practical implementation of Richards Controllers together with the development of a configurable educational hardware module dedicated to their experimental study. The proposed module allows rapid implementation of different state-transition diagrams through jumper-based reconfiguration. The developed system includes configurable test-condition emulation and mapping circuits, jump-address generation logic, optical output-state visualization, and selectable continuous or step-by-step operating modes. Experimental validation confirmed the correct functionality and educational utility of the proposed hardware module for laboratory activities involving sequential systems and finite-state machine synthesis.

ABSTRACT. This article investigates the methodology for detection and identification of the electrical faults that can occur in photovoltaic panels. The fault-free operation of photovoltaic systems is an achievable goal for increasingly longer periods of time. Preventing faults, improves the efficiency of electricity supply and increases the lifespan of the panels. This paper highlights the most common panel faults, using the multiple simulation method, based on scenarios that encompass each case of electrical fault. Emphasis is on describe and simulation the types of faults such as open circuit, cell is short-circuited, the rows of photovoltaic cells connected in parallel are short-circuited, system output is short-circuited, increase series resistance between two rows, etc. The simulation results and validation of the proposed technique are implemented by MATLAB/Simulink toolbox and PSpice. The purpose of simulating the types of defects is to demonstrate, based on the results, first of all, the accuracy of the values obtained, but also the direct impact on reducing the electrical power provided by the photovoltaic panels.

ABSTRACT. This paper presents a comparative study of two of the most widely used game engines in modern game development: Unity and Unreal Engine. The research focuses on their evolution, software architecture, scripting systems, asset management solutions, and overall development workflow. Unity uses a component based GameObject architecture together with C# scripting, offering a flexible environment well suited for rapid prototyping, indie projects, and cross platform deployment, especially on mobile devices. Unreal Engine, on the other hand, relies on C++ and the Blueprint visual scripting system, providing a more structured framework optimized for high performance applications and advanced real time graphics. The paper also compares gameplay architecture, data driven design approaches, and performance related aspects of both engines. In addition, a practical case study involving the implementation of a third person character controller is presented in order to highlight differences in workflow, code structure, and built in engine functionality. The aim of this study is to provide developers with a clearer understanding of the strengths and limitations of each engine and to support the selection of the most appropriate solution depending on project requirements, team experience, and target platforms

ABSTRACT. This paper presents a comprehensive comparative analysis between human and AI-assisted code review methodologies, focusing on efficiency, accuracy and their impact on software quality. As modern software development increasingly integrates artificial intelligence (AI) tools, the volume of generated code has significantly expanded, creating new challenges for quality assurance processes. While AI-based tools provide quick feedback and scalability, they often lack contextual understanding and produce lower-value suggestions. In contrast, human reviewers offer deeper insights into logic and architecture, but require significantly more time. This study evaluates both approaches through existing literature and practical analysis, highlighting their strengths and limitations. The findings suggest that a hybrid model combining automated and human review processes provides the most effective solution for maintaining code quality while preserving development speed.

ABSTRACT. This paper presents a hybrid control architecture for dual-axis photovoltaic (PV) tracking systems based on the integration of a Fuzzy Logic Controller (FLC) and a Battery Management System (BMS). The proposed approach addresses the limitations of conventional Light Dependent Resistor (LDR) and Proportional-Integral-Derivative (PID) methods under variable environmental conditions. The main contribution consists in a multi-parameter control strategy that combines differential illumination, global solar radiation, and battery state of charge to enable adaptive and energy-aware actuation. An adaptive deadband mechanism is introduced to reduce unnecessary movements and mitigate the hunting effect. The system also includes hardware enhancements for improved reliability and fault tolerance. Experimental results show that the proposed method reduces actuator activity by approximately 73% under partially cloudy conditions and improves net energy yield by 15–20% compared to classical approaches. The results demonstrate the effectiveness of combining intelligent control with energy management for efficient and sustainable PV tracking.

ABSTRACT. Industry 5.0 represents a paradigm shift from Industry 4.0, repositioning the human factor, sustainability and resilience as central values of industrial processes. Although recent literature extensively covers the individual technologies associated with this transition, a systematic approach is lacking that simultaneously maps the enabling technologies onto the three official pillars of Industry 5.0, assesses their maturity level and identifies tensions between the pillars. This paper proposes a pillar-based taxonomy, namely Human-Centricity, Sustainability and Resilience, as an original framework for classifying key technologies for Industry 5.0. The taxonomy integrates three dimensions of analysis: pillar membership, technological maturity level and inter-pillar conflict relationships, providing a structured framework for analyzing Industry 5.0 technologies. Based on the proposed taxonomy, cross-cutting technologies that simultaneously serve multiple pillars are identified, together with four key inter-pillar trade-offs. The paper concludes with a structured agenda of open research directions, derived directly from the taxonomy analysis. The results provide a useful framework for both researchers positioning their contributions in the context of Industry 5.0 and industrial practitioners assessing the technological maturity of available solutions.

ABSTRACT. Virtual reality (VR) has become one of the most important technological tools for educational innovation in recent years. This paper presents a modularly developed VR-based serious game: a survival simulator and its areas of application. The application exploits the possibilities of immersive VR in learning environments and problem-solving situations, for example, it provides an opportunity to acquire practical skills and competencies in safe and realistic situations; and it can improve motivation and increase creativity. The technological novelty of the application lies in its modularity, and thus its easy customizability. From an educational methodological point of view, its uniqueness lies in the fact that the content and the simulation scenario can be customized based on professional consultation before the development of the process. The application develops resource use, object recognition, spatial orientation and navigation skills through a survival simulation in nature. It can also be used well in behavioral assessment, as well as in the development of decision-making and problem-solving skills in various competencies. This article provides suggestions for the application and further development possibilities of the application.

ABSTRACT. This paper presents an operational capacity assessment of the reversible lane system on DN39 national road, Eforie Nord sector (km 12+370 – km 14+250, L = 1,880 m), Romania — a three-lane coastal corridor traversing a Black Sea resort with pronounced seasonal traffic asymmetry. Traffic monitoring data for 2025 document an Annual Average Daily Traffic (AADT) of 28,200 vehicles per day and a seasonality ratio of 2.08 between summer peak and winter baseline, with absolute daily peaks reaching 52,060 vehicles/day. The central lane currently operates under a Generation I manually reversible scheme. This study analyses the transition to an ITS-based automated management system integrating overhead Lane Control Signals, Variable Message Signs, and coordinated signalization of the sector's intersection and pedestrian crossings. Capacity and Level of Service assessments are conducted under the AND 584-2012 normative framework, cross-referenced against HCM 7 and NCHRP Synthesis 340. The ITS upgrade yields a 20.7% increase in effective lane capacity, a 17.1% reduction in volume-to-capacity ratios, and a Level of Service improvement from F to E on both directional flows under normative design conditions. A seasonal congestion duration analysis demonstrates a reduction from approximately 153 to 92 congested days per year (−40%), positioning the ITS upgrade as the optimal short-to-medium term intervention on this structurally constrained corridor.

ABSTRACT. Vehicle classification is a key functionality in Intelligent Transportation Systems (ITS), supporting traffic monitoring, infrastructure planning, access control, enforcement, tolling, and operational analytics. This paper presents a comparative case study of two practical approaches for vehicle classification: an intrusive solution based on inductive loops, represented by the FEIG VEK S4 detector, and a non-intrusive video-based solution, represented by the Adaptive Recognition Vidar camera equipped with Automatic Number Plate Recognition (ANPR) and Make and Model Recognition (MMR) functions. The study is anchored in a roadside deployment on the A0 Bucharest motorway, where both systems were installed in parallel on a two-lane carriageway with an emergency lane and a lane width of 3.75 m. The analysis indicates that inductive loop technology remains highly relevant when deterministic lane-level detection, speed estimation, and standard traffic classification are required, whereas camera-based ANPR/MMR systems provide richer semantic information and are better suited for advanced ITS services that require vehicle identification, evidence generation, and analytics-oriented processing. A mathematical comparison of the two inference models is also included.

ABSTRACT. The absorption of deuterium in CANDU (CANadian Deuterium Uranium) pressure tubes from the heavy water coolant during normal reactor operation leads to an increase in the concentration of hydrogen equivalent in the material volume. This, combined with the existence or occurrence of defects on the inner surface of the CANDU pressure tube, Zr-2.5%Nb (Zirconium 2,5% Niobium) alloy, results in the migration of hydrogen to locations of high mechanical stress. Under specific thermal cycling conditions, determined by the loading-unloading regimes of the nuclear fuel, localised hydrides may form at mechanical stress spots on the top of the volumetric flaw, thereby developing a fragile zone in these locations. Determining the threshold mechanical stress, from which the reorientation phenomenon of zirconium hydrides can occur, plays an important role in the structural integrity analyses, which are carried out on Fitness-for-Service criteria in the periods between two periodic inspections of CANDU 600 fuel channels. In this paper, a method is developed to obtain the threshold stress value from which the reorientation of localized hydrides occurs based on metallographic analyses and stress field analysis at volumetric flaws by using the finite element method (ANSYS).

ABSTRACT. The increasing adoption of Virtual Private Networks (VPNs) for secure communications has created new challenges regarding traffic monitoring, classification, and anomaly detection in resource-constrained environments. This paper presents an intelligent framework for VPN traffic analysis and classification deployed on a Raspberry Pi platform. The proposed system combines secure VPN connectivity using OpenVPN and WireGuard with real-time traffic monitoring and machine learning (ML)-based analysis. Network traffic is collected and processed using packet capture and inspection tools, while Random Forest (RF) and Support Vector Machine (SVM) models are employed to classify traffic flows and identify potentially suspicious activities. Furthermore, TensorFlow Lite is integrated to enable lightweight real-time inference on embedded hardware with limited computational resources. Experimental evaluation demonstrates the feasibility of deploying intelligent traffic analysis services directly on edge devices, providing a cost-effective solution for secure Small Office/Home Office (SOHO) environments. The obtained results indicate that ML techniques can effectively enhance traffic visibility and support cybersecurity monitoring while maintaining low resource consumption on Raspberry Pi-based platforms.

ABSTRACT. The increasing complexity of malware and ransomware underscores the necessity for sophisticated cybersecurity measures that extend beyond conventional protective strategies. Artificial Intelligence (AI) has transformed the landscape of cyber defense by utilizing Machine Learning (ML) techniques to identify, mitigate, and avert threats in real time. AI-based systems can process extensive datasets, detect irregularities, and address advanced threats such as polymorphic malware and zero-day exploits. Additionally, behavioral analysis enhances security protocols by forecasting and preventing attacks prior to inflicting damage. To address this evolving threat landscape, this project introduces AI-driven malware and ransomware detection system built using Long Short-Term Memory (LSTM) networks. Leveraging deep learning, the model is trained on both real-world and synthetic datasets containing static and behavioral features of executable files. Key attributes such as import functions, memory alignment, and system characteristics were extracted to enhance prediction accuracy. Nonetheless, issues such as adversarial tactics in AI and ethical dilemmas related to data privacy require careful consideration. Despite these challenges, AI-enhanced cybersecurity presents a formidable, proactive approach, empowering organizations with automation, predictive analytics, and resilience to protect sensitive information and ensure operational integrity.

ABSTRACT. Chronic low back pain (CLBP) is one of the leading causes of disability worldwide and significantly affects mobility, work productivity, and quality of life. Conventional rehabilitation approaches often rely on passive lumbar supports or single-modality therapies that provide limited long-term rehabilitation effectiveness. This study presents the design and development of a wearable multimodal rehabilitation belt integrating thermal therapy and vibration stimulation for non-invasive chronic low back pain management. The proposed rehabilitation system incorporates flexible heating pads, high-speed DC vibration motors, embedded microcontroller-based control, rechargeable power management, and ergonomic lumbar support within a lightweight wearable platform suitable for continuous home-based rehabilitation. The thermal subsystem operates within a controlled therapeutic temperature range of 37–45°C to improve blood circulation, reduce muscular stiffness, and promote muscle relaxation. Simultaneously, the vibration subsystem delivers localized mechanical stimulation for pain modulation and neuromuscular activation. An experimental evaluation was conducted using participants with different occupational backgrounds experiencing lower back discomfort caused by prolonged sitting, poor posture, or physical strain. The obtained results demonstrated noticeable improvement in rehabilitation comfort, pain reduction, and muscle relaxation, with satisfaction levels ranging from 50% to 85% and an average overall satisfaction score of 69%. The proposed system demonstrated several practical advantages, including portability, lightweight wearable design, rechargeable operation, adjustable therapeutic control, and suitability for continuous rehabilitation applications. Overall, the developed rehabilitation belt represents a promising wearable biomedical solution for chronic low back pain rehabilitation and supportive musculoskeletal therapy.

ABSTRACT. Securing user authentication while keeping device resources low remains an open problem in the Smart Environment of Things, particularly for the highly anticipated Internet of Health Things (IOHT). In this paper, we introduce a biometric-based, ultra-lightweight mutual authentication protocol for resource-constrained IoT. In addition to a lightweight cryptographic hash function and XOR operations, Kyber post-quantum lattice-based structures are used to secure the permission phase between a patient and a doctor, thereby allowing access to the EHR. Using the automated symbolic tool ProVerif, formal security verification shows that the protocol guarantees perfect session key secrecy, strong client-to-server mutual authentication, non-repudiation, and complete resistance to replay and man-in-the-middle attacks. Performance evaluations show that our scheme achieves efficient performance with a communication cost of 1344 bits, an ultra-low computation time of 0.0654 ms, and low energy consumption (0.43164 mJ). Network analysis also demonstrates remarkable performance, with total latency under 0.2 ms and server throughput exceeding 30,000 authentications per second. These observations validate that the newly suggested protocol offers a good trade-off between high security and scalability, making it an ideal candidate for real-world high-density IoT scenarios.

ABSTRACT. The ever increasing number of aging populace together with high number of chronic diseases has led to wide adoption of medical internet of things, such as wireless body area networks (WBANs). Through these networks, physiological data such as blood glucose, body temperature, ECG, blood pressure, heart rate and EEG are collected and transmitted to remote medical professionals for analysis, diagnosis and treatment. However, the deployed open wireless channels and IEEE 802.15.6 standardized protocol have vulnerabilities that can easily compromise the patient data. Although many security solutions have been presented based on techniques such as certificates and bilinear-pairings, these algorithms have either high overheads or still have numerous security gaps. In this paper, a lightweight encryption protocol is presented. It is shown to incur the least computation costs and average communication overheads. In addition, it offers anonymity, untraceability, authentication, session key agreement, perfect key secrecy, revocability and conditional privacy. Moreover, it protects against node capture, session hijacking, denial of sleep, denial of service, key compromise, side-channeling, impersonation, KSSTI, eavesdropping, session key disclosure, privileged insider, tracking, offline guessing, forgery and non-repudiation.

ABSTRACT. In the age of big data, large quantities of audio signals are output continuously every day. Following the rapid advancements in modern computer and social media technologies, we now have an alternative framework for building a more secure audio encryption system. As a result, data encryption technologies have evolved considerably. Several audio encryption techniques are now available, mainly because the quantum and chaos theories have become increasingly important lately. This study presents a cryptographic technique for protected speech communication. Our approach is novel because it integrates four separate audio encryption methods into the same structure, making it safer. These methods are dynamic DNA sequence scrambling, quantum chaotic maps, multi-chaotic maps, and graph theory. In the first stage, graph theory and a multi-chaotic map are used to generate multi-encryption keys. The creation of a multi-chaotic map involves three types of chaotic map, a 1D sine map, a 2D extended logistic map and a 2D Hanon map, which results in unexpected encryption keys. In the second stage, the speech signal vectors are scrambled twice: once using a quantum chaotic map, and again using a dynamic DNA sequence. Using the encryption keys established in the first stage, the speech signal obtained in the second stage is sent to the final module. In the third stage, the encryption key elements and scrambled speech vector values are transformed into a binary representation to compute an XOR operation between them. Next, a different encryption key is chosen randomly for each vector of the scrambled speech signal, to increase the difficulty for hackers attempting to retrieve the encrypted data. Numerous metrics are used to evaluate the process, including the key space, log-likelihood ratio, histogram, segmental signal-to-noise-ratio, spectrogram, correlation coefficient, and signal-to-noise ratio. The results show that this method is safer than several existing comparable speech encryption methods against various forms of attack.

ABSTRACT. E-government involves the deployment of information communication technologies to communicate and provide government services to various parties such as employees, citizens and businesses. To enhance governance and increase the availability of government services to citizens, developing countries must adopt e-government systems whose focus is on accelerating access to technological advancements. These e-government services are normally provided through electronic portals in a timely and user friendly manner. On the flip side, many developing countries especially Arab states face significant obstacles in implementing and adopting e-government programs. This paper analyzes the e-government concept as well as its key merits. In addition, the current state of e-government adoption is provided, highlighting the challenges and obstacles hindering its progress in developing countries. Moreover, a comprehensive review of e-government structures, specific applications and challenges facing the promotion of e-government initiatives are described. Basically, the aim is to identify strategies for overcoming these obstacles.

ABSTRACT. Although cloud computing has become an important aspect of data storage and retrieval, it is associated with significant issues of confidentiality, integrity, and trust in centralized service providers. Searchable encryption (SE) is a potentially appealing solution, as it supports keyword searches of encrypted data; however, existing SE systems still suffer from problems such as leakage of sensitive information, the production of non-verifiable results, and limited transparency. Recent studies have shown that a combination of blockchain and SE can enhance security through decentralization, immutability, auditability, and a verifiable search operation. The review presented in this paper provides a systematic categorization of blockchain-based SE schemes published between 2022 and 2025, including symmetric-based, public-key DE, attribute-based DE (ABDE), fuzzy private sets, and AI-aided models, as well as post-quantum-based approaches. We propose an efficiency-, security-, and practice-oriented comparative evaluation framework, and point out some of the most significant challenges, such as pattern leakage, blockchain overheads, smart contract risks, scalability bottlenecks, and multi-authority management. Motivated by these requirements, we argue for a unified next-generation design that includes on/off-chain verifiability, dynamic privacy, decentralized key management, post-quantum security, Layer 2 scaling, and hybrid exact–fuzzy searching. The aims of this survey are to present an in-depth assessment of the current state of the art and to provide a framework for the future development of secure, transparent, and practical blockchain-based SE systems that can support large-scale real-world cloud applications.

ABSTRACT. In the age of advanced technology and great development, in which information can be accessed at any place or time, a lot has been happening in the education sector because of software technology, cloud services and other related tools that would ensure speed, integrity and security. Despite its pluses, the education sector also faces challenges, such as data security and how to help prevent breaches. Accordingly, we must deploy a way to guarantee the access privileges and protect the security of information, because of existing issues in terms of ways to determine access privileges, how roles are distributed, cyber invasion and security function. In this paper, we have introduced a student authentication scheme that has the feature of being secure against security threats and maintains the privacy of students who can access their secure data and take an exam through a secure exam server. Moreover, our work benefits from security attributes such as mutual authentication, key management, forward secrecy and so forth to defend against famous attacks such as those related to insider attack, phishing attack and replay attack. Furthermore, the proposed scheme has strong security supported by formal analysis of Proverif and informal proof. In addition, we invoke BAN logic to authenticate the security proof and our work gives good results compared with similar works in terms of communication costs (620), computational costs (406.5) and throughput (12.2 Kbps).

ABSTRACT. Mobile Ad hoc NETworks (MANETS) are emerging as a favored approach for facilitating multimedia communications, especially in more infrastructure-less and highly dynamic settings. Nevertheless, reliable multimedia delivery in such environments is a real challenge, and the rapid topological changes in MANETs along with the unreliability of wireless links lead to packet loss that degrades the perceived quality of video at the user end. The proposed work formulates an improved routing protocol based on the Enhanced Dynamic Adaptive Forwarding Table (EDRAFT) technique, and the innovative approach of Hybrid Harris Hawks–Grey Wolf Optimization (HHGWO) is integrated to perform even better decision-making for choosing forward paths in MANETs. This optimization technique will help derive a multi-criteria routing metric based on multiple indicators of network reliability, such as packet loss and delay. Simulation results were obtained to analyze the proposed EDRAFT-HHGWO-based routing approach. The simulations presented consisted of 45 mobile nodes distributed over a square area of 120m x 120m, with overall simulation duration of 60 seconds. During each simulation, 1,534 video packets were generated and transmitted across the network. To obtain statistically reliable results, the average values of five independent simulation runs were computed. Simulation obtains deterministically higher multimedia communication performance compared to other works in the literature for the proposed EDRAFT-HHGWO-based routing method. The simulation trials showed that in the best trial, 1,302 video packets were successfully delivered, with stable and reactive routing behavior across varying network scenarios. Moreover, simulation trials results showed that the advocated method positively affects some Quality of Service (QoS), Quality of Experience (QoE) indicators; packet delivery ratio, packet loss ratio and video quality parameters (Peak Signal-to-Noise Ratio [PSNR], Structural Similarity Index [SSIM] and Mean Opinion Score). In conclusion, optimization algorithms combined with the EDRAFT routing protocol yield optimal paths by selecting fewer hops, achieving lower costs and higher performance for multimedia problems than traditional methods.

ABSTRACT. The rapid convergence of decentralized finance (DeFi), central bank digital currencies (CBDCs), and AI-driven cyber warfare has introduced unprecedented systemic risks to banking sectors, particularly within emerging economies. A new trend in hybrid warfare is to attack financial infrastructure with advanced digital attack vectors that reveal fragilities in traditional centralized risk assessment models based on unchanging datasets and standardized simulations. The research paper is a proposal of a new system to measure the financial resilience with blockchain-native intelligence and predictive analytics (BC-NI-PA). The study presents a novel model of Resilience-as-a-Smart-Contract (RSC) that combines on-chain analytics with machine learning to identify any initial signs of stress at the system level. The data on major blockchain networks including Ethereum and Polygon are looked at on a transaction level in 2023 to 2025 to track the liquidity changes, governance conduct, and wallet-level abnormalities. To detect temporal anomalies that come with cyber-induced disruptions, Long Short-Term Memory (LSTM) networks are used. An experimental simulation blockchain transparency and Monte Carlo prediction are established to establish an environment that simulates attack instances, such as governance manipulation, flash loan parasites, and RPC-layer attacks. The framework is topped with the design of a prediction of a liquidity crisis and institutional instability in decentralized-integrated banking ecologies through the On-Chain Early Warning Index (OC-EWI). It has liquidity stability index values of 0.64 to 0.94, gauge manipulation risk values of 0.45 to 0.88, anomaly detection score values of 0.42 to 0.85 and smart contract stress index values of 0.72 to 0.96.

ABSTRACT. Image splicing is a common attack vector in the digital world which is quite often deployed for malicious activities such as the dissemination of fake news. In order to preserve trust in visual evidence, these image manipulation technique need to be detected. Unfortunately, conventional image slicing detection cascades are primarily designed based on hand-crafted features, which is inefficient. In addition, conventional detection mechanisms experience some trade-offs between accuracy and localisation error. As such, many deep learning approaches have been developed for this task. However, these deep learning techniques are black-box models, which are characterized with low explainability and poor generalisation to other types of transformations. To address these problems, this paper presents a deep learning approach based on a Siamese Mask R-CNN for the detection of image splicing. The deployed Siamese architecture includes two symmetric networks with shared weights, which can be trained to compare pairs of inputs. As such, it results in better detection of image-splicing forgeries. Moreover, the deployed Mask R-CNN can detect splicing regions and their boundaries via pixel-level segmentation. The proposed method is evaluated on the CASIA v2 and Columbia datasets, achieving accuracy and precision values of 98.52% and 100% respectively. In addition, the proposed Siamese network with Mask R-CNN yields a significant boost in manipulation detection, particularly for image splicing. This technique renders the proposed model more attentive to image content, in addition to being much sensitive to changes. As such, it is capable of recognising splicing operations more efficiently and effectively.

ABSTRACT. The global growth in the use of digital educational platforms has exposed these platforms to various security risks. As such, there is heightened need for robust security and privacy solutions, especially in the face of quantum computing algorithms that can break some of the conventional cryptographic techniques. Unfortunately, even the modern networked systems rely heavily on these obsolete cryptographic operations for security protection. Across the world, educational infrastructures are increasingly dependent upon distributed cloud and fog technologies at local and global levels. As such, the cloud and fog computing paradigms handle sensitive academic records, identities, and intellectual property. However, the majority of the security measures deployed in these platforms have been shown to be susceptible to numerous security risks, including quantum-based attacks. This paper explores the design of quantum-secured networks, presenting crucial insights regarding forensics and security gaps that require immediate attention. The findings indicate that these network designs incorporate post-quantum cryptographic primitives, quantum-resistant authentication protocols, and decentralized trust models. As such, this paper recommends the use of quantum-secured educational networks, comprising of lattice-based post-quantum cryptography and hash-based post-quantum cryptography. Furthermore, recent approaches such as artificial intelligence-based behavioural authentication, adaptive role-based access control, and blockchain-based interoperable auditability must be deployed. In our work, we examine the network structure using the principle of classical and quantum adversarial threats. The major focus is on data protection provisions and traceability. Ultimately, this paper makes key contributions to formal security taxonomy, provides an integrated post-quantum architectural model, and identifies open challenges.

ABSTRACT. Obesity is a complex problem involving interactions of dietary, metabolic and behavioral factors that present a significant public health challenge. Machine learning techniques have been widely employed for obesity prediction; however, many existing approaches have limited interpretability, which prohibits their usage in future healthcare strategies. This study provides a novel data-mining framework that uses an ensemble of the interpretable association-rule mining algorithm FP-Growth to identify obesity-risk patterns from multiple-source data based on the National Health and Nutrition Examination Survey (NHANES). The framework collates demographic, dietary, laboratory, examination and questionnaire data into a single cohesive platform, yielding 9,813 records for complete participants. Continuous health variables were converted into binary risk indicators to facilitate efficient extraction of frequent patterns. Association rules were established with obesity as the target outcome and rated by support, confidence and lift. Fourteen major obesity-related association rules were identified from experimental results. The most significant predictors of obesity were dietary factors, especially high caloric intake, which produced a ceiling effect value of 2.83 and high sugar consumption. On the other hand, increased blood sugar had the greatest predictive accuracy at a confidence of 51.54%. Behaviors such as short sleep duration and a sedentary lifestyle contributed heavily to obesity risk as well. The transparent and clinically interpretable insights provided by the proposed framework may support healthcare analytics, obesity-risk assessment and public health decision-making.

ABSTRACT. The growing potential of wearable healthcare and artificial intelligence has propelled the evolution of continuous electrocardiography (ECG) monitoring into non-clinical settings in recent years. Despite this progress, the deployment of deep learning models for wearable applications remains limited due to constraints on computation, energy consumption, latency, and privacy and security issues associated with cloud-based processing. This is where Tiny Machine Learning (TinyML) comes to the rescue, enabling lightweight AI inference on even the lowest-power, resource-constrained embedded systems. This review provides a high-level organizational structure of wearable ECG monitoring systems using TinyML and summarizes recent developments in integrated artificial intelligence that aid cardiac health technology. First, this review summarizes the basics of wearable electrocardiogram (ECG) monitoring and the TinyML deployment ecosystem. Then, it examines lightweight neural architectures such as convolutional, recurrent-based, transformer-based, and hybrid circuits, particularly for low-power systems. This also includes optimizations of quantization, pruning, compression, and hardware-aware deployment for efficient execution under embedded constraints. In addition, this review covers popular public ECG datasets and evaluation approaches and discusses real-life challenges for deep learning in cardiac arrhythmia diagnosis, including interpretability, clinical validation, generalizability and deployment. Dimensions of future research towards reliable, energy-efficient and clinically-realizable wearable cardiac monitoring systems are discussed.

ABSTRACT. Automatic Speech Recognition (ASR) has emerged as one of the most transformative technologies in modern artificial intelligence, underpinning critical applications in healthcare, education, smart devices, robotics and human-computer interaction. Despite remarkable progress driven by deep learning, transformer architectures and self-supervised learning, a unified analytical framework covering the full trajectory from classical statistical models to contemporary foundation-model approaches remains absent from the literature. This paper addresses that gap by presenting a comprehensive review of ASR systems, tracing the evolution from traditional Hidden Markov Model (HMM) and Gaussian Mixture Model (GMM) pipelines to modern end-to-end, transformer-based and self-supervised architectures. The review synthesizes major ASR techniques, including MFCC feature extraction, DNN-HMM hybrids, CNN and LSTM acoustic models, attention-based encoder-decoders, Transformers, Conformers and self-supervised frameworks such as wav2vec 2.0, HuBERT and Whisper and critically evaluates their relative strengths and limitations. Benchmark corpora and evaluation metrics, including WER, CER, PER and RTF, are systematically examined. Beyond accuracy, the paper provides a multi-dimensional comparative analysis of computational complexity, robustness, multilingual scalability, latency and deployment feasibility across edge and cloud environments. Key challenges, including noise robustness, low-resource and multilingual ASR, accent variability, computational cost, algorithmic bias and data privacy, are thoroughly discussed. Emerging research directions encompassing foundation models, multimodal audio-visual ASR, edge AI and scalable self-supervised pretraining are also identified. By integrating architectural analysis, empirical benchmarking and open-challenge synthesis within a single structured framework, this review provides actionable insights for researchers and practitioners designing the next generation of efficient, equitable and deployable speech recognition systems.

ABSTRACT. This study introduces an all-metal based hexagonal terahertz metamaterial absorber designed for biomedical sensing applications. The proposed structure consists of two metallic layers: a steel plate at the bottom and an upper layer comprising four concentric hexagonal ring resonators of same metal. The absorber demonstrates four significant absorption peaks of 94.0%, 99.2%, 93.8%, and 92.6% at resonance frequencies of 2.852 THz, 3.031 THz, 3.352 THz, and 3.784 THz respectively. Notably, the second resonance peak at 3.031 THz is selected for sensing applications due to its maximum absorption efficiency. At this frequency, the proposed absorber exhibits a quality factor of 591.75 and a Figure of Merit of 359.22, both exceptionally high compared to conventional terahertz sensors. The device can effectively detect changes in the refractive index of biological tissues within the range of n=1.30 to 1.40. The Full Width at Half Maximum bandwidth is calculated to be 5.1221 × 10−3 THz, further confirming its suitability for precise biomedical sensing. In addition, the sensitivity of the structure is determined as 1.84 THz/RIU at 3.031 THz, where RIU represents the Refractive Index Unit. The absorption mechanism of the proposed absorber is thoroughly explained through the analysis of current distribution plots. The proposed absorber demonstrates potential applications in terahertz biomedical sensing.

ABSTRACT. In recent years, acoustic signal analysis has emerged as a promising approach for fault detection in electromechanical systems, providing a non-invasive, low-cost, and real-time alternative to traditional monitoring techniques. This paper presents the design and implementation of a low-cost embedded system for acoustic fault detection. The proposed system is built on a Raspberry Pi Nano platform, integrating a low-power microphone, real-time audio signal preprocessing, and lightweight machine learning models for classification of normal versus faulty operating conditions. The focus is placed on developing an efficient signal processing pipeline that includes noise reduction, feature extraction (time and frequency domain descriptors, Mel-frequency cepstral coefficients), and on-device classification using compact neural network architectures. The embedded setup enables autonomous monitoring without the need for external computation resources, making it suitable for edge deployment in industrial and IoT environments. Experimental validation is carried out using publicly available datasets such as MIMII (Malfunctioning Industrial Machine Investigation and Inspection), as well as preliminary real-time recordings. The results demonstrate that the system achieves reliable fault detection accuracy while maintaining low computational and energy costs, highlighting its potential for scalable deployment in smart maintenance applications.

ABSTRACT. We proposed a hierarchical locking scheme enabled by blockchain to assist nested operations within the decentralized setting. The proposed solution incorporates the blockchain capabilities of immutability, decentralized validation, and execution in smart contracts, unlike the traditional concurrency control mechanisms that use centralized lock management. Upon a subtransaction finishing its working, the respective lock is freed per the hierarchical locking protocol, and the next transactions authenticate the before- and after-images with blockchain verification. This is to guarantee consistency and safe concurrent execution of dependent transactions. The suggested solution enhances concurrency of systems in that, the number of nested transactions running at the same time can be achieved without affecting the correctness of the system because hierarchical dependency control is enforced. The framework offers transparency in execution, resistance to tampering, and trusted auditability by integrating models of layered transactions with blockchain-based validation. Generalized lock compatibility rules and hierarchical access control enable an efficient management of data objects of various granularities.

ABSTRACT. Manual processing of massive camera trap datasets remains a bottleneck in ecological research. Standard YOLO models often struggle with poor illumination and occlusion, ignoring critical metadata context used by ecologists. We propose a multimodal fusion framework integrating frozen YOLOv8 visual features with temporal-spatial metadata, addressing the persistent challenge of data scarcity in wildlife classification through transfer learning and parameter-efficient fusion. Using cyclic temporal encoding and spatial clustering, we fuse 512 dimensional visual features with 256-dimensional metadata embeddings, where is_crepu ∈ {0,1} captures crepuscular activity periods critical for dawn/dusk species identification. Evaluated on 85,247 images across 45 species, our additive fusion architecture achieved 86.97% top-1 accuracy and 98.86% top-5, outperforming concatenation and attention mechanisms while using 40% fewer parameters than competing methods. For edge deployment, our lightest model achieves 86.52% accuracy at 1110 FPS (0.90ms/frame), suitable for real-time inference on Raspberry Pi 5 and Jetson Nano. Ablation studies highlight the value of metadata: removing temporal features dropped accuracy by 4.2%, while unfreezing backbones caused an 8.9% decline due to catastrophic forgetting. Additionally, 26.3% of errors reflected biologically plausible morphological similarities among small mammals. This approach demonstrates that leveraging contextual data significantly improves automated species classification in challenging environments, even under noisy or incomplete metadata conditions. Code and pretrained models are publicly available at https://github.com/AnuruddhaPaul/METADATA_YOLO.

ABSTRACT. For error-tolerant applications like image processing, where minor errors are accepted in exchange for increased performance and energy efficiency, approximate computing has become a successful design model. The design and implementation of approximation adders specifically suited for image processing applications are presented in this study. To lower hardware complexity, a variety of approximation adder topologies are produced by streamlining carry propagation and logic structures. The suggested adder is compared to traditional approximate adders and assessed in terms of accuracy, delay, and power consumption. The approximation adders are included into breast cancer image processing techniques like image addition and filtering to show their usefulness.

ABSTRACT. The Voltage Difference Transconductane Amplifier (VDTA) has been traced out to be one of the active and efficient devices. These devices are also considered for designing of analog circuits to work on different functional aspects. This paper is mainly focused on presenting a newly devise of Schmitt trigger circuit by the use of the single active device, namely, VDTA and is realized without the involvement of passive element. This proved more propitious while fabricating an Integrated Circuits in VLSI design. The circuit, thus proposed avails a bias current of value 150 μA and works at ±0.9 V rail voltage. With the help of bias current, its amplitude can be electronically controllable/tunable. The circuit is greatly resistant to noise and temperature-sensitive. The devised circuit can be used in ADC’s, communication systems, waveform generators, etc. The devised circuit is simulated in a gpdk 180 nm CMOS process availing a Cadence Virtuoso tool. The circuit that has been designed provides a less rate of power dissipation which is of value 270 µW. The executed, results, however equaled with theoretical results which have been experimentally valid and proved by using commercially available ICs LM 13700

ABSTRACT. The Digital systems relays on full adders to perform arithmetic calculations and thus determine how fast and how complicated each computer system will be. In this document, we outline the design of an 18-transistor full adder that is compact in size and reduces the complexity of circuit design without sacrificing the correct logical functions of a full adder. The architecture for this full adder design can be derived directly from the truth table of the full adder; this means that no redundant transistor paths will be created like those performed by a conventional 28-transistor full adder. The full adder circuit was verified with Cadence simulations and all possible combinations for both the Sum and Carry outputs yielded correct results. The functionality of the proposed full adder was further evaluated within the context of a MATLAB based fingerprint matching environment by utilizing the full adder to generate popcount computations used to determine the Hamming distance between fingerprint descriptors. The test results determined that the reduced transistor full adder reliably completes arithmetic processes supporting the calculation of popcount metrics necessary for biometric matching systems. Therefore, an 18T full adder offers a compact arithmetic solution for all applications requiring popcount arithmetic operations, including biometric identification and binary neural networks.

ABSTRACT. Making sure only the right person votes has always been the biggest problem in any electoral system, and this issue gets even more difficult when voting is done online. Traditional e-voting systems have mainly relied on passwords, ID numbers, or one-time codes to manage this role- with these methods, although convenient, allowing impersonation and misuse of credentials. This article takes an alternative approach by putting the voter’s face at the center of the authentication procedure. The system being proposed acquires its biometric data during voter registration, it takes a live facial photo and generates a unique facial embedding for each user at the time of signup. At the moment of voting, the system matches the voter’s live face to three stored references- the registration photo, the facial embedding, and the profile face- before the voting position is given. A similarity cut-off point is used to make this decision, and any try which is less than the set value is totally rejected. Besides, the platform guarantees the vote’s authenticity by labeling each voter as ”voted” immediately after the acceptance of the vote, so that making repeat attempts is not even a possibility even if the subsequent authentications score quite well.Tests with 200 registered voters across five scenarios showed the system averaged 94.5% facial authentication accuracy. It blocked all duplicate votes and kept face detection working consistently. Does real-time biometric verification really outperform current credential methods? The data supports it. Voters get stronger protection when identity is verified by live facial features instead of passwords or IDs.

ABSTRACT. Panoramic dental radiography is a screening tool for detecting caries, periapical lesions, and impacted teeth. However, manual annotation of these images is not feasible for large-scale screening programs. Multimodal large language models (MLLMs) can process images via prompts to detect dental pathologies without task-specific fine-tuning. This study evaluates seven MLLMs—three from Anthropic (Claude), three from Google (Gemini), and one from OpenAI (Generative Pre-trained Transformer, GPT)—using 200 panoramic radiographs sampled from the DENTEX dataset. Performance is assessed at three levels: binary presence per pathology class, instance count matching, and identification of the most clinically severe pathology per image. Gemini 2.5 Pro demonstrates the highest image-level F1 score and the highest severity-detection accuracy. However, Gemini 2.5 Pro did not identify any of the 20 healthy images as normal, whereas Gemini 3.1 Pro, despite lower overall performance, correctly identified 12 of 20 healthy images.

ABSTRACT. In this paper, we present a study on multi-label fault classification in smart grid power systems using a LightGBM-based approach for fault identification. The proposed method is capable of detecting simultaneous faults involving ground (G) and phase conductors (A, B, and C) using three-phase current and voltage measurements obtained from existing datasets, as well as, derived datasets generated from real-world-inspired, large-scale synthetic data. Additionally, we conduct an ablation study in which twenty-five features are ranked according to their contribution to classification accuracy. The study reveals that symmetric component features, including zero-sequence and negative-sequence components, along with vector norms, are among the most discriminative feature groups for smart grid fault identification. The experimental results demonstrate that models trained on the 4.2-million-row synthetic bulk dataset achieve a subset accuracy of 97.89%–98.55% and a Hamming accuracy exceeding 99.4%, substantially outperforming models trained solely on existing datasets. The benchmark dataset used in this study was obtained from Kaggle, where cross-dataset evaluation achieved a subset accuracy of approximately 85.76%. These findings establish best-practice guidelines for scalable, accurate, and overfitting-resistant smart grid fault diagnosis systems.

ABSTRACT. Post-harvest quality assessment of pineapples is still highly dependent on manual inspection, which is slow, subjective, and difficult to scale for industrial sorting lines. This paper presents the Pineapple Status Prediction System (PSPS), an edgefirst IoT-enabled cyber-physical platform for automated pineapple forecasting, ripeness assessment, defect detection, grading, and sorting. The proposed system integrates four processing modules. First, a Cross-Attention Transformer predicts next-day harvest yield using a 10-day weather and production history. Second, a dualmodality ripeness classifier combines load-cell measurements with volatile organic compound sensing. Third, a fine-tuned YOLOv11n model detects visual defects from camera images. Finally, an MQTT-based IoT backbone coordinates sensor nodes, edge inference, dashboard monitoring, and mechatronic sorting. Experimental evaluation showed 1.68% MAPE for harvest forecasting, 93.3% ripeness classification accuracy on 120 physical fruits, and 0.952 mAP@0.5 for visual defect detection. The integrated system achieved 92.6% grading accuracy on 300 physical fruits, improving performance by 4.3% compared with visual-only grading. A 72-hour continuous validation also confirmed 99.97% QoS 1 MQTT delivery and 12 ms median message latency. These results show that the proposed PSPS can support low-cost, real-time, and locally deployable post-harvest automation for pineapple supply chains. Index Terms—IoT, MQTT, edge computing, pineapple quality grading, YOLOv11, Transformer, cross-attention, VOC sensing, FreeRTOS, cyber-physical systems, precision agriculture, sensor fusion

ABSTRACT. Cyber-physical energy systems rely on digital measurements, state estimation, and learning-assisted monitoring, yet these layers are vulnerable to stealthy false data injection attacks that can distort operator awareness without triggering residual alarms. This paper presents a physics-constrained deep generative framework for attack synthesis in power-system state estimation. Conditional generative adversarial networks, autoencoders, and variational autoencoders are compared under a unified physics-aware setting that embeds the nonlinear measurement function, local state-estimation sensitivity, residual preservation, and reconstruction consistency. The framework evaluates generated attacks on IEEE 14-bus, 57-bus, and 118-bus systems using convergence behavior, bad data detection bypass rate, and Jensen-Shannon divergence. Results show that the variational autoencoder and autoencoder achieve stronger residual evasion, whereas the conditional generative adversarial network yields more realistic measurement distributions. Confidential computing and zero-trust tokenization are incorporated as interpretive security layers for protected measurement handling, tokenized provenance, and joint residual distributional trust assessment within state-estimation security decision workflows.

ABSTRACT. This paper proposes a carbon-transparency blockchain framework for renewable-rich electricity, heat, and green hydrogen energy ecosystems. The framework couples Renewable Portfolio Standard obligations with Carbon Emission Trading through a ledger-enabled accounting layer that records renewable quotas, carbon allowances, green certificates, and emission settlements in an auditable form. A ladder-pricing mechanism translates the mismatch between assigned quotas and realized emissions into market signals for low-carbon operation. Power-to-Gas technology is embedded to convert surplus renewable electricity into hydrogen carriers, reduce wind curtailment, and improve cross-carrier flexibility among electricity, heat, gas, and hydrogen subsystems. The scheduling problem is formulated as a mixed-integer linear programming model and solved with the CBC solver. Simulation results indicate that joint participation in carbon and certificate markets lowers total operating cost, while Power-to-Gas deployment increases renewable utilization and supports more credible carbon-accountable hydrogen energy management. This structure provides a traceable basis for market-based decarbonization in industrial-park-level smart energy systems.

ABSTRACT. Language models acquire useful linguistic and generative behavior from the statistical structure of their training data. As public and synthetic corpora increasingly contain duplicated, weakly filtered, low-information, and recursively generated text, it becomes important to measure how low-signal contamination affects model capability under controlled conditions. This paper presents an experimental study of capability degradation in small language models under controlled low-signal textual corruption. A clean corpus is used to construct fixed-size dataset variants with corruption ratios of 0\%, 30\%, 50\%, 60\%, and 80\%. Corruption is generated through a formal taxonomy consisting of repetitive low-entropy text, semantically incoherent text, distractor-dominant text, recursive synthetic text, and shallow high-fluency filler. Small GPT-style decoder-only transformer models are trained from scratch under identical architecture, tokenizer, optimizer, training schedule, and evaluation settings across three random seeds per condition. Degradation is measured using clean-test perplexity, repetition behavior, lexical diversity, prompt adherence, bootstrap confidence intervals, benchmark-style multiple-choice probes, ablation analysis, and a Capability Retention Index. Results from 30 model-run rows show that mixed low-signal corruption increases mean clean-test perplexity from 1654.12 at 0\% corruption to 2199.29 at 80\% corruption, a 32.96\% increase. CRI remains below the clean baseline for every corrupted mixed variant and reaches 0.6139 at 80\% corruption, although it is not strictly monotonic under the current aggregate metric. Ablation results identify repetitive low-entropy and shallow fluent-filler corruption as the most harmful categories under the tested metrics. The study contributes a reproducible framework for data-quality degradation analysis while keeping claims limited to the tested small-model, three-seed setting.

ABSTRACT. This paper proposes a carbon transparency blockchain framework for renewable-rich electricity, heat, and green hydrogen energy ecosystems. The framework couples Renewable Portfolio Standard obligations with Carbon Emission Trading through a ledger-enabled accounting layer that records renewable quotas, carbon allowances, green certificates, and emission settlements in an auditable form. A ladder-pricing mechanism translates the mismatch between assigned quotas and realized emissions into market signals for low-carbon operation. Power-to-Gas technology is embedded to convert surplus renewable electricity into hydrogen carriers, reduce wind curtailment, and improve cross-carrier flexibility among electricity, heat, gas, and hydrogen subsystems. The scheduling problem is formulated as a mixed-integer linear programming model and solved with the CBC solver. Simulation results indicate that joint participation in carbon and certificate markets lowers total operating cost, while Power-to-Gas deployment increases renewable utilization and supports more credible carbon-accountable hydrogen energy management. This structure provides a traceable basis for market-based decarbonization in smart cities.

ABSTRACT. Ceylon cinnamon plays a vital role in Sri Lanka’s agricultural economy however, traditional cultivation practices often result in delayed disease detection, excessive fertilizer use, and reduced productivity. This research proposes SMARTCINNAMON, a smart precision agriculture framework that integrates Internet of Things (IoT), deep learning, machine learning, image processing, and geospatial analysis to modernize cinnamon farming. The system enables real-time monitoring of soil and environmental conditions and supports early identification of leaf fungal diseases using smartphone-acquired images. A convolutional neural network-based model demonstrated strong performance in detecting fungal infections at their nursery stages. The proposed approach shows a reduction in fertilizer usage compared to conventional fixed-schedule methods, without compromising soil quality or crop yield. Yield prediction and disease spread forecasting enhances proactive farm management and timely interventions. This research improves resource efficiency, lowers production costs, and promotes sustainable cinnamon cultivation, offering a practical and scalable solution for small- and medium-scale farmers.

ABSTRACT. Early identification of emotional dysregulation in children aged 4–6 is critical to preventing long-term behavioral and cognitive impairments. However, traditional diagnostic methods often rely on invasive digital technologies, such as facial recognition, which raise significant privacy concerns and lack contextual depth. This paper proposes a privacy preserving, multimodal framework designed to identify emotional states and dysregulation levels in Sri Lankan preschool children through a combination of artistic expression and use of emotion regulation checklist. The methodology integrates two distinct data streams. The first utilizes a Deep Learning pipeline where a Vision Transformer (ViT-L/16) extracts high-dimensional features from a custom dataset of 600 child-generated artworks. These features are optimized via Principal Component Analysis (PCA) and classified using a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel to identify emotions such as happiness, sadness, and anger. To address the "black-box" nature of AI, the system incorporates Explainable AI (XAI) features, including confidence rates and feature heatmaps, providing transparency for educators. The second stream processes the Emotion Regulation Checklist (ERC) via a Linear SVM to categorize dysregulation into low, medium, and high levels. Correlation analysis confirms that the ERC Lability/Negativity subscale serves as a robust ground-truth validator, showing strong correlations (r > 0.50) with total dysregulation scores. Implemented as a web-based diagnostic tool using React.js and Python, the system provides longitudinal time-series visualizations for parents and teachers. By bypassing invasive biometric monitoring in favor of culturally appropriate, non digital inputs, this research offers a scalable and ethically sound solution for early emotional intervention in urban educational settings

ABSTRACT. Traditional handwriting assessment in early childhood education relies on subjective teacher evaluation and becomes particularly challenging in bilingual Sri Lankan curricula involving both English and Sinhala. This paper introduces a multi-metric handwriting quality scoring pipeline for automated evaluation of preschool handwritten letters for children aged 4–6. The proposed system integrates contour-based segmentation, language-specific convolutional neural network (CNN)-based recognition (achieving 86.73% test accuracy on EMNIST for English and 99.04% on a custom Sinhala dataset), and an eight-metric structural analysis framework. The framework combines structural similarity (SSIM), topology preservation, shape invariants, proportion analysis, stroke continuity, alignment, pixel Intersection over Union (IoU), and a 9 × 9 grid-based spatial comparison. The weighted scoring approach shows strong agreement with teacher assessments, demonstrating the reliability of the proposed system. Designed for Sri Lankan early education contexts, the approach provides a scalable, interpretable, and objective solution, addressing the gap in automated Sinhala handwriting assessment tools.

ABSTRACT. Early childhood English proficiency is vital for cognitive and academic growth, yet assessment in low-resource, non-native contexts like Sri Lanka faces challenges from accent variations, child-specific acoustics, and manual scaling limits. We present a novel AI-based multi-stage pipeline using a voice assistant for personalized linguistic evaluation, featuring a custom 3.18-hour dataset (501 recordings from Sri Lankan children aged 4-8) to fine-tune Whisper ASR (Test WER: 27.27%). It integrates fine-tuned Flan-T5 for grammar error type detection and correction (91.47% accuracy), SBERT/spaCy for pragmatic relevance/tense matching, Q-learning for session adaptation, and LIME for interpretable feedback. Linking to Broca/Wernicke areas of brain and emotional influences, the system outperforms baselines in transcription, grammatical error classification (F1=0.82), and personalization, enabling scalable teacher/parent monitoring. This bridges the gap between clinical linguistics and classroom monitoring.

ABSTRACT. AI-powered travel assistants are evolving from static recommendation interfaces toward autonomous systems capable of itinerary planning, information retrieval, constraint handling, and adaptive interaction. However, most existing tourism assistants either remain monolithic or rely on loosely coordinated multi-agent interactions, resulting in weak execution control, limited reproducibility, fragile tool invocation, and poor observability under failure. Meanwhile, emerging interoperability mechanisms such as Model Context Protocol (MCP) and Agent-to-Agent (A2A) communication improve modularity but introduce new challenges in coordinating heterogeneous agents under dependency, safety, and latency constraints. This paper presents a protocol-aware Directed Acyclic Graph (DAG)-based multi-agent orchestration framework for intelligent tourism assistants. The proposed framework models user requests as dependency-constrained execution graphs, enabling deterministic scheduling, bounded parallelism, typed task contracts, structured failure handling, and protocol-level routing across heterogeneous tools. The architecture integrates key components including a multilingual language service, intent classification, a three-tier memory subsystem, a protocol-aware router, a DAG executor, and an NLI-based verifier–repair loop for reliability control. Unlike prior tourism systems that focus primarily on recommendation, this work treats orchestration as a first-class systems problem. The framework is instantiated in a Sri Lanka tourism domain and evaluated across multiple components. The intent classifier achieves a Micro-F1 score of 0.9403 and Macro-F1 of 0.9224 across 7161 samples, while the NLI verifier achieves a Macro-F1 of 0.936. A controlled transport benchmark demonstrates reduced tail latency under MCP Streamable HTTP (p95: 11342 ms) compared to A2A HTTP (p95: 18666 ms). These results demonstrate that explicit DAG-based orchestration improves reliability, execution control, and system observability compared to unstructured multi-agent approaches.

ABSTRACT. Fine motor skills and attentional focus are critical developmental indicators in early childhood, serving as foundations for academic success and emotional self-regulation. Traditional preschool assessments often rely on subjective observations, which can be inconsistent and prone to human error. Research indicates that the maturation of these skills is mediated by the Anterior Cingulate Cortex (ACC), where the dorsal region manages task persistence and the ventral region regulates emotional processing through the amygdala. This study proposes an automated computer vision framework to objectively quantify these markers using standard colouring and tracing activities. The system metrics, including line accuracy, stroke consistency, and pacing for tracing tasks, alongside coverage and boundary control for colouring. An age-adaptive algorithm adjusts evaluation thresholds for children aged 4–6, while a human-in-the-loop interface allows educators to refine scoring parameters based on professional expertise. The framework was empirically validated using 179 colouring samples and 62 tracing samples. The colouring assessment demonstrated high reliability (alpha = 0.8664), and the tracing assessment achieved acceptable reliability (alpha = 0.7900), confirming the internal consistency of the automated metrics. By classifying children into three developmental levels: Advanced, Developing, and Needs Support, this framework provides a standardized, data-driven tool for early intervention. This approach facilitates a more effective collaboration between parents and educators, ensuring targeted support for both fine motor development and emotional regulation.

ABSTRACT. Modern warehouse environments demand intelligent monitoring systems that ensure both operational safety and security. However, existing surveillance solutions remain largely reactive, relying on manual observation or isolated detection mechanisms that fail to address complex real-world challenges such as occluding theft behaviors, unsafe item placement, and varying lighting conditions. This research proposes an integrated vision-based framework that unifies warehouse safety monitoring and theft detection using advanced computer vision and deep learning techniques. The system combines object detection, human pose identification, human activity recognition (HAR), and multi-camera dynamic person re-identification in occlusion scenarios for theft detection and geometry-aware risk analysis and automated shelf edge detection to detect hazardous shelf conditions in real time. Theft-related activities such as loitering and abnormal human–item interactions are identified using activity sequences, while safety risks such as overhanging or fallen items are detected through segmentation-based object recognition and spatial boundary analysis. To enhance robustness, the framework incorporates multiple cameras for handling occluded situations and temporal stabilization techniques to reduce detection instability and false alerts. By integrating behavioral analysis with environmental risk assessment, the proposed system transforms traditional passive surveillance into a proactive monitoring solution. The framework aims to improve warehouse safety, reduce product damage, and enable early detection of theft through accurate, real-time alerts. This research contributes a scalable, multi-modal approach that addresses key limitations in existing systems, including lack of context awareness, poor occlusion handling, and absence of unified safety-security monitoring.

ABSTRACT. The persistent misalignment between academic preparation and evolving industry skill demands presents a critical challenge in employability assessment and career guidance. Undergraduate graduates frequently enter the workforce without the practical competencies and specialized technical skills required by modern industries. This research proposes an integrated employability intelligence framework that combines graph-based skill gap analysis, personalized learning recommendations, skill validation, and adaptive interview assessment to provide a comprehensive evaluation of candidate readiness for industry roles. The system models the employability ecosystem using a heterogeneous knowledge graph that represents relationships among candidates, skills, job roles, learning resources, and projects. A Graph Neural Network (GNN) is employed to predict missing skills through link prediction and compute a continuous readiness score for target roles. Based on the identified skill gaps, the framework generates personalized learning paths and skill-conditioned capstone project recommendations using a teacher–student knowledge distillation approach with LoRA-based fine-tuning. In addition to skill development, the framework incorporates a transcript-based skill validation module that evaluates academic evidence through evidence-weighted scoring and AI-generated quizzes. To further assess practical readiness, an adaptive interview intelligence module simulates job interviews using retrieval-augmented generation and evaluates responses using semantic relevance and speech emotion recognition models. The integration of these components forms a unified system capable of dynamic employability assessment, targeted skill development guidance, and preparation for real-world recruitment processes.

ABSTRACT. Malaria continues to pose a significant global health burden, where accurate and timely diagnosis is critical for effective treatment. Microscopic examination of Giemsa-stained blood smear slides remains the gold standard; however, its reliability depends heavily on staining quality, smear preparation, and operator expertise. These factors introduce variability, increase workload, and can lead to inconsistent diagnostic outcomes. This study presents an integrated system for Malaria diagnosis that incorporates staining quality assessment, blood smear quality analysis, and automated parasite detection within a sequential diagnostic framework. A multi-task deep learning model evaluates staining quality and recommends optimal staining duration, while a computer vision-based approach analyzes smear quality from mobile-captured images using features including smear dimensions, feather edge formation, spacing and positioning, and stain density. Only slides satisfying predefined quality criteria are forwarded to the parasite detection stage, where deep learning models perform infection classification, parasite localization, and species identification. The staining model achieved its highest per-class accuracy of 55.2% on Grade III, representing optimal staining conditions, while time prediction achieved a mean absolute error of approximately 4 minutes. The smear quality assessment module produced an average measurement error of 1.76 mm against expert-recorded ground truth values. The parasite detection and species identification models achieved mAP@0.5 scores of 0.812 and 0.897 respectively. By validating slide quality before microscopic examination, the system limits the exposure of detection models to substandard inputs, reducing both analytical overhead and the risk of misdiagnosis in routine laboratory settings.

ABSTRACT. Data preprocessing remains a persistent bottleneck in production machine learning pipelines, particularly when inputs are heterogeneous in modality, quality, and linguistic origin. Fixed-sequence pipelines apply the same operations regardless of whether the input actually requires them, wasting resources and risking degradation of already-clean data. This paper presents an agentic multi-modal preprocessing framework that addresses these inefficiencies through two complementary decision mechanisms: a locally-hosted large language model (LLM) that dynamically selects and sequences text-processing agents based on input characteristics, and a rule-based decision engine for images that evaluates quantitative quality metrics— Laplacian variance for blur, mean luminance for brightness, and standard deviation for contrast—to select only the agents a given image actually needs. Both pipelines incorporate post transformation validation. Text outputs are validated against a typed Pydantic schema. Image transformations are accepted or reverted based on PSNR and SSIM thresholds, ensuring that no processing step degrades the signal it was meant to improve. The entire system is built around an Apache Kafka streaming backbone that decouples ingestion, routing, processing, and persistence, enabling row-level parallelism for large text files and asynchronous image processing without blocking the API layer. Results are persisted to MongoDB and streamed to a React frontend via Server-Sent Events. The framework demonstrates that combining LLM-guided planning, quantitative quality assessment, and per-step validation produces a more adaptive and resource-efficient alternative to static preprocessing pipelines.

ABSTRACT. The increasing reliance of computer science students on AI-generated code without adequate understanding of underlying concepts presents a significant pedagogical challenge. Students often submit functionally correct code while remaining unable to explain the algorithms, data structures, or design patterns involved, leaving them underprepared for industry-level technical assessments. This paper presents OptiCode, a multi-component intelligent code analysis platform that integrates four AI-powered modules: (1) a Code Concept Extractor using a hybrid Abstract Syntax Tree (AST) and Large Language Model (LLM) pipeline; (2) an Adaptive Weakness Detector that classifies recurring conceptual gaps through machine learning; (3) a Code Refactoring and Risk Analyser employing deterministic ASTbased transformations; and (4) an AI Interview Simulator combining real-time facial emotion recognition with NLP scoring. Experimental evaluation demonstrates that the hybrid AST-LLM approach achieves a macro F1-score of 0.83, outperforming the LLM-only baseline by 16.9%, confirming that static structural analysis substantially improves concept recall. The integrated platform provides educators and learners with actionable, real-time feedback on code comprehension and technical readiness.

ABSTRACT. Cloud-native architectures have fundamentally shifted where security risk lives from runtime systems into version-controlled text files. IaC misconfiguration now accounts for 67% of publicly disclosed cloud breaches [2], yet the tooling most teams rely on was never designed with ZeroTrust Architecture (ZTA) in mind. Enforcing continuous, leastprivileged verification across every infrastructure layer demands more than static rule-matching can offer. We built ZeroGuard to close that gap: a five-component generative AI system that couples proactive IaC misconfiguration repair with continuous ZTA policy generation and enforcement. At its core, a hybrid Transformer-Graph Attention Network (T-GAT) model evaluates per-resource risk alongside ZTA pillar compliance. A Code-LLMbased Generative Remediation agent then produces corrected IaC templates and OPA Rego policies directly from those scores. An Identity Management Agent handles ongoing permission rightsizing, while a Continuous Compliance Orchestrator carries the generated policies through to enforcement. Evaluated on 52,000 templates spanning Terraform, CloudFormation, Kubernetes, and Azure Bicep [1], ZeroGuard achieves detection F1 = 0.947, Zero-Day F1 = 0.658, a 78.4% reduction in critical findings, and a ZTA posture improvement of ∆Ξ = +0.380 [6]

ABSTRACT. Managing modern cloud-native infrastructure is, frankly, a problem that has grown faster than our tools for handling it. Continuous delivery pipelines, ephemeral containers, and sprawling multi-cloud IaC ecosystems create an operational surface that expands combinatorially with organisational scale and static rule-based monitoring simply cannot keep up. This paper presents InfraAgent, a five-agent agentic AI framework de- signed for predictive infrastructure management and autonomous deployment operations in heterogeneous multi-cloud settings. The framework brings together: a Temporal Graph Attention Network (T-GAN) agent that jointly models infrastructure de- pendency topology and multi-variate telemetry for deployment failure prediction across 1 hour, 6 hour, and 24-hour horizons; an LLM-powered Remediation Planning Agent (RPA) that generates context-aware IaC patches, rollout prescriptions, and rollback directives directly from risk scores; a Deployment Safety Agent (DSA) that enforces configurable pre-flight autonomy gates; a Capacity Forecasting Agent (CFA) built on a Temporal Fusion Transformer (TFT) for probabilistic 24 hour to 72 hour demand forecasting; and a Continuous Operational Intelligence Orchestra- tor (COIO) that provides closed-loop governance, cryptographic auditability, and online model adaptation. We evaluate InfraAgent on 47,800 deployment events and 310,000 telemetry snapshots drawn from AWS, Azure, and GCP production environments. The results are encouraging where a Mean Time to Detection of 4.3 minutes (an 85.2% reduction over rule-based baselines), a deployment failure prediction F1 of 0.931, a 76.8% reduction in production incidents per 100 CI/CD runs, and 91.4% capacity forecast accuracy at the 72- hour horizon. Together, these numbers suggest that coordinated agentic reasoning grounded in graph-temporal learning can meaningfully shift cloud operations from reactive firefighting toward continuous, self-correcting infrastructure assurance.

ABSTRACT. Policy-constrained generative AI in financial services sits at the intersection of two regulatory regimes that disagree on almost everything except cost of failure: a model can be plausibly correct and yet violate KYC, suitability, fair-lending, or recordkeeping obligations in ways that scale with throughput. Existing safety layers, whether prompt-engineered guardrails or static keyword filters, neither prove compliance nor leave the audit trail an examiner will eventually ask for. We present PolicyAgent, a five-agent generative AI framework that couples retrieval over a structured policy corpus with formal automated reasoning checks and tamper-evident audit. A Policy Ingestion and Indexing Agent (PIIA) compiles institutional policy documents into a typed reasoning graph; a Reasoning Verification Agent (RVA) discharges per-response obligations through an SMT/automated-reasoning backend; a Generative Drafting Agent (GDA), built on a 13B Code-Llama variant fine-tuned with LoRA, produces responses constrained by the RVA feasibility envelope; an Audit Trail Orchestrator (ATO) records every action in a SHA-256 chained, append-only log; and a Continuous Policy Evaluation Coordinator (CPEC) runs the FinPolicyBench evaluation harness against the live system. Evaluated on FinPolicyBench, a reproducible benchmark of 47,200 policy-constrained queries spanning seven obligation categories across a 90-day controlled CI/CD pipeline, PolicyAgent achieves a policy-compliance F1 of 0.926, a 71.3 percent reduction in policy violations per 1,000 responses relative to a prompt-guardrail baseline, a mean reasoning-discharge latency of 412 ms, and a Policy Compliance Score Pi = 0.918, with full ablation studies confirming each agent contributes independently.

ABSTRACT. Multi-agent generative AI deployments inside enterprises increasingly straddle two opinionated platform stacks at once: a Customer Relationship Management (CRM) system that owns the business object model (Salesforce, Microsoft Dynamics, HubSpot) and a hyperscaler cloud (Amazon Web Services [AWS], Microsoft Azure, Google Cloud Platform [GCP]) that owns the compute and model substrate. The orchestration patterns that make these deployments work, sequential pipelines, hierarchical supervisors, peer collaboratives, blackboards, and the rest, are reinvented by every team because no published catalog describes them at the level of granularity required for engineering reuse. We present PatternForge, a five-agent generative AI framework that turns a curated catalog of seven orchestration patterns into a deployable artifact. A Pattern Selection Agent (PSA) maps a structured task specification to a candidate pattern using a learned scorer over historical orchestrations; a Coordination Topology Agent (CTA) instantiates the chosen pattern as a typed message-bus topology; a Hyperscaler Adapter Agent (HAA) compiles the topology against AWS, Azure, or GCP primitives; a CRM Integration Agent (CIA) materialises the business object bindings against Salesforce or Dynamics; and a Pattern Evaluation Orchestrator (PEO) measures the realised orchestration against a multi-dimensional pattern-quality score Lambda. Evaluated on 37,400 orchestration tasks across three enterprise scenarios (lead triage, dispute resolution, renewal forecasting) over a 90-day controlled Continuous Integration / Continuous Deployment (CI/CD) pipeline, PatternForge achieves a pattern-selection F1 of 0.913, a Lambda score of 0.904 averaged across the seven patterns, a 68.4 percent reduction in median time-to-first-completed-task relative to a single-agent baseline, and a CRM-write integrity beta of 0.993 across 374,000 business-object writes. Single-agent ablation confirms each component contributes independently.

ABSTRACT. This paper introduces a novel adaptive, attention-coupled Long Short-Term Memory (LSTM) architecture developed specifically for real-time scenario recognition and prediction in complex automotive electrical/electronic (E/E) systems. Modern vehicles generate rapidly growing data streams from signals such as current, voltage, and temperature. We address this by monitoring critical signal patterns via a fused LSTM. The proposed dual-path methodology comprises a trend path for long-term pattern modeling and a motif path for short-term pattern recognition, coupled via a bidirectional, attention-based gating mechanism that enables dynamic information exchange. The outputs provide a reliable basis for initiating high-resolution data capture or adaptive system responses once a scenario is identified with high confidence. Experimental results demonstrate significant reductions in mean squared error compared to the individual values and interpretable attention weights that reveal information-exchange patterns. The proposed approach enables robust, noise-resilient forecasts and allows for efficient, data-driven development for future EE architectures.

ABSTRACT. Optimal pressure management in water distribution systems (WDSs) is one of the most efficient approaches to regulate water leakage for water utilities worldwide. The pressure control aiming to minimize excessive pressure can be accomplished through regulating operations of pressure reducing valves (PRVs) placed in the WDSs, and this engineering problem can be formulated as a nonlinear program optimization problem (NLP). Using the Hazen-Williams formula in the formulated NLP will lead to the problems of singular jacobian when solving it by standard nonlinear program solvers. To deal with this problem, the Hazen-Williams formula (H-W) will be approximated by a quadratic function, which can be readily used for formulating and solving the NLP. However, the proper choice of the flow range for taking the approximation, which results in feasible solution when applying to the WDS using H-W model, has not been discussed yet in the literature. In this paper, we propose an approach to improve the NLP solution by determining an effective velocity for taking approximation of quadratic pipe model. With such the velocity value, the solution of the formulated NLP will guarantee minimum pressure satisfaction as applying to the WDSs with H-W model and, at the same time, it results in low excessive pressure. A benchmark case study is taken to demonstrate the efficiency of our proposed approach. The results have revealed that, for each WDS, it is possible to find an effective velocity providing feasibility and high quality of NLP.

ABSTRACT. Water quality monitoring is a crucial task in ensuring water security. With global climate change and the emergence of new pollutants, surface water contamination is becoming increasingly serious. These factors pose challenges in improving monitoring efficiency and developing methods for early pollution detection to alleviate the spread of pollution. This paper proposes an effective solution for monitoring and early detection of water pollution over large areas using unmanned aerial vehicles (UAVs) based in-situ measurement, water sampling and remote sensing. The system operates on the principle of data collection, modeling to estimate water quality parameters, and early pollution detection methods. UAVs can carry direct sensors and sampling devices to survey large areas and provide data for constructing water quality estimation models using remote sensing imagy. The estimation models and multispectral imagery help to create water quality maps and trends of water quality change within the managed area. Areas with unusual changes in water quality over a large area could be sources of pollution. Managers can identify these locations, and UAVs can fly there to directly measure and collect samples, aiding in the verification of pollution sources. A system for monitoring and early pollution detection has been practically implemented in one lake in Hanoi, and the results demonstrate its effectiveness

ABSTRACT. Internet of Things (IoT) sensors continuously generate large volumes of data that must be transmitted to data centers. The resulting high bandwidth demand calls for methods to reduce the number of data points while preserving data quality. In this context, the DREAM (Dynamic Regression Algorithm) has shown that the original data points can be represented by a smaller number of samples while maintaining acceptable accuracy. This paper presents an application of DREAM in an intelligent IoT data collection system. IoT devices fit linear regression models over recent measurements and continuously evaluate the coefficient of determination to decide when data should be transmitted. At the beginning of each segment, two data points are sent to establish the trend of the data. When the coefficient of determination remains above a predefined threshold, no additional data are transmitted; when the coefficient of determination falls below the threshold, the previous data point is sent to mark the end of the current segment. The missing values are then reconstructed at the backend by linear interpolation between consecutive transmitted points. A system of two nodes (indoor and outdoor) collecting temperature and humidity data using different sampling periods is used to compare the proposed approach and the baseline method in sending all data in terms of data sending reduction. The results show that DREAM can reduce the number of transmitted data points while maintaining acceptable accuracy.

ABSTRACT. This paper presents a fault-tolerant human detection system that integrates a 77 GHz FMCW radar module and a visible-light camera through a hybrid sensor fusion architecture. The camera subsystem employs YOLOv4-tiny to perform real-time person detection, while the radar subsystem, built around the Texas Instruments AWR1843 ES2.0 single-chip SoC, leverages on-chip signal processing to offload the host CPU and provides range, angle, and velocity measurements as a processed point cloud. Under normal operating conditions, detected bounding boxes from the camera are fused with projected radar point cloud data to estimate target distance. A three-criterion camera health monitoring mechanism continuously evaluates image quality through focus measure and edge density analysis. When the camera is deemed unreliable due to low illumination, occlusion, or hardware failure, the system autonomously transitions to a radar-only detection mode employing DBSCAN-based clustering and human signature classification, with mode switching completed within 333 ms. This hybrid approach ensures continuous human detection regardless of camera availability. Inter-module communication is established via a TCP-based real-time data stream using a lightweight JSON protocol. Experimental results demonstrate that sensor fusion raises detection confidence from 0.5 in radar-only mode at 1.7 m to 0.9 in fusion mode at 1.7 m, while providing metrically grounded distance estimates that neither modality can deliver alone. The complete system sustains approximately 30 Hz throughput on a CPU-only embedded platform, confirming that robust, fault-tolerant human detection is achievable without dedicated GPU hardware. These findings highlight the proposed system's potential for safety-critical ADAS and indoor monitoring applications.

ABSTRACT. Accurate modeling of Electric Power Steering (EPS) systems is essential for predicting electrical power demand in modern vehicle electrical and electronic (E/E) architectures. This paper investigates sequential deep learning approaches for data-driven EPS torque prediction on real-world driving data (∼2.2 million samples, 4988 driving sequences), using a sequence-level evaluation protocol with block-bootstrap confidence intervals and permutation tests. Explicit temporal modeling via LSTM significantly outperforms non-sequential MLP baselines, with medium-to-large effect sizes (Cohen’s d = 0.70–1.12, p < 0.001), confirming that sequential processing of vehicle dynamics signals is essential for accurate torque estimation. This paper further evaluates whether attention mechanisms, including Simple, Additive (Bahdanau), and Scaled Dot-Product variants, can improve upon the LSTM baseline. None yields a practically relevant improvement, with effect sizes remaining negligible (|d| < 0.1) for Simple and Additive Attention, while Scaled Dot-Product Attention significantly degrades performance (p < 0.001). Analysis of the learned attention weights reveals two failure modes, recency collapse and uniform collapse, explaining why attention cannot contribute beyond what the LSTM hidden state already encodes for short automotive time series. Additionally, a small LSTM with 85K parameters achieves near-identical performance to medium models with 598K parameters at 3.5× lower inference latency, establishing it as the recommended architecture for deployment in resource-constrained automotive systems.

ABSTRACT. Biometric-enabled authentication system are the secured mechanisms in modern era computing. Among all, palmprint authentication systems are rapidly employed for human recognition. Moreover, these systems are vulnerable to a variety of spoof attacks such as printed photographs, video replay, and masks. In this research, we present a hybrid approach for palmprint liveness detection. The approach utilizes the potency of a pre-trained learning model i.e. MobileNetV3. The conception of fine-tuned MobileNetV3 is deployed to extract deep level features from palmprint images. Thereafter, highly discrimination features are selected by using Principal Component Analysis (PCA). A robust Support Vector Machine (SVM) classifier is used to categorize the given image as Bonafide or fake trait. Our approach is trained and evaluated on self-created palmprint spoofing dataset. The proposed approach is lightweight as well as efficient mainly due to usage of MobileNetV3. The proposed approach demonstrates promising performance in terms of Equal Error Rate (EER) 3.08%.

ABSTRACT. Face seal leaks are a major cause of protection loss for N95 filtering facepiece respirators (N95 FFR), yet current fit-testing methods provide only a global fit factor and no information about leak location. Infrared (IR) imaging offers a way to visualise leaks by capturing temperature variations near the mask seal, but most existing approaches operate at the level of full-face images and do not explicitly model the temporal dynamics of breathing or the strong class imbalance between leak and non-leak samples. In this work, controlled infrared videos of N95 FFRs under calibrated leak scenarios, acquired on a respiratory test bench, are used to construct patch-level sequences centred on leak locations along the seal. A Bi-Level Class Balancing (BLCB) strategy is introduced, combining global resampling with targeted preservation of hard negative sequences. Three spatiotemporal deep architectures are evaluated, including two CNN-LSTM models and an attention-based model, all trained with a focal-type loss on balanced sequence sets. Experiments show that the attention-based model achieves high sequence-level performance (up to 98\% accuracy and AUC-ROC close to 0.99 on several regions), while the BLCB scheme effectively mitigates class imbalance. Cross-device evaluation further indicates that models trained on one configuration maintain strong discriminative ability when transferred to the other, with only a modest drop in accuracy and F1-score. These results demonstrate the potential of patch-level spatiotemporal modelling combined with dedicated class balancing for accurate and robust infrared leak detection on N95 FFRs.

ABSTRACT. Driver distraction is a leading cause of road traffic fatalities worldwide. While deep learning-based approaches have significantly advanced the field, most existing methods rely on end-to-end convolutional architectures that are computationally demanding, difficult to interpret, and sensitive to variations in scale and viewpoint. This paper presents a lightweight geometric framework for binary driver distraction detection. Upper-body keypoints are extracted using a High-Resolution Network (HRNet)-based pose estimation model, from which 31 geometric features are derived, including 21 normalized pairwise Euclidean distances and 10 joint angles. The multi-class distraction problem is reformulated as a binary classification task distinguishing normal driving (c0) from all distracted states (c1--c9), and evaluated under a Leave-One-Subject-Out Cross-Validation (LOSO-CV) protocol. Experiments conducted on the State Farm Distracted Driver Detection dataset and the AUCDD (American University in Cairo Distracted Driver) benchmark achieve mean accuracies of 94.60\% (Area Under the Curve -- Receiver Operating Characteristic, AUC-ROC = 0.9806) and 83.17\% (AUC-ROC = 0.9028), respectively. These results demonstrate that compact skeleton-based geometric representations provide an interpretable, scale-invariant, and computationally efficient alternative to deep convolutional models for real-time driver monitoring applications.

ABSTRACT. Hand gesture recognition is an important component of human–machine interaction for robotic control and wearable systems. Unlike vision-based approaches, wearable inertial sensing provides a robust and low-cost solution that is less sensitive to illumination changes and occlusions. This paper presents a subject-independent hand gesture recognition framework based on wearable inertial sensors and deep learning. A dedicated acquisition protocol involving 15 participants and nine dynamic gestures was designed to construct a balanced dataset containing 2,700 gesture sequences. The inertial signals were temporally standardized and enriched using signal magnitudes and temporal derivatives. Four deep learning architectures were evaluated under a rigorous Leave-One-Subject-Out (LOSO) protocol: Multilayer Perceptron (MLP), one-dimensional Convolutional Neural Network (CNN1D), Bidirectional Long Short-Term Memory (BiLSTM), and a hybrid CNN–BiLSTM model. Experimental results show that sequence-aware models significantly outperform the MLP baseline, with CNN1D achieving the best performance of 86.85\% accuracy and 86.70\% macro F1-score. The results highlight the importance of subject-independent evaluation and demonstrate that lightweight convolutional architectures provide an effective trade-off between recognition performance and computational efficiency for future embedded robotic interaction systems.

ABSTRACT. Automatic Facial Expression Recognition (AFER) plays an important role in affective computing and human--computer interaction. However, developing robust AFER systems remains challenging under limited-data conditions due to intra-class variability, dataset imbalance, and reduced generalization capability. In this work, we propose a controlled deep feature learning framework that combines transfer learning with adaptive dimensionality reduction and classical machine learning classification. Facial representations are extracted from a partially fine-tuned ConvNeXt-Tiny backbone and projected into a compact discriminative subspace using dataset-adaptive Principal Component Analysis (PCA). The reduced feature vectors are subsequently normalized using Min--Max scaling and classified using a linear Support Vector Machine (SVM). The proposed framework is evaluated on four benchmark AFER datasets, namely JAFFE, KDEF, WSEFEP, and RaFD, using a stratified 10-fold cross-validation protocol. Experimental results demonstrate stable and competitive performance across datasets, achieving recognition accuracies of 93.14\%, 93.06\%, 89.05\%, and 97.23\%, respectively. Additional cross-dataset experiments further highlight the robustness and transferability of the proposed approach under varying acquisition conditions. The obtained results demonstrate that combining deep feature representations with adaptive dimensionality reduction and lightweight linear classification provides an effective and computationally efficient solution for AFER under constrained data conditions.

ABSTRACT. Infrared (IR) imaging provides a non-invasive approach for evaluating the fit of N95 filtering facepiece respirators (N95 FFRs) and detecting potential leakage. This work presents an optimized U-Net architecture for N95 FFR segmentation from infrared images, a critical component for reliable leakage analysis. The proposed architecture integrates bilinear upsampling, Squeeze-and-Excitation (SE) blocks, and normalization techniques to improve feature representation, stabilize training, and reduce computational cost. The model is trained on one participant dataset and evaluated on a separate unseen dataset containing mask types not encountered during training, enabling assessment of cross-participant generalization and robustness. Experimental results demonstrate that the proposed model achieves an IoU of 98.61\% and a precision of 99.94\%, outperforming the standard U-Net across all evaluated metrics while demonstrating strong generalization capability for practical healthcare and industrial applications.

ABSTRACT. Precise delineation of wood surface defects is essential for automated grading and industrial quality assessment, yet most existing deep-learning approaches focus on image-level classification or bounding-box detection rather than pixel-level localization. This paper presents a semantic segmentation framework for wood defect analysis and, to the best of our knowledge, the first reproducible pixel-wise benchmark on the large-scale public VSB-TUO sawn timber dataset. The proposed pipeline relies on defect-centered Region of Interest (ROI) extraction to reduce foreground-background imbalance while maintaining consistency between training and inference distributions. Five segmentation architectures, including U-Net, attention-gated Fully Convolutional Networks (FCN-AG), and DeepLabV3+, are evaluated under a unified protocol using Dice score, Intersection over Union (IoU), Precision, Recall, and computational cost metrics. Experimental results show that all evaluated models achieve Dice scores above 92.9\%, with FCN-AG-VGG16 obtaining the best performance at 93.46\% Dice and 88.37\% IoU. Despite substantial architectural differences, all models converge within a narrow performance range, revealing a dataset-level performance ceiling where data characteristics dominate segmentation quality more than model complexity. The results suggest that future improvements are more likely to arise from higher-quality annotations, higher-resolution inputs, and downstream defect-severity analysis than from increasingly complex segmentation architectures.

ABSTRACT. Palmprint-based biometric authentication has attracted increasing attention due to the rich discriminative patterns contained in palm texture and line structures. However, most existing deep learning approaches rely on multi-class classification strategies that require full system retraining whenever new users are enrolled, limiting scalability in real-world deployments. This paper proposes a scalable palmprint authentication framework combining deep feature extraction with user-specific one-class modeling. Deep representations are extracted using a pre-trained ConvNeXt network, followed by a two-stage dimensionality reduction process based on global Principal Component Analysis (PCA) and user-specific One-Class PCA (OC-PCA). Unlike conventional multi-class approaches, the proposed method relies exclusively on genuine samples during training, enabling independent user enrollment without retraining the entire system. The proposed framework is evaluated on the publicly available CASIA-Palmprint-V1 and IITD-Palmprint-V1 benchmark datasets using a five-fold cross-validation protocol. Experimental results demonstrate competitive authentication performance, achieving Correct Recognition Rates (CRR) of 95.13\% and 91.11\%, respectively. The obtained results confirm that combining deep feature representations with one-class modeling provides an effective and scalable solution for biometric authentication systems.

ABSTRACT. Audio deepfake detection has become a critical challenge as recent advancements in text-to-speech and voice conversion technologies increasingly threaten the integrity of voice-based biometric systems. Existing countermeasure systems often struggle to generalize beyond the controlled conditions of their training data, exhibiting poor performance when exposed to real-world channel distortions, codec compression, and unseen synthesis methods. In this paper, we propose a progressive multi-phase hybrid framework for audio spoof detection that systematically evolves from handcrafted spectral features towards self-supervised learning representations. Our proposed framework/ architecture combines a task adaptive fine-tuned Wav2Vec 2.0 Base encoder as a front-end feature extractor with a two-layer Bidirectional Gated Recurrent Unit (BiGRU) as a sequence-level back-end classifier, trained on the ASVspoof 2019 Logical Access (2019 LA) dataset. The proposed system achieves an Equal Error Rate (EER) of 0.84% and a minimum tandem Detection Cost Function (min t-DCF) of 0.0069 on the ASVspoof 2019 LA evaluation set, demonstrating competitive in-domain performance. The cross-dataset evaluations on ASVspoof 2021 LA (2021 LA), ASVspoof 2021 Deepfake (2021 DF), the In-The-Wild dataset (ITW), and the Fake-or-Real (FoR) datasets reveal systematic generalization challenges under telephony channel conditions, codec-compressed audio, and physically re-recorded speech, with EERs ranging from 9.56% to 29.16%. Moreover, attack-wise analysis across thirteen spoofing systems confirms strong detection of neural vocoder and waveform concatenation attacks, while highlighting the sensitivity of the model to unseen acoustic channel conditions. These findings motivate future work on cross-domain robust training strategies and multi-condition augmentation for real-world deployment of voice anti-spoofing systems.

ABSTRACT. This Artificial Intelligence (AI) is increasingly used in recruitment processes by IT companies to enhance efficiency, reduce time-to-hire, and support decision-making. However, the use of AI-based recruitment systems also raises concerns related to organizational readiness and the risk of algorithmic bias, which may lead to unfair or discriminatory hiring outcomes the purpose of this study is to examine the level of AI readiness and assess bias risk in AI-based recruitment practices in selected IT companies in Bangalore. This study aims to examine the relationship between artificial intelligence (AI) readiness and perceived bias risk in AI-based recruitment, with a particular focus on the role of human factors in influencing organizational readiness within the Information Technology (IT) sector. A mixed-methods approach was adopted, combining quantitative survey data and statistical analysis. Primary data was collected through a structured questionnaire administered to HR professionals, recruiters, and technical staff involved in AI-supported recruitment. Statistical tools such as descriptive analysis, reliability testing, correlation, and regression were used for data analysis. The findings indicate human factors, including employee skills, training, and technological competence, have a significant positive impact on AI readiness. Furthermore, findings suggest that AI-based recruitment systems are perceived to introduce bias in candidate selection processes. The study also highlights that higher levels of AI readiness can contribute to mitigating bias risks, emphasizing the importance of organizational preparedness in ensuring ethical AI implementation. While IT companies exhibit moderate to high AI readiness, gaps exist in governance frameworks and systematic bias mitigation practices. A significant relationship was observed between AI readiness and effective bias risk management. The study highlights the need for stronger ethical guidelines, governance mechanisms, and continuous bias audits to ensure fair and responsible AI-based recruitment.

ABSTRACT. This paper explores the processes that are fueling the Artificial Intelligence (AI) adoption in digital marketing contexts by considering the relationship between the AI perception, the usage behaviour, and the behavioural intentions. The research relies on modelling equation (SEM) in the analysis of data by using structural Data on 450 marketing professionals in various industries. This article validates holistic paradigm that will justify the transformation of AI perceptions into the behaviour of use and plans in the future and customizing them based on the relevant demographic and experience variables. In favour of measurement validity, the SEM analysis indicates a good fitment of the model (CFI = 095, RMSEA = 0048) and high factor loading. (074-088) The Results section demonstrates that there is positive significant relationship between AI usage, Behaviour and perception intentions. The Control variable analysis indicates that AI familiarity (0.41), experience (0.28), and data analysis capabilities (0.36) have a strong impact on the AI perceptions, whereas demographic factors have insignificant impacts. The model accounts 31.5 and 42.1 percent of the variance in behavioural intentions and AI perceptions respectively. These results apply technology acceptance theories to AI settings and give useful information to organizations adopting AI technologies. The findings indicate that the optimal approach to AI adoption must involve both perception-making and facilitating the usage experiences, with the significance of experiential learning being the key to further AI interest in digital marketing settings.

ABSTRACT. The initialization process plays a significant role in the exploration capacity of metaheuristic search algorithms in the high-dimensional search space. In this paper, we proposed three hybrid initialization frameworks: "Sequential Application," "Quality-Based Selection," "Adaptive Hybridization," as well as the conventional "Merge & Combine" for the Mountain Gazelle Optimizer algorithm by combining random initialization with ten advanced initialization techniques: OBL, GPS, Halton, Chaotic Map, LHS, DUD, DLU, WELL, Torus, and Sobol. The performance of the proposed initialization frameworks was tested using thirteen benchmark functions in 100D space by conducting 30 independent runs using best, worst, average, and STD fitness value criteria. The proposed hybrid initialization frameworks dominated the performance of the eleven individual initialization techniques on ten out of thirteen test functions, while producing identical top results on the remaining three functions. "Merge & Combine" using DLU produced the best results in the average rank table, which was further validated by the statistical analysis using the Friedman and Wilcoxon tests, showing no significant difference between the proposed frameworks but proving the practical efficiency of the proposed hybrid initialization techniques in multimodal optimization.

ABSTRACT. Abstract— Optimizing the hyperparameters of a machine learning model remains a critical bottleneck in achieving peak performance from a model. In this paper, we introduce Fuzzy-Adaptive Bayesian Hyperparameter Optimization (FABHO), a novel algorithm for automatically tuning the hyperparameters of an XGBoost model. FABHO was evaluated against six different approaches on four classification and two regression datasets. The results of the one-way ANOVA and Bonferroni’s paired t-test tests show that FABHO achieves the highest predictive performance on three of the six datasets and the second-highest on the remaining three datasets. These results indicate that FABHO is a robust and generalizable approach for hyperparameter optimization in machine learning.

ABSTRACT. This study investigates the impact of cell aging on current distribution and State of Charge (SOC) estimation accuracy within a 2p2s lithium-ion battery pack configuration. Unlike conventional studies focusing on series-connected strings, this research specifically examines the weakest link shift phenomenon in parallel configurations. A first-order RC equivalent circuit model (ECM) was developed to simulate the dynamic behavior of healthy and aged battery branches. A robust Coulomb Counting (CC) approach was employed to analyze the direct effects of internal resistance mismatch. Simulation results, conducted under a repetitive current profile, reveal a 60% current imbalance between parallel branches due to the resistance disparity. This imbalance leads to a 12% SOC divergence by the end of the cycle. The findings highlight a critical security risk: while the aged cell limits the system initially, the healthy branch eventually becomes the new limiting factor due to accelerated current stress, reaching the lower cutoff voltage faster. These quantitative results emphasize the necessity for BMS algorithms to adopt branch-specific monitoring to prevent secondary aging of healthy cells.

ABSTRACT. Artificial Intelligence (AI) and hybrid work models have drastically changed organizational productivity and employee effectiveness in the workplace. Using job engagement as a mediating variable and digital literacy and organizational support as moderating factors, this study investigates the combined effects of AI integration and hybrid work on employee performance. A structured questionnaire was used to gather data from 300 IT workers using a quantitative research design. Reliability, correlation, regression, mediation, and moderation tests were among the analyses carried out using SPSS. The findings show good sample adequacy (KMO = 0.912, p < 0.001) and robust reliability (Cronbach's α = 0.846–0.889). Regression research shows that employee performance is highly impacted by AI integration (β = 0.29, p < 0.001) and hybrid work (β = 0.21, p < 0.01), with work engagement emerging as the biggest predictor (β = 0.45, p < 0.001). Organizational productivity is highly impacted by employee performance (β = 0.79, p < 0.001). Work involvement partially mediates the linkages between AI, hybrid work, and performance, according to mediation analysis. The results of the moderation show that the associations are strengthened by digital literacy (β = 0.18, p < 0.01) and organizational support (β = 0.20, p < 0.01). The results underline the significance of combining organizational and technology resources and support the Job Demands–Resources model. By creating an integrated framework that combines AI and hybrid work, this study contributes and provides useful insights for raising productivity and performance in contemporary enterprises.

ABSTRACT. Coverage map estimation is fundamental in planning wireless communication systems to enhance service quality and minimize operational costs. Although traditional deterministic propagation models offer high precision, their computational costs and processing times increase exponentially, especially in urban areas. In this study, a ResNet-based Conditional Variational Autoencoder (ResNet-CVAE) architecture is proposed for rapid and accurate coverage map generation in scenarios involving multiple diffractions. In order to generate dataset, a dynamic ray-tracing code integrating Geometric Optics (GO) and the Uniform Theory of Diffraction (UTD) was developed. By processing obstacle geometry and transmitter locations as both numerical and spatial condition information, the proposed deep learning model successfully captures abrupt signal level drops and physical shadowing effects behind obstacles. Experimental results demonstrate that the ResNet-CVAE model produces high accurate results with significantly lower computational overhead compared to traditional methods and adapts effectively to complex obstacle configurations. This approach offers significant potential for real-time analysis in network planning and base station placement processes.

ABSTRACT. Accurate and efficient prediction of signal coverage is of critical importance for wireless network optimization and base station planning. This study advances a baseline one-dimensional convolutional neural network (1D-CNN) architecture previously proposed by the authors by integrating a attention mechanism. The proposed multi-branch architecture is trained on a physically consistent dataset generated using the Uniform Theory of Diffraction (UTD) and Geometrical Optics (GO) methods. The results demonstrate that the integration of the attention mechanism reduces the average root mean square error (RMSE) from 1.267 dB to 0.954 dB, corresponding to a 24.7% improvement in overall performance. Furthermore, error distribution analyses reveal that the proposed model produces significantly more stable and lower-variance predictions, particularly in complex propagation scenarios. In addition, the model achieves more than 200,000× speed improvement compared to conventional ray tracing methods, providing a powerful solution for real-time coverage prediction at 1-meter spatial resolution.

ABSTRACT. This study presents a high-performance framework for the identification of 15 different UAV remote controllers using raw Radio Frequency (RF) signals. To address the security risks posed by the rapid proliferation of Unmanned Aerial Vehicles (UAVs), a specialized 1-Dimensional Convolutional Neural Network (1D-CNN) architecture is developed. This model eliminates the computational overhead of manual feature engineering by learning discriminative features directly from time-domain data. A dedicated burst detection algorithm is implemented to capture signal onsets, ensuring optimal data representation and noise reduction. The proposed methodology achieves a classification accuracy of 99.87% demonstrating significant discriminative power even between hardware-identical controllers. Experimental results indicate an average inference latency of 10.42 ms confirming the system's suitability for highspeed real-time identification. These findings suggest that the proposed approach offers a scalable and efficient alternative to traditional feature-based methods for critical defense applications.

ABSTRACT. This study presents hierarchical 1D-CNN-based models that learn directly from raw audio waveforms for three dimensional positional estimation of Unmanned Aerial Vehicles (UAVs). Unlike conventional feature extraction methods, the proposed approach employs three specialized models for azimuth, distance, and height estimation, enabling high-precision modeling of phase and amplitude characteristics in acoustic signals. Experimental results on the UaVirBASE dataset demonstrate accuracies of 98.44% (1.41° MAE) for azimuth estimation, 98.05% (0.20 m MAE) for distance estimation, and 94.14% (0.59 m MAE) for height estimation. Furthermore, the system processes 5-second audio segments within 11.9–18.6 ms, demonstrating its capability to meet real-time operational requirements for UAV tracking. The results indicate that the proposed hierarchical architecture outperforms traditional methods such as MFCC, STFT, and LFCC in both positional accuracy and computational efficiency.

ABSTRACT. The rapid growth of unmanned aerial vehicles (UAVs) has raised critical security concerns, especially in restricted and urban areas. Identifying drones through radio frequency (RF) signals has therefore become an important research focus. In this study, a feature-based classification approach is proposed to distinguish different drone controllers using their RF signal characteristics. The method extracts representative features from the meaningful portion of each signal. These features capture statistical, spectral, and envelope properties, enabling a compact and informative representation. The resulting dataset is evaluated using multiple supervised learning algorithms, including XGBoost, Support Vector Machines (SVM), and Artificial Neural Networks (ANN). To ensure reliable evaluation, a group-based data splitting strategy is applied to prevent data leakage. The experimental results show that the proposed approach performs well in classification tasks, with XGBoost achieving the strongest results.

ABSTRACT. Sound-based fault detection has gained significant attention due to its non-intrusive nature and low cost, as it does not require physical intervention in industrial systems. This study experimentally evaluates the performance and generalization capability of one-dimensional convolutional neural networks (1D-CNNs) under limited data and noisy conditions. The experiments were conducted using audio data from four different machine IDs (ID00, ID02, ID04, ID06) in the industrial fan category of the MIMII dataset. Within the scope of this study, two scenarios are considered: (i) training and testing on the same machine ID, and (ii) training and testing on different machine IDs, enabling the evaluation of both intra-device performance and cross-device generalization. The obtained results demonstrate that the proposed 1D-CNN model achieves high performance even under challenging conditions with a low signal-to-noise ratio (SNR) of −6 dB. In tests conducted on the same machine ID, the model attains an accuracy ranging from 99% to 100%, successfully distinguishing the normal class with 100% accuracy and the abnormal class with over 98% accuracy. These findings indicate that 1D-CNN architectures can effectively learn discriminative features directly from raw audio signals despite the presence of noise. However, performance degrades significantly when evaluated across different machine IDs, highlighting the impact of domain shift. In conclusion, while the proposed model performs effectively under controlled conditions, it exhibits notable limitations in cross-device scenarios. Future work will focus on improving generalization through self-supervised representation learning to extract domain-invariant features and domain adaptation techniques.

ABSTRACT. This paper presents the design, implementation, and evaluation of MarketMatic, an AI-powered multi-tenant Smart Assistant leveraging Retrieval-Augmented Generation (RAG) to deliver accurate, context-grounded customer support for Small and Medium Enterprises (SMEs). Each business tenant uploads domain-specific knowledge base documents (PDF, DOCX, TXT), which are processed through a section aware text chunking pipeline, embedded using the Nomic Embed-Text model (768 dimensions), and stored in per-tenant ChromaDB vector collections enforcing strict data isolation. User queries are resolved through an adaptive nine-stage retrieval pipeline: semantic vector search with dynamic threshold adjustment, near-duplicate deduplication via overlap coefficient filtering (threshold = 0.88), diversity-aware chunk selection, keyword-based fallback retrieval, and optional cross encoder reranking using BAAI/bge-reranker-base. Response generation employs LLaMA 3.1 8B on a Modal A10G GPU, with transparent failover to the Groq Cloud API. The platform further incorporates a three-tier role-based access control model, an agentic tool layer with five configurable business intent tools, cross-session persistent memory through LLM extracted fact distillation, AI-generated session analytics, and automated knowledge-gap detection. Experimental evaluation across eight business domain categories demonstrates a mean retrieval confidence of 0.81 and a 13.2 % improvement over fixed-window chunking, with high functional accuracy on domain-specific queries with complete source traceability, though broader evaluation across additional domains remains a direction for future work.

ABSTRACT. Large Language Models (LLMs) hold increasing potential for supporting autistic learners through personalised instruction, adaptive content generation, communication scaffolding, and structured learning support. However, most existing AI educational systems are developed without grounding in the lived realities of parents, tutors, and educators who provide educational support to autistic learners, particularly within low-resource contexts such as India. Current systems also frequently reflect neurotypical communication assumptions that create accessibility barriers for neurodivergent users. This study employed a qualitative requirements analysis to investigate how LLM-based educational tools may be more effectively designed for neurodiversity-informed educational contexts. The study combined semi-structured interviews with eight Indian parents, teachers, and tutors, alongside a netnographic analysis of 14 global online sources. Findings revealed that participants already engage in extensive adaptive educational practices involving interest-based learning, multimodal instruction, visual scaffolding, and flexible pacing. The analysis identified key system-level requirements for neurodiversity-aligned LLM tools, such as increased interest-based anchoring, goal-oriented interactions, privacy settings, multilingual accessibility, and simplified interfaces for parents. The study also identifies the “neurotypical default problem” and proposes mitigation through neurodiversity-affirming communication styles. The paper contributes empirically grounded design requirements and proposes a human-in-the-loop architectural framework positioning caregivers as mediators of AI-assisted learning.

ABSTRACT. Neurodivergent individuals, particularly those with ADHD and autism, frequently experience difficulties related to time awareness, task management, planning, and temporal regulation in everyday life. Existing productivity and assistive technologies often rely on neurotypical assumptions of consistent attention, rigid scheduling, and linear task management, limiting their effectiveness for neurodivergent users. While prior research has examined temporal cognition and executive functioning, limited work has explored how AI-based assistive systems can support neurodivergent experiences of time within non-Western contexts such as India. This study investigates time awareness, compensatory strategies, and technology use among neurodivergent adults in India through semi-structured interviews with 19 participants. Using content analysis, the study examines how participants manage time-related demands, engage with digital tools, and conceptualise ideal forms of technological support. Findings reveal heavy reliance on external temporal scaffolds, cognitive offloading systems, and interpersonal accountability, alongside dissatisfaction with rigid, impersonal, and cognitively demanding productivity applications. Based on these findings, the paper proposes a conceptual framework for AI-based personal assistants designed for neurodivergent users, emphasising proactive outreach, contextual personalisation, multimodal interaction, human-like scaffolding, and integration within familiar communication platforms such as WhatsApp.

ABSTRACT. One of the most common and deadly forms of cancer worldwide is skin cancer. Exposure of the skin to ultraviolet (UV) rays from direct sunlight is the primary cause of this type of cancer. For early-stage diseases to be successfully treated, early and precise identification is essential. In this paper, we propose a technique that extracts global features from an efficient ViT-16 model by turning the image into a sequence of $16 \times 16$ patch tokens, supplemented by positional embeddings. These tokens pass through twelve Transformer Encoders using Multi-Head Self-Attention to produce the final 768-dimensional feature vector which will be used for the final classification task. In order to test the ViT-based classifier, a challenging skin lesion dataset with seven classes was used which is called HAM10000 database. The dataset consists of 10,015 images and it is highly imbalanced. The proposed lightweight model was trained on 80\% of data and tested on the remaining 20\% with 100 epochs utilized during training stage. It achieved a high classification accuracy of 92.62% which outperforms state-of-the-art methods applied on the same dataset.

ABSTRACT. This paper aims to explore an MVVM (Model View ViewModel)-based mobile application that plots a graph spanning multiple data points. In doing so, readers will gain insights into how the trading apps draw charts and manipulate the data line function to provide a perfect visual representation of how the data is being shaped. The paper will also provide a comparison of UIKit implementation. It will also provide an Architectural framework that readers can use to develop their custom mobile applications. The source for the same will be provided in form of a Github repository.

ABSTRACT. This research study focuses on improving the performance of AC voltage regulators using a variety of control methodologies. Three main control systems are investigated: Phase Angle Control (PAC), Pulse Width Modulation (PWM), and Integral Cycle Control (ICC), also known as On-Off Cycle Control. Currently, Phase Angle Control and PWM methods are widely used but generate a large number of harmonics, resulting in poor circuit performance. The paper describes the design and modeling of a thyristor-based AC voltage controller that uses Integral Cycle Control, Pulse Width Modulation, and Phase Angle Control approaches. Simulation tasks were carried out using three software platforms: MATLAB, Multisim, and PSCAD. The use of the FFT block in the PSCAD software permitted the extraction of the direct harmonic spectrum, which improved the analysis process.

ABSTRACT. Voltage-Oriented Control (VOC) is a widely adopted control strategy for three-phase active rectifiers based on the Voltage Source Rectifier (VSR) topology. In this paper, a cascade control structure is developed, in which a linear Proportional–Integral (PI) controller is employed in the outer DC-link voltage control loop to ensure power balance, while the inner grid current control loop is designed using a Linear Quadratic Gaussian (LQG) controller. By integrating an optimal Linear Quadratic Regulator with a Kalman state estimator, the LQG controller enables current control based on estimated system states in the presence of measurement noise and model uncertainties. The proposed cascade PI–LQG control structure aims to enhance current waveform smoothness, improve Pulse Width Modulation (PWM) modulation quality, and increase robustness against operating disturbances. Simulation results demonstrate that the proposed method maintains high grid-side current quality and achieves improved stability under voltage and load variations, indicating its potential applicability in three-phase active rectifier systems requiring high power quality.

ABSTRACT. Mental health conditions such as depression and anxiety are increasingly reflected in how people behave and communicate on social media platforms. However, most existing automated detection systems focus only on what users write, ignoring important behavioral patterns such as when they post, how often, and where. This study proposes a hybrid deep learning framework that classifies Reddit users into at-risk and severe mental health risk categories by combining both text and behavioral signals. A fine-tuned Bidirectional Encoder Representations from Transformers (BERT) encoder captures the meaning behind user posts, while a two-layer Bidirectional Long Short Term Memory (BiLSTM) with attention models patterns in user behavior over time. The key contribution of this work is the fusion of these two complementary signal types - linguistic and behavioral - into a unified multi-factor severity scoring system, enabling more personalized and context-aware risk assessment. The model is trained on a balanced dataset of 3,500 Reddit users and achieves an accuracy of 82.49%, precision of 76.44%, recall of 92.97%, F1-score of 83.90%, Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 91.36%, sensitivity of 92.97%, and specificity of 72.40% at an optimized threshold of 0.380. These results show that incorporating behavioral context alongside text significantly improves identifying behavioral risk patterns associated with mental health, with potential to support early intervention efforts in community-facing early awareness settings

ABSTRACT. Conventional vehicle diagnostic tools, such as On-Board Diagnostics (OBD), may not recognize small mechanical problems with the engine's mechanical parts. In this paper, the authors propose the use of an artificial intelligence-based engine fault diagnosis system based on acoustic signals as a predictive vehicle maintenance approach. The engine states that the proposed system can recognize include Normal, Clicking, Knocking, Rod Knock, and Timing Belt. Audio signals were obtained from online sources such as TikTok, YouTube, Facebook, and Hugging Face. Three deep learning models were tested: MobileNetV2, ResNet50, and YAMNet. The results were obtained in terms of accuracy, precision, recall, and F1 score. The best model among all was MobileNetV2, with a high accuracy of 94.38% for training, 84.38% for validation, and 67.86% for real-world applications. The model correctly identified 19 out of 28 vehicles. The ResNet50 model had a high accuracy of 78.12%, while the YAMNet model had an accuracy of 43.75%. It was also observed that the fuel type had an impact on the classification results. The system achieved an accuracy of 75% in gasoline engines, whereas the system achieved an accuracy of 50% in diesel engines.This is because there were more gasoline engine recordings in the dataset. A web-based diagnostic tool was also implemented to record, analyze, and report engine sound in real-time to enable non-technical users to utilize it. The results indicate that deep learning with acoustic analysis is a non-invasive method for early fault detection in engines.

ABSTRACT. Investment decision-making in emerging markets such as the Colombo Stock Exchange is constrained by fragmented analytics, limited interpretability, and delayed incorporation of market-relevant information. Investors commonly rely on isolated financial indicators and short horizon price models that lack an integrated decision-support structure. This study proposes InvestWise, an artificial intelligence driven portfolio planning platform tailored to the Colombo Stock Exchange. The framework integrates four components: a two-stage XGBoost model for detecting and classifying market-moving financial news; a forward-looking financial risk module using Support Vector Regression to forecast Altman Z Score trajectories; a dividend cut prediction model based on Elastic Net logistic regression; and an ensemble price forecasting model combining XGBoost, LightGBM, and CatBoost for five-day-ahead predictions. Empirical evaluation using Colombo Stock Exchange data from 2012 to 2025 demonstrates consistent predictive stability and improved directional accuracy. By embedding explainable artificial intelligence mechanisms across modules, the proposed system provides a scalable, transparent, and context-aware decision support platform for portfolio management in emerging markets.

ABSTRACT. The identification of vehicle information has increased significantly over the past few years to help detect stolen vehicles and traffic violations. One important application of computer vision technology is vehicle License Plate (LP) detection and recognition. Currently, Iraq uses two types of plates: the first contains numbers, letters, and Arabic words, while the second is in English, a future system to be used throughout Iraq. The proposed system uses deep learning to detect vehicle LPs, employing the You Only Look Once version 8 (YOLOv8) algorithm and the PaddleOCR algorithm to recognize LP types. Initially, the image containing the LP is segmented using a predefined grid, and the LP is then localized using a bounding box. After that, the determined LP is extracted to determine its type and identify the text it contains. The system achieves a 98% recognition rate using the well-known Open Images dataset for the vehicle category. On the other hand, the accuracies of detecting Iraqi LPs of the first and second forms are 97.5% and 100%, respectively. Besides, the accuracies of identifying the two types of Iraqi LPs are 96.58% and 98.75%, respectively. Compared to related work, the proposed system performs better in LP detection and recognition.

ABSTRACT. Breast cancer remains a highly deadly disease that spreads rapidly among women worldwide. A significant control measure is needed to limit its rapid spread and save women's lives. When tumor sensitivity is uncertain, early detection becomes critical. Various computer technologies and models have been developed to detect breast cancer at an early stage. This study presents a hybrid model that combines a classical Convolutional Neural Network (CNN) with quantum deep learning, using the INbreast mammography database across three models. The first model is a fully CNN classifier. The second integrates CNN feature extraction with adjustable quantum-classification circuits, thereby forming a CNN-quantum hybrid model. The third model improves performance by combining a CNN with a quantum classifier. The first has achieved an accuracy 97.5%, the hybrid has achieved 99.16%, and the third achieved 99.58% accuracy with flawless sensitivity. These consequences highlight the impact of quantum machine learning on the recognition of complex patterns in medical imaging data and support additional research using superior datasets and advanced quantum systems.

ABSTRACT. Handwritten medical prescriptions remain a significant source of medication errors in developing regions, where illegible handwriting and complex pharmacological terminology cause misinterpretation of drug names, dosages, and administration frequencies. Existing optical character recognition systems focus on professional transcription and fail to bridge the comprehension gap between clinical notation and patient-accessible language. This paper proposes a modular multi-stage transformer based pipeline that integrates YOLO11s for prescription region detection, TrOCR with a Vision Transformer encoder for handwritten text recognition, BioBERT for clinical named entity recognition across drug names, dosages, frequencies, and durations, and GPT-4.1-mini for safety-constrained explanation generation that expands medical abbreviations into simplified instructions at a sixth-grade reading level while reducing unsafe inference. The fine-tuned TrOCR-ViT achieves 98.4% character accuracy and 97.8% word accuracy on the best seed. BioBERT yields a micro-averaged F1 of 0.94 across all entity types. End-to-end evaluation on 100 handwritten prescription samples achieves 76% complete entity-level interpretation accuracy.

ABSTRACT. Squirrel Cage Induction Generators are widely used in renewable energy systems, but their reliability is threatened by stator short-circuit faults. Traditional steady-state diagnostic methods are often rendered ineffective in modern applications because closed-loop controllers actively mask fault signatures. To address this challenge, this work introduces an event-driven fault classification framework specifically designed to detect and classify masked stator faults, namely inter-turn and inter-winding short-circuits. Instead of computationally expensive continuous signal analysis, the proposed methodology isolates brief, high-frequency transient disturbances using an adaptive, derivative-enhanced event detection mechanism. Discriminative features are extracted from these localized transient intervals, with a focus on derivative-based voltage representations. Finally, a novel confidence-weighted aggregation strategy is introduced to combine event-level predictions into robust file-level decisions. Experimental validation using a high-resolution dataset demonstrates that the proposed framework significantly improves diagnostic accuracy. The results show that derivative-based voltage features provide superior class separability, achieving a file-level Area Under the Curve of approximately 0.89, demonstrating the effectiveness of the targeted transient evaluation strategy in closed-loop systems.

ABSTRACT. Purpose: This investigation aims to investigate the conjoint impact of FinTech adoption (FA), green finance (GF) and financial literacy (FL) on the development of the sustainable consumer behaviour (SCB) in Indian digital financial landscape. It further scrutinises the mediating role of financial literacy in the transformation of the digital financial engagement and green financial awareness into sustainability-oriented behavioural outcomes. Design/methodology/approach: A quantitative cross-sectional design was adopted whose results were based on the survey data collected from 430 active users of financial digital services in both urban and semi-urban areas in India. Partial Least Squares Structural Equation Modelling (PLS-SEM) was used through the software SmartPLS 4.0 to measure construct reliability, convergent and discriminant validity, structural relationships and indirect mediation relationships. Findings: The analysis supports strong psychometric properties on all constructs and find that both FA and GF significantly improve on FL as well as on SCB. Green finance has a large predictive relevance of FA, FL, and SCB. Financial literacy was a powerful mediator in the relationships of the FA - SCB and the GF - SCB, which evidenced meaningful indirect and sequential mediation effects. The structural model has a high explanatory power (R2-SCB = 0.692), thus shedding light on the important role of literacy and awareness to promote sustainable financial behaviour. Originality/value: This research adds an integrated model linking the financial technology utilisation, green financial orientation and financial capability towards sustainability driven consumer behaviour which is an underexplored area in emerging economies. The findings provide valuable information for immediate action, for policymakers, educators, and FinTech providers looking to enhance sustainability outcomes using digital financial ecosystems.

ABSTRACT. Choosing a suitable hairstyle is often challenging due to differences in facial structures, personal preferences, and limited access to professional hairstyling advice. Existing hairstyle recommendation applications still face limitations in personalization and realistic visualization, which can make users uncertain about trying suggested styles. To address this gap, this study presents HairMixer, a web-based hairstyle recommendation system that applies deep learning and computer vision to analyze facial structures and generate personalized hairstyle suggestions with a virtual hairstyle overlay feature. The system uses a Convolutional Neural Network (CNN) model to classify face shapes from uploaded facial images. User-provided hair attributes such as texture, thickness, and desired length are applied as filtering constraints to refine recommendations. The system outputs a detected face shape, a list of recommended hairstyles, and a virtual overlay preview that allows users to visualize selected styles on their own image. System performance was evaluated through user testing and professional hairstylist validation. Based on user survey results, the system achieved an overall weighted mean score of 4.52 out of 5, interpreted as "Very Satisfied". Hairstylist evaluation showed a 76.3% agreement rate on generated recommendations, with a Fleiss Kappa value of 0.62, indicating substantial agreement with professional judgment.

ABSTRACT. Pharmacy prescription verification is often managed using first in, first out (FIFO) processing. FIFO is simple to operate, but it can delay prescriptions whose value depends on timely verification when the queue contains many routine orders. We study pharmacy verification as an online scheduling problem in which each prescription is prioritized using two sources of information available at order time: medication urgency and patient deterioration risk. The medication signal is a rule based criticality score that maps each prescription to a 1 to 5 urgency scale. The patient signal is a Platt calibrated classifier that estimates the probability of ICU transfer within 24 hours from order time clinical features, including vitals, laboratory values, and care unit. The patient model excludes medication names, so its score reflects patient state rather than medication identity. Across LightGBM, XGBoost, Random Forest, and Logistic Regression, AUROC ranges from 0.66 to 0.73, suggesting that the patient risk signal is not tied to a single model family. We compare FIFO with three priority policies: medication score only, patient risk only, and a combined policy. Evaluation is performed in a discrete event, single server simulator on MIMIC IV demo prescriptions across five random seeds at 40% utilization. A key finding is that 59.6% of prescriptions for patients who transfer to the ICU within 24 hours are routine by medication criticality. These Type B prescriptions cannot be identified from medication urgency alone. The combined policy reduces mean wait time relative to FIFO by 82±15 minutes for urgent medication prescriptions associated with deterioration and by 75 ± 15 minutes for routine medication prescriptions associated with deterioration. It improves on either single signal policy by 20 to 25 minutes in both subgroups while also reducing wait time for non deteriorating prescriptions. The two signals contribute different scheduling information: medica- tion urgency prioritizes drugs whose clinical benefit depends on timely verification, while patient risk identifies unstable patients whose medication orders otherwise appear routine. Code, data preparation scripts, and the analysis pipeline are released for reproducibility.

ABSTRACT. This study introduced an automated edge-computing system to evaluate sweet corn maturity, addressing a vital need for precise harvest timing to maximize crop value in the Philippines. Operating on a Raspberry Pi 4 with an external camera, the standalone device utilized a robust two-stage deep learning pipeline. It employed YOLOv7 for real-time object detection alongside a ResNet-50 convolutional neural network to accurately classify the corn as mature, immature, or over-mature. Trained on a custom dataset of 2,400 locally sourced images, the system demonstrated exceptional reliability in a controlled environment. It achieved a flawless 100% detection rate and an 88.89% multi class classification accuracy, surpassing the 85% target. Ultimately, this framework provided farmers a highly practical solution to standardize crop quality and optimize their harvest schedules.