ABSTRACT. The rapid pace of digitalization and the emergence of Generative Artificial Intelligence (GenAI) in higher education have shifted the responsibility for cognitive monitoring entirely to students. This context can trigger the Internal Academic Saboteur (IAS), a dysfunctional self-protective pattern that predisposes individuals to cognitive offloading and an illusion of competence. This study evaluates the effectiveness of the Adaptive Self-Regulated Monitoring Model (ASRMM), implemented through a weekly digital journal kept over a 10-week period. The study employed a quasi-experimental design with repeated measures on a sample of N = 159 psychology students (M_age = 20.25 years; SD=1.52), divided into an experimental group (n = 79) and a control group (n = 80). The results partially refuted the initial hypothesis, revealing an unexpected positive correlation between IAS and pre-test performance (r = .38, p < .001), which can be explained by mechanisms of maladaptive perfectionism and strategic pessimism. However, the mixed ANOVA analysis revealed significantly greater academic progress in the post-test phase for the experimental group (F = 60.11, ηp2 = .277). Furthermore, hierarchical regression analysis confirmed a robust moderating effect (ΔR2 = .078, β = .286, p < .001), demonstrating that high adherence to the ASRMM log mitigates and neutralizes the negative impact of IAS on academic progress. In conclusion, the weekly log ceases to be a passive tool and becomes a form of digital scaffolding that prevents the transition from latent vulnerability to defensive avoidance strategies, thereby facilitating deep learning.

ABSTRACT. The digitalization of quality management has become a critical priority for knowledge-based organizations operating in environments characterized by rapid change, high complexity, and continuously expanding data volumes. This paper examines the key challenges faced by such organizations in their efforts to digitally transform quality management systems. Major obstacles include resistance to organizational change, fragmentation of information systems, insufficient data standardization and data quality, as well as gaps in digital and artificial intelligence (AI) competencies among employees. Against this background, the study highlights the role of artificial intelligence as an essential enabler for enhancing quality management practices. AI-based solutions support automated data processing, advanced decision-making, predictive identification of nonconformities, and systematic performance improvement. The paper discusses both managerial and technical approaches to addressing these challenges, including AI-oriented training and reskilling initiatives, data governance frameworks designed to support intelligent algorithms, modular IT architectures, and agile methodologies for the implementation of AI-driven solutions. Furthermore, the analysis emphasizes that fostering an organizational culture centered on innovation, continuous learning, and the responsible use of artificial intelligence is a fundamental prerequisite for the sustainable integration of digital technologies into quality management processes.

ABSTRACT. Resource allocation is one of the most important activities that appears in any organization, especially in transportation systems. Railway transport operators must solve complex problems such as ticket pricing adjustments. Due to the complexity of ticket pricing problem, the traditional solving of this problem involved a fixed per-kilometer price depending on season. Artificial intelligence and machine learning techniques can bring a strong contribution for dynamic ticket pricing in railway transport system. This paper presents a regression-based machine learning prototype that can support optimal ticket price prediction, based on historical information given by the user. The proposed software requires two CSV files including historical data and the data on which the price prediction must be realized, then trains five regression models: Linear Regression, Ridge Regression, Lasso Regression, Random Forest Regression and Support Vector Regression. Model performance is evaluated using standard regression metrics, allowing users to identify the most suitable predictive method for their scenario. The experimental analysis is conducted on a synthetic dataset generated through Monte Carlo simulation, serving as a preliminary proof‑of‑concept in the absence of publicly available pricing data. While the results indicate strong predictive accuracy across all models, the exclusive use of synthetic data represents a key limitation, and real‑world validation remains necessary. The prototype is implemented in Python and can be extended to operational datasets in future work.

ABSTRACT. Medical information systems now carry a twofold promise: better patient care and more precise data. This paper compares the main architectural archetypes of these systems, the integrated monolith, the interoperability-oriented system and the modular one, illustrated with three solutions present on the Romanian market. It does not rank them. The aim is structural, not commercial. The central argument is that data precision is a product, not a sum. Two factors multiply, infrastructure and organisation. Infrastructure sets a ceiling on quality, through interoperability, imaging integration, resilience and structuring. The organisational layer decides how much of that ceiling is reached, that is who enters the data, how and how often. When either factor is weak, the result collapses. Drawing on recent official assessments of the Romanian health system and on the international literature, the paper argues that the dominant constraint is not the type of system. It is the shortage of staff and the absence of a clear assignment of responsibility. The conclusion belongs to quality management. Better data do not come from the newest system. They come from the best-integrated one, used under firm governance of data entry, organised as a measurable plan-do-check-act loop.

ABSTRACT. This paper proposes a hybrid AI-driven conceptual framework that integrates predictive quality analytics with Failure Mode and Effects Analysis (FMEA) to support zero-defect strategies in automotive supply chains. The proposed model links historical supplier performance data, process parameters, defect records, and inspection outcomes to identify pre-failure conditions and emerging risk patterns before nonconformities occur. By incorporating predictive risk indicators into the FMEA structure, the framework enables a dynamic reassessment of occurrence and detection priorities, moving beyond the limitations of static and retrospective risk evaluation. The model supports data-driven prioritization of preventive actions, refinement of control plans, optimization of supplier audit resources, and adaptive incoming inspection strategies. The paper contributes to the development of intelligent decision-support mechanisms for supplier quality management by demonstrating how predictive analytics can enhance traditional risk assessment tools. The proposed framework provides a structured basis for proactive failure prevention, reduced quality costs, and improved resilience in automotive supply chains.

ABSTRACT. Artificial intelligence (AI) architectures are automated systems that, based on a collection of input data, produce predictions, decisions or recommendations as output, which have a potential impact on physical or digital environments. Their variety in terms of autonomy and adaptability generates important challenges in relation to their technological and ethical governance. This article aims to analyze the fundamental principles and key requirements that underlie an effective and responsible governance of AI, focusing on the integration of international standards ISO/IEC and IEEE. We distinguish between these two: while ISO/IEC standards provide the technical framework for system reliability, the IEEE 7000 series offers a value-based engineering approach. This distinction is vital for military AI, where technical performance must be guided by clear ethical boundaries. The study provides a comparative-normative analysis, highlights the gaps existing within the taxonomies of the defense sector – in particular OSRA/EDA – regarding the concept of “trusted AI”, and differentiates the limits found in relation to more advanced developments in the civilian sector. Seven main requirements are also identified and discussed: human oversight, technical robustness, data governance, transparency, fairness and non-discrimination, social and environmental sustainability, and accountability. The present paper highlights the importance of a clear distribution of responsibilities between the actors involved in the AI life cycle, namely: developers, implementers, end-users and society as a whole. Thus, by proposing a cross-sectoral ethical governance framework, the article supports efforts to harmonize technical standards with fundamental rights, arguing that trustworthy AI cannot be achieved solely through compliance, but requires the proactive integration of ethical considerations at all stages of design, implementation and monitoring, especially in high-risk applications.

ABSTRACT. The rapid advancement of digital technologies has transformed recruitment practices, with Artificial Intelligence (AI) playing a crucial role in modern talent acquisition. Traditional methods are often time-consuming and costly, prompting organizations to adopt AI-enabled digital platforms. Social media platforms such as LinkedIn, Facebook, and Twitter, integrated with AI, enable automated screening, candidate matching, and predictive analytics. This study evaluates the effectiveness of AI-enabled social media recruitment using a quantitative methodology and survey data analysis. The results indicate improved recruitment efficiency, reduced hiring time, expanded candidate reach, and enhanced candidate-job matching. The study emphasizes the importance of adopting AI-driven recruitment systems.

ABSTRACT. This study investigates how artificial intelligence, particularly conversational agents, shapes interpersonal communication in the university environment. Drawing on responses from 120 students at Politehnica University of Bucharest, the research examines stylistic and content-related changes in AI-mediated communication, users' perceptions of authenticity and empathy, and AI's role in facilitating or inhibiting interpersonal exchange. The findings show high familiarity with AI among students, moderate and selective use of tools such as ChatGPT, and widespread concern about the social implications of AI. The paper proposes recommendations for responsible conversational system design that prioritizes transparency, user autonomy, and the integrity of human communication.

ABSTRACT. Conventional e-commerce architectures maintain Enterprise Resource Planning and Customer Relationship Management systems as isolated data environments, synchronizing them through scheduled batch processes that introduce decisional latency and limit the scope of predictive analytics. This paper presents the operational instantiation of a multi-layer intelligent Enterprise Resource Planning and Customer Relationship Management architecture within an electronics retail environment, extending it through the incorporation of the e-commerce platform as a third data source at the ingestion level. Behavioral signals generated at the platform layer, session activity, browsing patterns and cart abandonment events, are absent from both ERP transaction records and CRM interaction logs, yet directly condition demand patterns and customer disengagement. Three operational scenarios are examined: seasonal demand forecasting through Long Short-Term Memory based multivariate modeling, silent churn detection through ensemble classification on a unified cross-source feature set and multi-channel overselling prevention through Robotic Process Automation driven event-level stock synchronization. In each scenario, the failure mode of the baseline configuration is shown to be structurally unavoidable without the proposed integration layer. The results establish the Semantic Integration Layer as the architectural precondition for the capabilities demonstrated and identify empirical validation in a live deployment as the primary direction for future research.

ABSTRACT. This paper presents a comparative study of low-frequency magnetic shielding using spherical and cylindrical shields made of copper and steel. Finite element simulations were performed using Comsol Multiphysics finite element software considering shield thicknesses of 1 mm, 2 mm, and 3 mm in the frequency range of 50 Hz–10 kHz and compared with Kaden’s analytical model. The shielding efficiency was evaluated in terms of magnetic field attenuation, while the distribution of induced eddy currents was used to analyze the shielding mechanism. The results show that spherical shields provide higher shielding efficiency than cylindrical shields for all investigated configurations. In both geometries, shielding performance increases with frequency due to the growing contribution of eddy currents. A good agreement between numerical and analytical results was obtained for most cases. Larger deviations were observed for steel shields at 10 and 10 kHz, particularly for thicknesses of 2 mm and 3 mm, where strong skin-effect conditions lead to very high attenuation levels and increased sensitivity to numerical and analytical assumptions.

ABSTRACT. In this paper a review of the main techniques applied for description and modeling the behavior of shielding materials in near and far fields is done. A comparative analysis is developed regarding the obtained shielding effectiveness of different conventional and non-conventional materials. Then, the processes and power losses in electromagnetic interference shielding materials are analyzed, to develop appropriate criteria for their selection and optimal design. A comprehensive synthesis is made of functional and multifunctional shielding materials, in the form of metals, carbon, ceramics, cement-based materials and polymers that exhibit high shielding efficiency. This review shows the current challenges and trends regarding the passive shielding techniques developed to improve both shielding performance and power transfer efficiency.

ABSTRACT. Power to analog perturbation is a challenge in mixed signal PCBs, where high dv/dt switching signals run in traces near measurement paths. Based on my previous passive measurement study presented at ISSCS 2025, this work extends the analysis through detailed extraction and simulation of parasitic elements. Four PCB variants with different routing strategies are characterized by computing trace resistance, inductance, mutual capacitance, and trace to ground capacitance from layout geometry; these parameters are used to build SPICE aggressor‑victim models. Simulations reproduce experimentally observed behavior. The results demonstrate that PCB geometry dominates the coupling mechanism and confirm that the proposed simulation approach reliably predicts power to analog interference in mixed‑signal systems.

ABSTRACT. This paper presents the design and experimental validation of frequency-scalable square open-loop resonator (SOLR) microstrip bandpass filters operating at 2.4 GHz, 3.5 GHz, and 5.8 GHz. The study extends previous single-frequency investigations by demonstrating the applicability of the proposed design approach to multiple frequency bands of interest in modern wireless systems. The filters are designed using different dielectric substrates and analyzed through three-dimensional electromagnetic simulations. Fabricated prototypes are experimentally validated by frequency-domain measurements to evaluate the filter performance as a function of operating frequency and substrate properties. The experimental results show good agreement with the simulated responses and highlight the multi-band behavior of the SOLR filters. The comparative analysis confirms that the proposed structures enable effective frequency scalability while maintaining suitable performance for wireless and electromagnetic compatibility applications in the sub-6 GHz frequency range.

ABSTRACT. This study proposes a cold plasma cryptography model for secure communication scenarios. Its core innovation is that key material is extracted directly from physically observable parameters of low-temperature plasma discharge. The supporting experiment adopts a hollow cathode structure, with helium, hydrogen, and deuterium used as working gases, a gas flow rate of 300 sccm, and an input power of 250 W. A single Langmuir probe and optical emission spectroscopy are applied to characterize the discharge and verify the properties of the cold plasma. This study serially compresses four categories of parameters of time-resolved plasma probes (including electron density, ion density, Debye length, and other related metrics) and acquisition duration. This compressed data is processed by the SHA-256 hash function to generate a seed, which serves as the input source for the AES-256-GCM and RSA-2048 encryption primitives. The full scheme is implemented in Python. Pairwise Hamming distance analysis is conducted to verify the key’s reproducibility, uniqueness, and inter-source separation degree. The results confirm that this coupled scheme can guarantee communication security and is feasible for real-world implementation.

ABSTRACT. Optical fibers have evolved from light-guiding structures into versatile sensing media. The fundamental aspects of optical fibers are reviewed, including their structure, light propagation, and classification, together with key transmission characteristics such as attenuation, scattering, and dispersion. Attention is given to distributed optical fiber sensing, in which Rayleigh, Brillouin and Raman scattering mechanisms enable the measurement of physical parameters such as temperature, strain, and dynamic perturbations along the fiber. The paper further emphasizes the relationship between propagation phenomena and sensing performance, providing a structured perspective that evolves from optical fiber fundamentals to sensing technologies.

ABSTRACT. This paper proposes a dynamic wireless network solution based on NRF24L01+ radio communication modules and Arduino microcontrollers, intended for collecting data from digital and analog sensors, in distributed environments. The network does not subject to a fixed topology, but allows for automatic node discovery and efficient routing of data packets, avoiding congestion of the single radio communication channel. Elements of system architecture, the protocols used, as well as the tests carried out in a controlled environment are presented.

ABSTRACT. This paper presents a real time human activity classification system using three dimensional point clouds produced by millimeter wave radar and a PointNet based deep learning model. A Texas Instruments IWR6843ISK radar operating between 60 and 64 GHz was used to collect data from 10 subjects, including 5 female and 5 male participants, performing five distinct gestures: waist bending, arm flapping, arm swinging, forward bending, and lateral bending. Data were recorded through Robot Operating System 2, converted into sliding window segments of 500 points with a stride of 250 points, quality filtered, and normalized before being fed to a shared multilayer perceptron and global max pooling classifier. The final curated dataset contains 4,155 segments split into 3,324 training samples and 831 test samples. The model achieves 98 percent test accuracy and a macro F1 score of zero point ninety seven. In addition to a contextual comparison with literature baselines, an in house RadHAR reproduction on the same dataset reached 93 percent accuracy, confirming the advantage of direct raw point processing for this dataset. Real time inference runs at approximately 35 frames per second on Jetson Nano and approximately 25 frames per second on Raspberry Pi 4, demonstrating suitability for embedded and contactless monitoring in smart living, rehabilitation, and human machine interaction scenarios.

ABSTRACT. In contrast to relational database management systems, which require numerous joins and intricate queries to handle complex interconnected data, graph databases excel precisely due to their storage approach which involves keeping the data as nodes and edges, effectively representing entities and the relationships that connect them. However, querying graph databases can be perceived as a cumbersome process, especially by the non-technical audience, since it requires not only an understanding of data structures but also the technical specifications of the used protocol, usually SPARQL. To overcome this challenge, Question Answering over Linked Data systems provides users with the ability to query datastores using natural language. The current work presents a new architecture of the onIQ system, which, in addition to enhancing the syntactic parsing mechanism brings users the ability to easily enrich their statistical model with new named entities. The results of evaluating the system against the QALD-8 test dataset are promising, showing the highest precision parameter value among the analyzed systems.

ABSTRACT. Domain-Driven Design (DDD) structures complex business software around the domain model, making it particularly effective in finance, healthcare, and e-commerce. Java's strong typing and mature ecosystem enable robust DDD implementations. This paper examines how compile-time and runtime annotation processing complement DDD: compile-time processors enforce structure and generate boilerplate, while runtime validation (e.g., Bean Validation) enforces business rules without polluting domain classes. We present patterns, code snippets, and trade-offs for eventless DDD aligned with modular monoliths and microservices.

ABSTRACT. Cloud storage remains concentrated among a handful of infrastructure providers, leaving the confidentiality and governance of user data dependent on operators that users must trust not to read, lose, or restrict access to their files. Decentralized storage networks remove the single operator but treat stored content as opaque, while the Solid Protocol restores user-level access control yet binds storage to a single Pod provider and leaves content unprotected below the access-control layer. This paper presents a decentralized storage platform that unifies these directions by combining the Solid Protocol, the InterPlanetary File System (IPFS), and blockchain-based coordination under a client-side encryption model. Files are encrypted on the user's device before reaching any external component, so that no infrastructure operator observes content in the clear; a two-layer key scheme enables file sharing through elliptic-curve key agreement without exposing encryption keys to the platform; storage obligations between consumers and providers are enforced through smart contracts that release payment only after cryptographic verification of file availability; and provider identity is attested through verifiable credentials anchored on-chain. A working prototype was implemented and evaluated; off-chain operations met their performance targets, with on-chain confirmation latency bounded by the underlying test network rather than the architecture, and security tests confirmed correct rejection of unauthorized requests, expired signatures, and invalid credentials.

ABSTRACT. This paper addresses the optimization of an oil-water separation process that is an integral part of an onshore crude oil extraction, collecting, and treatment plant. The process receives three-phase emulsion streams from primary well separators, combines them in a collector tank, and routes the mixture through a heating separator before delivering treated crude oil to distribution clients. A simplified mathematical model is developed that captures the key process variables. An economic objective function is formulated to maximize the operational profit, subject to a nonlinear quality constraint limiting residual water content. Three complementary optimization strategies are applied and compared: Random Sampling Monte Carlo, Latin Hypercube Sampling Monte Carlo, and gradient-based optimization via MATLAB's fmincon with an interior-point algorithm. The optimization method converges to the optimal profit, representing an improvement over the conservative baseline operating point. The results support a hybrid strategy for industrial deployment: Latin Hypercube Sampling Monte Carlo for global exploration of the feasible space, followed by fmincon for precise local refinement. The analysis also reveals that multiple parameter combinations generate the same maximum profit, highlighting practical flexibility in operating point selection for the separation plant.

ABSTRACT. Supply chain networks exhibit highly coupled dynamic behavior resulting from uncertain demand patterns, delayed information exchange, and decentralized operational decisions. These characteristics often lead to instability phenomena, most prominently the bullwhip effect, where small variations in customer demand become significantly amplified as orders propagate upstream through the network. This study introduces a geometric nonlinear control framework aimed at stabilizing multi-echelon supply chain systems operating under stochastic demand disturbances and structural delays. The inventory dynamics are reformulated within a nonlinear control-affine structure, allowing the use of structural controllability analysis together with Lyapunov-based stability verification. A feedback control law that combines integral regulation with a nonlinear damping component is proposed in order to attenuate oscillatory inventory behavior and limit variance propagation across supply chain stages. To assess the effectiveness of the proposed approach, a synthetic dataset is generated using an autoregressive demand model enriched with seasonal patterns and disruption events. A comprehensive Monte Carlo simulation framework is employed to evaluate system performance across multiple uncertainty scenarios, including demand shocks, transportation delays, and volatility changes. The numerical results reveal substantial improvements in dynamic performance. In particular, the proposed controller reduces the bullwhip amplification ratio by approximately 68%, shortens settling time by more than 60%, and significantly enhances robustness against delay-induced instability when compared with conventional linear feedback strategies. The findings demonstrate that nonlinear geometric control concepts can provide a rigorous analytical foundation for stabilizing complex supply chain networks and improving operational resilience in automated logistics environments.

ABSTRACT. Growing structural complexity within Multi-access Edge Computing topologies challenges the rollout of fifth-generation advanced networks and the shift toward sixth-generation ecosystems. These networks demand ultra-low latency, decentralized execution, and autonomous continuity. Standard orchestration policies remain reactive, resolving faults only after they occur, with little capability to foresee systemic degradation in hostile settings. This paper introduces the Autonomous Resilience Hyperstructure Omega, a distributed cognitive architecture built to fortify fifth-generation and sixth-generation Multi-access Edge Computing deployments. The proposed framework fuses multi-agent reinforcement learning, operational entropy-based control loops, semantic service migration, and predictive self-reconfiguration. We model the infrastructure as a dynamic graph of intelligent edge nodes defined by heterogeneous computational capacity, communication latency, energy profiles, and instability indices. To mathematically measure resilience dynamics, we introduce two metrics: the Adaptive Cognitive Resilience Potential and the Temporal Resilience Drift. These indicators evaluate adaptive capacity and track systemic degradation over time. Empirical tests via simulation demonstrate increased architectural stability, reduced operational entropy, and self-stabilizing behaviors, proving that the Autonomous Resilience Hyperstructure Omega successfully shifts orchestration toward a predictive, autonomous, and anti-fragile operational model customized for artificial intelligence-native sixth-generation infrastructures.

ABSTRACT. Lead-cooled fast reactors (LFRs) are considered one of the most promising Generation IV nuclear technologies due to their potential to provide enhanced inherent safety, improved thermal performance, and more efficient utilization of nuclear fuel resources. Within these reactor systems, steam generators represent critical components, as they enable heat transfer between the secondary and tertiary circuits and facilitate the conversion of thermal energy into electrical power. The severe operating conditions encountered in LFRs, including elevated temperatures, high heat-transfer rates, and aggressive corrosive environments, impose stringent requirements on both the materials and structural designs of steam generators. In conventional CANDU (CANadian Deuterium Uranium) reactors moderated with heavy water, steam generators operate under well-established conditions typical of pressurized water systems. By contrast, steam generators designed for liquid lead-cooled fast reactors must withstand significantly harsher environments generated by the liquid metal coolant and the demanding thermal regime. This paper presents an overview of the main steam generator concepts proposed for Generation IV reactors, with particular emphasis on helical coil and bayonet configurations. The advantages, limitations, and operational characteristics of each design are examined. Furthermore, a comparative analysis is carried out between the steam generator technologies currently employed in CANDU systems and the advanced concepts developed for LFR applications. The study also highlights future research directions and development opportunities, underlining the importance of technological innovation in improving the safety, reliability, and long-term durability of steam generators intended for next-generation nuclear power systems.

ABSTRACT. Continuous personnel tracking that survives the passage between the open sky and the interior of a building remains an open practical problem: satellite positioning degrades sharply indoors, while purely inertial estimation drifts without bound. This paper presents a hybrid localization platform for fieldpersonnel tracking that operates seamlessly across the indoor– outdoor boundary and, equally important, across the online– offline boundary imposed by intermittent connectivity. Outdoors, the position is obtained from the fused satellite provider; indoors, a two-dimensional Kalman filter combines a pedestrian deadreckoning prediction with three independent radio corrections— a BLE weighted centroid, a WiFi k-NN fingerprint, and a reverse BLE anchor—under an explicit confidence hierarchy. The regime switch is governed by a geometric geofence around the building. A local-first synchronization layer guarantees that no measurement is lost when the network is unavailable, buffering records on the device and reconciling them in chronological order on reconnection. The system was evaluated on a real Android terminal in an instrumented building. Indoor accuracy reached a mean error of 1.8m under full fusion, the indoor–outdoor transition succeeded in 9 of 10 crossings, and offline buffering recovered 100% of captured records across outages of up to two hours, while the backend sustained the target of 100 concurrent clients with a 95th-percentile latency below 15 ms. Security and resilience mechanisms— jamming and spoofing watchdogs, token-based authentication, and role-based access control—are integrated by design rather than added as an afterthought.

ABSTRACT. Wastewater treatment plants have the role of ensuring environmental protection and public health by eliminating or reducing pollutants from wastewater. Modeling parameters, such as dissolved oxygen (DO), nitrogen (in various forms, such as ammonia, nitrite, and nitrates), play an important role in optimizing the performance of wastewater treatment plants (WWTPs). This paper presents certain optimization strategies through modeling for oxygen and nitrogen compounds, with a focus on predictive modeling and process control. The method for optimizing automatic control in biological wastewater treatment processes uses: identification of the main parameters subject to automatic control; presentation of the way these methods are implemented in the General Process Simulator – Extended (GPS-X) simulation environment; comparison of the performance obtained through different control methods; formulation of recommendations regarding the optimization and extension of the application of automatic control in real operation. The methodology used in this paper presents the study of dynamic models used in the Activated Sludge Model (ASM1) simulations, the creation of automatic control scenarios in the GPS-X environment and the comparative interpretation of the results.

ABSTRACT. Recent advancements in laboratory automation and process integration have opened opportunities for enhancing the efficiency, reproducibility, and scalability of lithium isotope separation methods. This study presents the development of an integrated automated platform that combines lithium salt processing, controlled crystallization, and ICP-MS quantitative analysis. The system features a cascade electrochemical setup connected to a double-wall crystallization vessel cooled to 4 °C, along with a DOBOT MG400 robotic arm for automated transfer operations. Experimental validation demonstrated stable cooling, reliable robotic handling, and seamless integration of electrochemical processing, crystallization, and analysis stages. Overall, this platform offers a scalable solution for laboratory automation and an open road for AI-driven optimization of lithium isotope separation processes.

ABSTRACT. This paper investigates how fifth-generation mobile networks integrated with Multi-Access Edge Computing sustain operational resilience during critical communication failures. By partitioning the network into distinct radio access, transport backhaul, core, and edge processing layers, our modeling approach maps end-to-end latency to isolate specific processing bottlenecks. We use Packet Delivery Ratio as the core reliability metric to fulfill Third Generation Partnership Project Ultra-Reliable Low Latency Communications requirements. To evaluate system performance under stress, we develop two distinct analytical frameworks: the Mission Continuity Score for tracking temporal service persistence during load spikes, and the Resilience Elasticity Index for measuring degradation paths and self-healing speed. Our methodology relies on extensive Monte Carlo simulations to analyze architectural stability under volatile traffic profiles and congestion. This iterative stochastic approach generates probability density functions that reveal tail latency behaviors during cascading network outages. The resulting statistical distributions prove that positioning user-plane functions at the network edge curtails transport delays and stabilizes system predictability. These findings define the structural limits of resilient, edge-based computing frameworks designed for time-critical operations.

ABSTRACT. Unmanned aerial vehicles (UAVs) are increasingly used in precision agriculture to monitor crop health and detect diseases using high-resolution imagery. However, continual visual data transmission from UAVs results in significant communication overhead and raises energy consumption, particularly in rural areas with limited bandwidth. A reinforcement learning-based adaptive communication system for UAV-assisted agricultural IoT monitoring is proposed in this paper. The proposed approach dynamically adjusts transmission parameters such as image compression level, frame transmission rate, and edge–cloud processing decisions based on real-time network conditions and UAV system status. To balance monitoring accuracy, energy consumption, and communication latency, a Deep Q-Network (DQN) agent learns the best communication strategies. The suggested approach dramatically lowers bandwidth and energy consumption while maintaining good crop disease detection accuracy, according to experimental evaluation using the PlantVillage dataset and a simulated UAV communication environment. The findings demonstrate how intelligent network adaptation can enhance the effectiveness and scalability of agricultural monitoring systems provided by unmanned aerial vehicles.

ABSTRACT. Energy awareness is increasingly important for shaping user behavior in energy communities. The Common European Reference Framework (CERF) targets energy-saving applications for monitoring and managing end-user energy consumption. Within this framework, the present paper investigates notification-based mechanisms for increasing user energy awareness relative to an energy community. The proposed approach combines classical signal filtering with NN-assisted event de-spamming and has been evaluated within the Eclipse Digital project for the Romanian pilot site. The main objective is to reduce excessive daily notifications in order to avoid behavioral saturation while maintaining user engagement in voluntary energy-consumption adaptation. This paper presents a proof-of-concept implementation of an NN-assisted event filtering and notification prioritization mechanism designed to limit redundant alerts generated from smart-meter event streams. Preliminary results indicate that the proposed approach can reduce the number of notifications generated from smart-meter event streams while preserving the operational capability of the monitoring framework.

ABSTRACT. Climate-related hazards increasingly threaten critical infrastructure systems, particularly transport and energy networks that are geographically co-located and functionally interdependent. Effective climate resilience planning requires the integration of heterogeneous datasets, including hazard simulations, infrastructure networks, and environmental indicators. However, these datasets are often distributed across different platforms and formats, limiting their accessibility and usability for infrastructure planners and decision-makers. This paper presents the design and implementation of a web-based geospatial visualization platform developed within the ReCharged project to support climate risk analysis of transport and energy infrastructure systems. The platform integrates multiple geospatial datasets provided by project partners, including climate hazard layers, infrastructure networks, and exposure indicators, into a unified interactive environment. Through a modular architecture and a layer-based mapping interface, users can explore hazard scenarios, visualize infrastructure exposure, and analyse potential impacts across different case study locations. The platform supports multiple European case studies and enables the dynamic visualization of hazard layers such as flood depth, flow velocity, and mudflow extent alongside transport and energy infrastructure elements. By combining geospatial data management with interactive web-based visualization, the system facilitates clearer interpretation of complex risk scenarios and enhances communication between technical experts and decision-makers. The proposed platform demonstrates how digital geospatial technologies can improve accessibility to climate risk information and support more informed infrastructure resilience planning.

ABSTRACT. The energy sector is undergoing an accelerated digitalization process in which continuous monitoring of critical infrastructure has become both a technical and operational requirement. Hydroelectric power plants, as strategic units within the national energy system, contain numerous electrical cabinets whose parameters must be permanently tracked to ensure operational availability and safety. This paper presents an Industrial Internet of Things (IIoT) open-source architecture dedicated to the supervision of such a cabinet. The system integrates a Janitza UMG 104 industrial power analyzer and an XY-MD02 temperature and humidity sensor through a Waveshare ESP32-S3 industrial gateway, with field-level communication carried out via Modbus RTU over an RS-485 bus. Data are aggregated by a Mosquitto MQTT broker running on a Home Assistant Operating System platform hosted on a Raspberry Pi 5, while remote access is provided through a Cloudflare tunnel with HTTPS encryption. The proposed solution offers low cost, high flexibility, and a security level adequate for an industrial application, while eliminating single-vendor dependency. Experimental validation demonstrated stable communication between field devices and the supervisory platform, reliable remote monitoring through encrypted HTTPS connections, and real-time actuator control with low operational latency. The proposed architecture represents a scalable and cost-effective alternative to proprietary SCADA solutions for medium-complexity industrial applications.

ABSTRACT. Imperfections and obstacles on roads can cause real discomfort in driving and can cause road accidents, vehicle damage, and transportation inefficiencies. Traditional methods of road inspection, especially manual ones, require a lot of time and are much more expensive. This research proposes an automated pothole detection system using deep learning architectures, including EfficientNet, NasNet-Mobile, and ShuffleNet implemented in MATLAB. The system uses image datasets captured with a real-time camera under different environmental conditions, running on the Jetson Nano platform for detection and pothole classification. The images were imported into MATLAB, and using different predefined architectures, we identified different training epochs, each with different iterations, resulting in several accuracy scores. Two training and validation classes with asphalt defects and a roadway without imperfections were used. An experimental evaluation of 246 real-world road images demonstrated that EfficientNet attained the highest validation accuracy of 83.78%, succeeded by NASNet-Mobile at 81.08% and ShuffleNet at 79.73%. This highlights that EfficientNet attains higher classification accuracy, NASNet delivers strong feature extraction capabilities, and ShuffleNet enables faster performance suitable for real-time use. The suggested framework can aid intelligent transportation systems and smart road maintenance solutions.

ABSTRACT. Urban environments face increasing challenges related to road safety and infrastructure maintenance, particularly as cities adapt to the growing prevalence of autonomous vehicles. By leveraging high-quality datasets and machine learning techniques, we aim to contribute to safer and more efficient navigation for both autonomous vehicles and drivers, ultimately fostering smarter, safer cities. This paper investigates the various YOLO (You Only Look Once) models, specifically YOLOv5, YOLOv11, and YOLOv26, in detecting road irregularities, such as denivelations, as part of smart city initiatives for autonomous vehicles. The comparative analysis demonstrates that YOLOv11 consistently excels in all versions, making it a reliable choice for real-time assessment of road conditions. However, as the dataset size increased from 244 to 616 and then to 1998 images, a significant decrease in the average accuracy (mAP@50) was observed, falling from approximately 0.94 to 0.67-0.70. This reinforces the idea that the quality and consistency of the annotations have a significant impact on the model performance more than the sheer volume of data. Furthermore, YOLOv26 did not demonstrate significant improvements over YOLOv11, suggesting that the model is sensitive to factors such as motion blur, variable angles, and inconsistent illumination, which are key challenges in detecting road bumps for autonomous vehicles. The results of this study highlight the critical importance of high-quality datasets in developing effective computer vision models aimed at real-time monitoring of road conditions within smart cities, ultimately supporting the safety and effectiveness of autonomous driving systems.

ABSTRACT. Backup on mobile devices is usually performed on a fixed schedule or at the user’s initiative, which leaves the most recent data unprotected at exactly the moments when the device is under threat. This paper presents a mobile backup system that inverts this logic: instead of a timer, backups are driven by security events observed on the device, so that a current snapshot is preserved when the risk of loss is elevated. A set of background Android components monitors signals such as repeated failed authentication, SIM changes, untrusted networks and suspicious application installations, and a trigger engine initiates a backup while suppressing redundant executions. Because a backup exposes a large and sensitive cross-section of personal data, all information is encrypted on the device before transmission, and a searchable index is built locally so that the backup can be queried by metadata without revealing that metadata to the server. The searchable component uses a deterministic HMAC-based construction that favours low clientside cost, and we make explicit the trust model under which its confidentiality guarantees hold. We describe the architecture, the triggering and encryption mechanisms, and the practical limitations encountered in building such a system on Android, which we believe are instructive for similar work.

ABSTRACT. District-heating failures in dense smart-city districts require rapid service restoration while limiting emissions and operational cost. This paper proposes a techno-economic framework for selecting the economically preferred thermal point to serve when several outages occur simultaneously. The case study is based in Bucharest on the SMARTELTER concept, a green mobile service that combines renewable charging, battery containers, a 1 MWt thermal-energy module and an AI-based management platform. The system extracts records from the public Termoenergetica outage page, matches affected thermal points with an urban thermal-demand inventory, filters candidates within the thermalmodule capacity, estimates route duration from the fixed base, schedules three battery containers under a 15% state-of-charge reserve rule and ranks candidates by operational profit. The results show that dispatch decisions depend jointly on demand, distance, battery autonomy and season-specific tariff assumptions. A single emergency service price is used to assess whether temporary outage response remains economically viable while preserving continuity of thermal comfort.

ABSTRACT. Microring resonators are key building blocks in integrated photonic circuits, where device performance is strongly governed by the efficiency of optical coupling between the bus waveguide and the resonant cavity. This paper presents a systematic beam propagation method (BPM)–based comparative analysis of three microring resonator configurations: without grating, with non-uniform grating, and with uniform subwavelength grating in the coupling region. Optical field evolution, confinement characteristics, and resonant circulation behavior are investigated at an operating wavelength of 1550 nm using OptiBPM. Two-dimensional and three-dimensional field intensity distributions are analyzed to examine coupling dynamics and power leakage mechanisms. Quantitative evaluation is performed using normalized peak field intensity and coupling efficiency as performance metrics. The results demonstrate that uniform subwavelength grating-assisted coupling provides smoother effective refractive index transitions, enhanced mode matching, and significantly improved optical field confinement compared to both no-grating and non-uniform grating configurations. Furthermore, the extracted simulation-based field metrics form a structured and physically meaningful dataset that is well suited for data-driven surrogate modeling and AI-assisted optimization of grating-assisted photonic devices. The proposed framework offers practical design insights for integrated photonic applications and supports emerging trends in computational intelligence–enabled photonic system design.

ABSTRACT. Zero Trust Architecture (ZTA) eliminates implicit network trust and requires continuous verification, yet many deployments still rely on static policy logic that reacts slowly to emerging threats. This paper proposes a multi-component anomaly detection approach that leverages an Isolation Forest and a Dense Autoencoder, integrated through a decision-level OR fusion strategy, to enhance continuous verification in ZTA environments. The two detectors are trained exclusively on normal-class traffic and operate on disjoint feature subsets: the Isolation Forest processes network-level features while the Autoencoder captures behavioral session patterns. This complementary design creates a clear division of labor, the Isolation Forest specializes in detecting DoS and Probe anomalies, while the Autoencoder captures R2L and U2R intrusions that manifest in session-level behavior. Decision-level OR fusion preserves both detection signals without requiring score calibration, addressing a demonstrated failure mode of score-level averaging between heterogeneous detectors. Evaluated on the NSL-KDD benchmark, the fused system outperforms both standalone components and the selected supervised baselines (Random Forest and Gradient Boosting) on the distribution-shifted test sets. Bootstrap confidence intervals and McNemar's tests confirm the statistical robustness of the reported improvements. The full pipeline achieves throughput well within real-time network monitoring requirements. These results position decision-level fusion of complementary one-class detectors as a practical and scalable mechanism for anomaly-informed policy decisions within the ZTA continuous verification loop.

ABSTRACT. Insider threats combine legitimate access rights with malicious intent in ways that defeat perimeter-centric defenses. This paper presents a one-class anomaly detection framework integrated with Zero Trust Architecture (ZTA) for the detection and mitigation of insider threats in enterprise environments. The proposed system comprises three sequentially integrated components: (1) an LSTM-based autoencoder trained exclusively on normal user behavior to detect temporal anomalies in daily behavioral sequences without requiring labeled threat data; (2) a Trust Score Engine (TSE) that computes a dynamic, temporally smoothed per-user trust score from the autoencoder reconstruction error; and (3) a ZTA policy layer that maps continuous trust scores to tiered access control decisions. Evaluated on the CERT Insider Threat Dataset r5.2, the framework detects insider threats on average over a week before confirmed malicious activity using only normal user behavior as a training signal, enabling preemptive access restriction. The Trust Score Engine achieves competitive anomaly detection performance without access to labeled threat data at inference time, while the ZTA policy layer concentrates the majority of confirmed malicious activity into a small fraction of monitored and high-risk user-days. Comparison against supervised and unsupervised baselines documents the performance trade-offs between label-available and label-free detection conditions.

ABSTRACT. Critical infrastructures, particularly in the healthcare sector, are increasingly vulnerable due to their dependence on third-party software providers. The Hipocrate incident stands as one of the most severe cyberattacks in the history of the Romanian healthcare system. The compromise of the servers belonging to the Hipocrate Health Information System provider paralyzed the operations of 26 medical units and forced the preventive disconnection of 79 others, affecting approximately 19% of Romania's operational hospitals. The ransom demand of 3.5 BTC (~€157,000) was refused. This paper reconstructs the attack chain using attack trees and the MITRE ATT&CK v14 framework, evaluates post-incident resilience through MTTD, MTTR, and a proposed Longitudinal Resilience Index (LRI), and quantitatively validates — through Agent-Based Modeling simulation (AnyLogic 8.9.8) — the impact of NIS2 compliance: the LRI at 72 hours increases from 0.539 in the real scenario to 0.674 in the post-NIS2 scenario, while the infection peak drops from 94% to 57% of network nodes. The decisive vulnerability: the absence of multi-factor authentication for remote vendor access. The factor that limited the disaster: the institutional transparency of DNSC, with IoCs published within 72 hours.

ABSTRACT. This work analyzes major Software Supply Chain Attack (SSCA) incidents from the period January 2025 – May 2026. The 10 new incidents in the first five months of 2026 were examined along three dimensions: attack vectors, the effectiveness of OSINT (Open Source Intelligence) sources in early detection, and the implications for cyber resilience. The comparison with the 2025 landscape reveals a paradigm shift: self-replicating worm campaigns have given way to coordinated multi-vector operations targeting npm and PyPI simultaneously, with new extensions to DevSecOps tooling and AI/ML (Artificial Intelligence / Machine Learning) ecosystems. OSINT has proven effective for public package registries—with response times in the order of hours — and systematically ineffective for the vendor/installer vector, where the non-detection interval has reached 30 days. Based on these findings, we developed and preliminarily validated an ML framework that uses NLP (Natural Language Processing) to extract entities from unstructured OSINT sources and LSTM (Long Short-Term Memory) to classify packet behavior - with the aim of detecting attacks before a CVE (Common Vulnerabilities and Exposures) is issued.

ABSTRACT. Vibe coding is an emerging approach in which developers and non-technical users describe desired functionalities in natural language, delegating the actual code generation to an artificial intelligence model. This paper explores the concept of vibe coding through a practical experiment, using Visual Studio Code and Anthropic’s Claude model, to build a functional web application and analyze the efficiency of the process. The results are compared with traditional software development methods, highlighting significant reductions in implementation and modification time. At the same time, the study identifies the inherent risks of this paradigm, in particular security vulnerabilities, potential exposure of sensitive data and the difficulty of auditing automatically generated code that the user does not have a deep understanding of. We conclude that vibe coding democratizes access to software development but requires rigorous oversight frameworks to be adopted responsibly.

ABSTRACT. Multi-tenant wired networks based on the DOCSIS 3.1 standard must simultaneously serve tenants with heterogeneous and time-varying service level agreement requirements. Classical static resource partitioning fails under dynamic traffic conditions, leading to either wasted capacity or service level agreement violations. This paper proposes a dynamic resource slicing framework that employs a Genetic Algorithm augmented by a Random Forest traffic predictor for demand-aware warm-start initialization to optimize the joint allocation of downstream bandwidth, transmit power, and Orthogonal Frequency Division Multiple Access minislot assignments across multiple tenants in real time. The proposed framework is evaluated against three baseline algorithms Round Robin , Weighted Static, and MaxMin Fairness over 60 simulation episodes with heterogeneous traffic profiles. Results demonstrate that the proposed GA significantly outperforms the Round Robin algorithm, delivering higher average throughput, lower packet loss, and better overall Quality of Service satisfaction. An impact analysis isolating the ML component demonstrates that the trained Random Forest predictor consistently outperforms a static heuristic seed, leading to improved QoS satisfaction and higher aggregate fitness, thereby directly quantifying the value of data-driven prediction within the overall framework.

ABSTRACT. The transition from Advanced Driver Assistance Systems (ADAS) to fully Autonomous Vehicles (AVs) has generated unprecedented challenges regarding safety validation and certification. Traditional testing methods, relying exclusively on distance-based testing (mileage-based), have proven statistically impractical for demonstrating a reliability superior to that of human drivers. This paper presents a critical analysis of the current state-of-the-art in autonomous vehicle testing methodologies, highlighting the limitations of existing approaches from the perspective of the Safety of the Intended Functionality (SOTIF - ISO 21448). The study evaluates the efficacy of hybrid testing environments, such as X in the Loop (XIL) architectures, and analyzes critical vulnerabilities caused by the Sim to Real gap. As its main contribution, this paper systematizes the research gaps within the specialized literature and proposes future research directions focused on developing a unified performance evaluation framework. This framework integrates dynamic risk metrics and real-time assessment of sensory degradation. Ultimately, this synthesis provides a robust foundation for future adaptive testing standards in autonomous vehicle engineering.

ABSTRACT. This paper presents a cellular automata-based hash function for data integrity applications. The proposed algorithm uses a sponge-inspired structure with a 1600-bit internal state which generates a 256-bit digest. The internal permutation process combines elementary cellular automata rules, lane rotations, lane permutations and nonlinear transformations to ensure diffusion and confusion within the internal state. In addition, the constant injection stage is implemented using a hybrid cellular automaton, replacing fixed predefined round constants with dynamically generated values. The proposed algorithm was implemented in C# programming language using the Visual Studio development environment and experimentally evaluated for data integrity verification. Preliminary results indicate a strong avalanche effect and efficient execution due to parallel operations performed on 64-bit lanes, highlighting the potential of cellular automata in the design of experimental hash functions.

ABSTRACT. The increasing connectivity of automotive systems through Vehicle-to-Everything (V2X) communication and Mobile Ad Hoc Networks (MANETs) has created new vulnerabilities to Distributed Denial of Service (DDoS) attacks, threatening the availability of safety-critical vehicle communications and infrastructure services. This re-search addresses the challenge of protecting MANET-based automotive infrastructure by developing an integrated security architecture combining Network Detection and Response (NDR), Security Information and Event Management (SIEM), and Security Orchestration, Automation, and Response (SOAR) capabilities. Risk analysis was conducted using the NIST Cybersecurity Framework 2.0, mapping security controls across its six core functions. A laboratory proof of concept validated the architecture using CYBERQUEST (SIEM) and NETALERT (NDR) platforms to detect and automatically mitigate a simulated volumetric DDoS attack against a static network node. The integrated detection chain successfully identified abnormal connection volumes, correlated alerts across multiple sources, and executed automated blacklisting responses without human intervention. The results demonstrate that commercially available security platforms can be effectively adapted for MANET environments when properly integrated, providing rapid automated response capabilities aligned with European regulatory requirements including the NIS2 Directive and UNECE Regulation No. 155.

ABSTRACT. Distributed medical sensor networks that are used for continuous healthcare monitoring produce heterogeneous traffic that has strict quality-of-service requirements regarding latency, reliability, and energy efficiency. On the other hand, communication static approaches that are traditionally used cannot efficiently manage dynamic network conditions and priority-sensitive medical data, especially when congestion occurs. This study introduces an innovative, adaptive communication framework designed to manage distributed medical sensor traffic by leveraging traffic-aware scheduling and dynamic resource allocation. The main idea of the proposed approach is to divide medical data into different priority levels, such as routine, alert, and critical traffic, and to change transmission parameters dynamically based on network load and packet priority. The primary objective is to enhance end-to-end communication reliability while minimizing transmission delay and energy consumption. To validate the proposed method, simulations were conducted under realistic wireless medical network scenarios. The results of the experiments show that the adaptive strategy proposed by the authors can substantially decrease latency and packet loss for critical medical traffic compared to static communication schemes while making sure that energy use stays efficient. The current results confirm that the performance and reliability of future distributed healthcare monitoring systems can be greatly enhanced by using adaptive traffic management techniques.

ABSTRACT. In resource-constrained economies across Asia and Africa, where approximately 700 million individuals subsist below the international poverty threshold of $2.15 per day (World Bank, 2025), government-sponsored public distribution systems (PDS) for subsidized rations to underprivileged populations are perpetually undermined by pervasive corruption inherent in antiquated manual and paper-based authentication protocols. Such vulnerabilities precipitate egregious data manipulation, fraudulent entitlements, and prodigious fiscal leakages, exacerbating socioeconomic inequities. This manuscript delineates an advanced WEB Server and IoT-Enabled RFID ATM Ration Vending System, meticulously engineered to engender unparalleled transparency, operational efficacy, and scalability not only for developing nations but also for affluent countries pursuing welfare imperatives for economically marginalized cohorts. The architecture pivots on an Xtensa LX7 dual-core microcontroller orchestrating RFID readers, relay modules, and auditory feedback mechanisms, conjoined with dual proprietary vending paradigms: a precision spring-actuated dispenser for pre-packaged rations spanning 0.5 to 5 kg, and a solenoid-governed hopper apparatus for volumetric grain disbursement from 1 to 5 kg. Family-specific RFID credentials interface with a robust SQL and python backend via a Tkinter-constructed Python graphical user interface (GUI), facilitating wireless command propagation through a dedicated webserver for instantaneous quota modulation and immutable transaction archival. This paradigm obviates human intermediaries conventionally requisitioned for shop oversight, ration enumeration, fiscal computation, and archival stewardship, thereby precipitating a 60–70% diminution in manpower expenditures. Empirical validation attests to 99.8% dispensing fidelity and unassailable data provenance, furnishing a paradigmatic blueprint for global PDS metamorphosis, fortifying governance accountability, and ameliorating poverty mitigation stratagems.

ABSTRACT. As a renewable and sustainable energy source, solar panels have seen a big increase in new customers. Adoption of PV (Photovoltaic) technology is at an all time high. Silicon cells are the majority of the PV cell market because they are both financially viable and reliable. There are problems associated with PV cells that can affect efficiency, such as high operating temperature. For the majority of manufacturers and installers, the potential efficiency losses are substantial and occur at a rate of approximately 0.4 - 0.5% per 1°C increase in operating temperature. This issue becomes even more problematic with low-concentration photovoltaic (LCPV) systems due to the concentrated solar energy producing excessive heat on the solar cells. The aim of this project is to develop a system to actively cool an LCPV solar cell with the help of the Internet of Things (IoT). An IoT-based cooling system is being developed that will monitor the solar cell and other environmental conditions such as the surrounding air temperature and solar irradiance to help protect the solar cell from the effects of heat. Sensors connected to microcontrollers will collect data on the variables listed above and transmit that information, via the Internet of Things, to a cloud service. Using the temperature data collected, an active cooling method, i.e. air or liquid, will be initiated to ensure the optimal operating conditions of the solar cell. Compared to conventional cooling methods, the proposed system provides enhanced temperature control, real-time monitoring and automated control, resulting in higher efficiency and power output from the solar cell and an extended operational life of the solar cell.

ABSTRACT. Water rationing has a significant effect on the lives of residents in Zamboanga City, as irregular supply times cause missed collection times, overflows, and constant human observation. To address these issues, this study was created to develop an innovative IoT-based automatic water refilling system, which is HydroLink. This is a smart device meant to optimize water management by utilizing automation and remote observation. The proposed system utilizes an ESP32-S3 microcontroller, which includes an HC-SR04 ultrasonic sensor to measure water levels precisely and a YF-S403 water flow sensor to control a solenoid valve. A critical calibration offset of 7.5 cm was used to overcome the dead zone, allowing for a high accuracy of 97-100% to be achieved. Additionally, a web interface, combined with the OLED display, has enabled the user to access analytics, refill history, and system status through a Firebase database. Through the research design, the HydroLink was tested under a range of pressure conditions, validating the prompt response time of the system, at 1-4 seconds, well within the 5-second window for efficient rationing. This study has proven that the HydroLink is a highly reliable system, accurate to 99.8%, allowing for the efficient elimination of the need for human intervention, preventing overflow, and providing a reliable volumetric water metering solution for the complexities of water rationing in urban areas, utilizing the IoT framework.

ABSTRACT. The decarbonization of marine auxiliary power systems requires efficient renewable-integrated electrical architectures. This work presents the design and hardware implementation of a DC-coupled marine power system based on a regulated common DC bus. Solar photovoltaic and wind energy outputs are conditioned through power electronic interfaces before integration into the DC link. To improve system dynamics, a hybrid storage unit consisting of a 12 V battery and a supercapacitor is connected across the DC bus. The battery provides sustained energy support, while the supercapacitor compensates for short-duration power fluctuations, maintaining DC link voltage stability. A microcontroller-based control scheme generates PWM signals to drive MOSFET switching devices through opto-isolated drivers. Voltage and current sensing provide feedback for duty-cycle control and protection. Experimental evaluation demonstrates stable DC bus regulation, effective renewable integration, and reliable DC load operation under varying input conditions. The proposed DC-coupled configuration simplifies power coordination and is suitable for renewable-assisted marine auxiliary power systems.

ABSTRACT. The growing scale and complexity of Distributed Denial-of-Service (DDoS) attacks continues to pose a significant threat to modern IoT & networked systems by impacting service availability. While deep learning–based intrusion detection systems have showed great promise, their effectiveness is usually constrained by high-dimensional and redundant features of network traffic. To overcome this challenge, the proposed paper introduces an efficient DDoS detection framework based on a deep autoencoder–based dimensionality reduction followed by a neural network classifiers. The proposed method makes use of a deep autoencoder to learn compact and informative latent representations from high-dimensional network traffic flows. The extracted latent features will be the input for MLP and DNN classifiers to perform binary DDoS detection. The proposed framework is validated on the CICDDoS2019 dataset and the performance is measured through standard evaluation metrics such as accuracy, precision, recall and F1-score. Experimental results show that using autoencoder-based dimensionality reduction improves training stability and detection performance. The reduced feature space is beneficial for both classifiers; nevertheless, DNN showed performance superiority over MLP in all evaluation metrics with improved detection accuracy and robustness. Results confirm that the neutralisation of nonlinear features improves the performance of neural net–based DDoS detection with respect to more complex hybrid architectures. Deep representation learning for high-dimensional network traffic data enlightening DDoS solution in IoT networks state-of-the-art model of scalable, effective and reproducible framework.

ABSTRACT. The development towards beyond-5G (B5G) and 6th generation wireless systems requires the fronthaul networks to have ultra-high bandwidth, very low latency, high spectral efficiency and operate in an energy-efficient manner. The traditional Radio-over-Fiber (RoF) systems, Analog RoF (A-RoF) and Digital RoF (D-RoF), have inherent latency-robustness-bandwidth trade-offs. A-RoF is low-latency but acute sensitive to the optical impairments, whereas D-RoF has high robustness at the expense of more fronthaul bandwidth and power consumption. To cope with these drawbacks, in this paper, we propose an innovative Elastic Digital–Analog Radio-over-Fiber (EDA-RoF) concept for next generation wireless access. The proposed architecture includes both analog and digital transmission paths in the same system and presents a performance-oriented Elastic Control Unit (ECU) to enable dynamic swapping of modes on account of instantaneous SNR, traffic load and delay requirements. An analytical model is developed for analog, digital and hybrid signal transmission. These results are extensively validated by MATLAB based simulations under realistic B5G/6G fronthaul simulations for performance metrics including SNR, BER, EVM, end-to-end latency, spectral efficiency and power consumption. The results show that the EDA-RoF can improve SNR (2–4 dB), decrease BER by up to one order magnitude, reduce latency (about 30–45%) than D-RoF, enhance spectral efficiency (20–30%), save power consumption between 25% and 40%, under different resource allocation strategies. The results validate that elastic analog–digital switching can well balance the reliability, efficiency, and scalability and EDA-RoF is a promising candidate for green and high-performance B5G/6G fronthaul networks.

ABSTRACT. Wireless Sensor Networks (WSNs) are well-established, smart systems whose significance is rapidly escalating. Geographic routing offers many advantages in terms of network reliability and performance. However, they tend to fail in high congestion and the presence of energy holes. To address these challenges, a new geographic routing technique has been introduced to enhance network performance. The proposed routing algorithm is built upon the Technique for Order of Preference by Similarity to the Ideal Solution (TOPSIS). TOPSIS will use four criteria to forward data reports to the best-neighbor node without causing excessive congestion. In addition, the proposed algorithm will be embedded with a technique that retard the energy hole formation and can bypass them effectively without expanding them. The Path Deviation (PD) technique will prevent repetitive sending along the same paths. PD technique removes the previous forwarding node from the routing table. The proposed routing algorithm outperforms the benchmark algorithms, delivering notable improvements in packet error rate, reaching 29.3% and 16.4% compared to the FTR and EEGR algorithms, respectively. Moreover, the proposed routing algorithm demonstrates enhanced performance over the benchmark algorithms, achieving gains in packet delivery ratio of 2.6% and 1.16%, respectively. At the same time, it outperforms the network lifetime of the FTR algorithm by 28.4% and the EEGR algorithm by 11.7%. Lastly, it reduces latency by 12.2% compared with FTR and 18.2% compared with EEGR.

ABSTRACT. The integration of Artificial Intelligence (AI) in the field of healthcare has significantly impacted the way medical consultation and diagnosis are done in the past few years. This paper focuses on the engineering and development of an AI-powered Smart Health Assistant capable of conducting medical remote consultation through the comprehension of the patient's symptoms to provide diagnostic consultation. The proposed AI model utilizes Natural Language Processing (NLP) to understand the patient's verbal descriptions, Machine Learning (ML) algorithms to determine the possible diseases, and API integration to retrieve medically authenticated information from sources trusted to provide information smoothly. The proposed model, which represents a new form of telemedicine, offers personalized, instant, and context-aware conversations with a high level of human simulation, different from normal telemedicine services. The study revealed the need to develop accountable healthcare AI models that are ethical in nature to ensure trust and reliability. The model presented consists of modules for device symptom evaluation, disease prediction, and answer generation in a conversational Python API environment. The performance evaluation tests conducted on symptom-disease datasets indicate the accuracy of prediction and user experience at a high level. The research outputs unveiled that AI-driven conversational agents could be instrumental in the timely identification of diseases, thus substantially enhancing healthcare accessibility, which is a basis for the upcoming telemedicine innovations employing AI.

ABSTRACT. Text summarizing is the process of reducing the length of an input document without changing its main ideas. Extractive summarization is choosing the most significant sentences from the input text. In contrast, query-based multi-document summarization is a complex task that faces more challenges than single-document summarization. It involves overlapping information between sentences from various documents and selecting the most relevant and diverse sentences from a collection of documents while minimizing redundancy and maintaining relevance to a specific query. This paper suggests integrating Latent Dirichlet Allocation (LDA) and K-means clustering to address these challenges. In this paper, researchers assess the probability distribution of words and their respective subjects at the cluster level to identify the most relevant topics, therefore addressing the drawbacks of conventional topic modeling. This method avoids sentence overlap and effectively detects important sentences. Also, LDA is utilized to extract and understand the underlying topics within the documents, facilitating the identification of content most relevant to the query. K-means clustering is then applied to group sentences based on their topic distributions, allowing for the selection of representative sentences that enhance the diversity of the summary. The Maximal Marginal Relevance (MMR) criterion was utilized to re-rank the picked sentences, balancing query relevance and redundancy reduction. By combining LDA’s ability to model topics with K-means’ capacity for clustering, this approach effectively handles the challenges of relevance, diversity, and redundancy, resulting in a more comprehensive and informative summary. By using CNN/Daily Mail benchmark data, the experiences show this hybrid method is particularly suited for large-scale document collections; there are notable improvements in performance compared to current text summarizing methods. Our strategy when evaluating the created summaries is automatic. shows significant improvement in ROUGE metrics (F-measure, accuracy, and recall), providing a scalable solution for generating high-quality summaries. The introduced system evaluates the CNN/Daily Mail datasets. By experiments, the proposed system obtains 62.8 ROUGE-1, 54.5 ROUGE-2, and 40.0 ROUGE-L scores, an enhancement over previous models. Human evaluation also illustrates that the produced system has higher-quality summaries. This paper introduces the integration of the Topic LDA model, the K-means clustering algorithm, and the MMR technique, leveraging them to generate a robust extractive query-based summarization. Section 2 produces the most challenges that can be faced in the extractive query-based summarization task. Section 3 mentioned most of the previous studies that deal with extractive text summarization, referring to some strategies and approaches to perform this task. Section 4 includes the background and definitions of the keywords mentioned in this paper. Section 5 describes the proposed system introduced by this paper for query-based summarization. Section 6 explains the results and the metrics used for evaluating the proposed system, and Section 7 is the conclusion for this paper.

ABSTRACT. One serious issue with road safety is accidents caused by intoxicated drivers and sleepy drivers. An intelligent car safety system that uses the ESP32 microcontroller to keep tabs on the driver and their environment in real-time is showcased in this project. When the system detects blood alcohol concentrations (BACs) that are higher than a certain point, it will disable the ignition system. A sensor that detects when a driver is getting sleepy triggers a warning system that includes a buzzer . To warn of potential danger, an ultrasonic sensor scans the area for adjacent obstructions. The Gyroscope sensor continuously monitors the movement of vehicle and cut off the vehicle ignition when accident occurs. The ESP32 is responsible for handling sensor data and facilitating wireless connectivity for alerts sent to mobile devices. The suggested solution is affordable, dependable, and extensible for sophisticated Internet of Things (IoT)-based car safety applications.

ABSTRACT. The reliable operation of hybrid renewable energy systems, which simultaneously produce hydrogen and meet electricity needs, depends on efficient energy management. This paper proposes a rule-based energy management strategy (EMS) for a hybrid photovoltaic (PV) and wind system connected to a proton exchange membrane (PEM) electrolyzer and integrated with a shared DC bus. Safety, load satisfaction, battery protection, hydrogen production, and energy consumption limitation are among the system objectives that the proposed EMS dynamically prioritizes. It ensures stable and efficient system operation by coordinating power distribution between the PV modules, wind turbines, battery storage, and electrolyzer under various production and load conditions. A complete MATLAB/Simulink model of the hybrid system and its control approach was created and evaluated under various operating conditions, including load transients and the intermittency of renewable energy sources. The simulation results show that by redirecting excess power to hydrogen production, while respecting the operating limits of the batteries and electrolyzer, the proposed energy management system (EMS) increases the utilization of renewable energy. Furthermore, this plan ensures increased hydrogen production, constant DC bus voltage stability, and reliable power supply to the loads. These results confirm the robustness and efficiency of the proposed energy management system for hybrid renewable energy systems with integrated hydrogen production.

ABSTRACT. While the idea of electronic voting is not a new one, most implementations are far from ideal, rushing means legislative, procedural and technical issues that come with risks spanning from low trust in elections to even cancelled elections. The solution, validated during last national elections as a scope-limited pilot program, allows for a more gradual approach to electronic voting with the introduction of electronic smart ballot boxes, a hybrid system architecture that preserves the paper ballot as the legally authoritative artefact, yet enables near real-time aggregation with strong auditability. and serves as an early warning system for anomalies. As the current legal voting system still relies on paper ballots, so does our solution. The design separates voter identification from vote interpretation and counting and enforces data minimisation by avoiding long-term storage of individual vote results. This is done by maintaining integrity, traceability, and resistance to manipulation while mitigating ballot-secrecy leakage. This approach attempts to satisfy mandatory conditions, such as maintaining vote secrecy required by both national regulations in the case of countries such as Romania and similar EU electoral mandates. While the use of such a system proposes many challenges, a successful implementation can bring significant improvements to electoral systems around the world and serve to remove barriers in the public perception of electronic voting as a less safe or private form of voting.

ABSTRACT. This DC-DC non-inverting buck-boost converters are widely used in modern power electronic applications due to their capability to operate in both buck and boost modes while maintaining positive output polarity. Also, fast voltage regulation with low switching losses is still difficult in low-cost embedded systems, particularly for low-cost embedded platforms. This paper presents an adaptive fuzzy tuned Continuous Control Set Model Predictive Control (CCS-MPC) strategy for a non-inverting buck–boost converter with real-time implementation using an ESP32 microcontroller. The proposed controller updates the weighting factors of the predictive cost function according to the voltage tracking error and duty cycle variation using a fuzzy inference mechanism. In addition, a simplified absolute value-based cost function is adopted to reduce computational complexity and reduce execution time for real-time operation on low-cost embedded hardware. To reduce switching activity and switching losses, the converter operating region is divided into buck, boost, and buck-boost modes according to the relationship between the input voltage and the reference voltage. The buck-boost operating mode is restricted to a narrow region around the reference voltage (±0.5 V), allowing the converter to operate with a single active switching device whenever possible. This operating strategy reduces switching activity by approximately 50% while maintaining smooth transitions and stable output voltage regulation. The proposed controller was validated using both MATLAB/Simulink simulations and practical hardware implementation. The experimental setup includes an ESP32-based control platform with practical voltage sensing, low-pass noise filtering, and ADC protection circuits. Comparative results demonstrate that the proposed adaptive fuzzy CCS-MPC improves transient response, reduces overshoot, and achieves faster settling time compared with conventional fixed-parameter MPC and PID controllers under reference voltage and load variations.

ABSTRACT. Wireless sensor networks are commonly used for mobile target tracking, where distributed sensor nodes send their measurements to a fusion center for state estimation. But, using the proximity or information gain criterion for choosing sensors will inevitably activate the same nodes or the same spatial region along the target trajectory repeatedly. This results in unequal energy consumption across the network, as some portions of the network are being leveraged heavily while others are relatively underutilized, potentially resulting in less long-term tracking accuracy, or even losing the target path. To address this problem, this paper proposes a hybrid information- and energy-aware sensor scheduling method for EKF-based target tracking. The proposed method not only selects the informative sensors but also considers residual-energy distribution so as to minimize repeated use of the same nodes. This is compared to the All-in-Range, Closest-N, Random-N, and Info-only scheduling methods. The results indicate that the Hybrid method has tracking accuracy that remains close to the Info-only method. It outperforms the baseline methods by up to 99.03% and 96.39% at N_s=10 in reducing the node-level KL divergence and spatial KL divergence, respectively. The results prove that the proposed method is able not only to maintain the load fairness at the cell level, but also at the node level, and still preserve the tracking accuracy.

ABSTRACT. PHOTON-256/32/32 is a lightweight cryptographic hash function designed for resource-constrained applications such as Internet of Things (IoT) devices and embedded systems. However, the MixColumnsSerial transformation represents one of the most hardware-intensive components of the algorithm due to the finite field multiplication required for diffusion. This paper presents an FPGA implementation of PHOTON-256/32/32 that integrates a previously published Log/AntiLog-based GF(2⁸) multiplication technique into the MixColumnsSerial stage. The objective is to reduce hardware resource utilization while preserving the original functionality and diffusion properties of the PHOTON algorithm. Both the conventional and optimized architectures were implemented and evaluated under identical synthesis conditions on Xilinx FPGA platforms. Experimental results on a Kintex-7 XC7K160T-3FBG676 device show that the proposed architecture reduces Slice Registers from 295 to 41, Slice LUTs from 3605 to 306, and Occupied Slices from 1453 to 127. In addition, the maximum operating frequency increases from 360.188 MHz to 508.518 MHz, while the critical path delay decreases from 2.776 ns to 1.966 ns. The obtained results demonstrate that the proposed architecture provides a highly area-efficient implementation of PHOTON-256/32/32, making it suitable for lightweight cryptographic applications where hardware resource minimization is a primary design objective.

ABSTRACT. This research introduces a novel game theoretic proportionality assessment model for military operations con- sidering mappings from collateral damage descriptors and mil- itary advantage to proportionality outcomes. To this end, two perspectives are considered. First, a physical only model that treats collateral damage as injuries, deaths, and damage to civilian objects. And second, an extended model that incorporates psychological harm as part of the collateral damage component which conducts to a more restrictive proportionality boundary. The learned assessors are embedded as constraint/penalty mech- anisms within a Bayesian game between an attacking force and an adaptive adversary, enabling computation of disproportionate risk probabilities under uncertainty and strategic interaction. Moreover, for demonstration purposes, the model is evaluated through simulation demonstration on a counter-UAV (Unmanned Aerieal Vehicle) scenario in which a military Cyber Operation exploits a software vulnerability to seize control and force a safe landing, and is compared against kinetic intercept and RF jamming alternatives under adversarial postures. The results obtained show that the underlying ruling system can shift robust action selection toward control-oriented neutralization options when adversaries amplify collateral risk, and that explicitly accounting for psychological harm can materially affect permis- sibility assessments in otherwise borderline contexts.

ABSTRACT. The industrial transition of internal combustion engines to zero-carbon alternative fuels, such as hydrogen and ammonia, renders traditional engine calibration paradigms inadequate due to complex combustion kinetics. Historically, static lookup tables populated through steady-state dynamometer testing and traditional proportional-integral-derivative controllers fail to adapt to nonlinear dynamics during transient operations, such as sudden load increases; this limitation leads to severe knock risks and massive emission spikes. On the other hand, training autonomous artificial intelligence algorithms, such as deep reinforcement learning, directly on a physical engine poses risks of catastrophic hardware damage due to the trial-and-error mechanism of the agent. To overcome the manual calibration bottleneck, this conceptual study evaluates the potential of an autonomous control approach that integrates deep reinforcement learning agents with physics-guided digital twins. In these conceptually examined architectures, it is envisioned that engine control is formulated as a Markov decision process in a continuous action space, requiring the use of the soft actor-critic algorithm. It is supported by literature findings that the autonomous exploration process of the agent must be constrained by a multi-objective penalty and reward function embedded with thermodynamic rules. Literature findings and theoretical evaluations confirm that this rapidly developing simulation-to-reality transfer methodology can simultaneously optimize interdependent multiple variables, improve the integral absolute error in transient states by 20%, and reduce months of experimental calibration effort by 71%. Consequently, by presenting a digital twin perspective ranging from data synthesis via generative adversarial networks to autonomous decision-making via deep reinforcement learning, this study establishes a new industrial standard vision for autonomous engine management in the alternative fuel era.

ABSTRACT. Traditional CCTV systems rely on manual monitoring, resulting in delayed responses to unauthorized access. This study presents a Reactive CCTV System that integrates AI-powered facial recognition and logic-based behavior analysis on consumer-grade edge hardware. The system utilizes FaceNet for identity verification and YOLO11n combined with an Intersection over Union (IoU) spatial algorithm to track Human-Object Interactions (HOI). Tested in a controlled office environment, the FaceNet module achieved a 100% precision rate, ensuring strict access control despite a highly conservative 13.5% recall. The behavioral recognition module achieved 79.54% precision and 68.4% strict recall in identifying interactions with target assets such as laptops, bags, and cellphones. While the system proved highly effective for isolated interactions, the study identifies "proximity occlusion" as a primary limitation in cluttered environments where bounding boxes overlap. Ultimately, the integration of lightweight spatial logic and pre-trained models provides a viable, real-time automated security solution without the computational overhead of complex action-recognition networks.

ABSTRACT. Automated detection of human fatigue is of critical importance for a wide range of applications including occupa- tional safety, driving safety, and human performance monitoring. While existing fatigue detection systems predominantly rely on image processing and physiological signal-based approaches, the rich acoustic and linguistic information available in speech signals remains largely underexplored. In this study, we propose a multimodal deep learning framework for binary fatigue clas- sification (energetic vs. fatigued) using speech signals. A purpose- built dataset was constructed via neural speech synthesis under controlled acoustic parameters and prosodic disfluency injection to simulate energetic and fatigued speaking conditions. Three multimodal fusion strategies—Early Fusion, Mid Fusion, and Late Fusion—are evaluated, each combining acoustic features (Mel-spectrogram and handcrafted audio descriptors) with lin- guistic features derived via automatic speech recognition (ASR) followed by a BERT language model. An audio-only CNN+MLP baseline (82.5% accuracy) is also provided for comparison. Experimental results demonstrate that the Late Fusion model achieves the highest performance with 94.7% accuracy and 0.978 AUC-ROC. These findings establish that speech-based multimodal approaches offer a powerful and complementary alternative to camera-dependent and physiological methods for fatigue detection.

ABSTRACT. Nowadays, domain-specific question-answering systems require accurate and reliable responses, yet Large Language Models (LLMs) tend to produce hallucinated answers when lacking domain knowledge [4]. In this study, we propose a Retrieval-Augmented Generation (RAG)-based hybrid question answering system for Turkish hotel information and evaluate it on a pilot dataset representing Antalya-region hotels. The proposed system blends dense retrieval, BM25-based lexical retrieval [7], reciprocal rank fusion (RRF) [12], and cross-encoder reranking [9] with local LLM inference. In order to demonstrate the performance of the proposed model, 239 hotel records are collected, a controlled comparative analysis of chunking strategies and embedding models is conducted, and four local LLMs are evaluated. Experiment results show that the combination of Section-Based chunking, multilingual embeddings, and hybrid retrieval exhibits the best results with 32.6% MRR improvement, 82.6% F1 improvement over the LLM-only baseline, and eliminates observed hallucinations in the evaluated test set under the defined evaluation setup. These results highlight the importance of retrieval design and document segmentation in domain-specific Turkish QA systems.

ABSTRACT. Accurate automated segmentation of root canal fillings in periapical radiographs represents a critical bottleneck in scaling endodontic quality assessment, yet most existing deep learning approaches prioritize accuracy while neglecting deployment feasibility in resource-constrained clinical settings. This study evaluates deep learning-based segmentation models for automated segmentation of root canal fillings in periapical dental radiographs. Five widely used architectures—U-Net, U-Net++, Feature Pyramid Network (FPN), LinkNet, and SegFormer—were compared in terms of both segmentation performance and computational efficiency using a dataset of 597 annotated periapical radiograph images. Experimental results demonstrated that U-Net achieved the highest segmentation accuracy with a mean Intersection over Union (IoU) of 74.57 ± 1.34% and a Dice score of 85.42 ± 0.90% across five independent training cycles. Lightweight models such as FPN and LinkNet provided a favorable balance between performance and efficiency for resource-constrained clinical environments, with LinkNet achieving a GPU latency of 7.52 ms and throughput of 132.9 images/s. Additionally, pruning experiments on the best-performing U-Net model revealed that moderate unstructured pruning (10–30%) led to limited performance degradation recoverable through fine-tuning, whereas aggressive pruning and structured pruning at higher ratios resulted in substantial accuracy loss. These results highlight the trade-off between segmentation accuracy and computational efficiency and provide practical insights for the deployment of deep learning models in clinical dental imaging workflows.

ABSTRACT. Quantum computing is one of the most disruptive technologies to emerge this millennium. This research paper provides a complete study of quantum computing, touching upon all aspects of quantum computing from theory to practical application and scalability issues. Moreover, this paper offers an exhaustive bibliometric analysis of quantum computing over the time frame from 2020 to 2026, as well as a comprehensive literature review of 20 recent articles, including extensions up to 2026. The literature review section includes a comprehensive list of articles that have been chosen for review. All articles reviewed have been included in a table, which contains such information as publication year, venue, algorithms, accuracy/performance results, limitations found, and the most important of all – research gaps that should be studied in the future. The key insights gained during the analysis include evidence of tremendous growth within the domain (Compound Annual Growth Rate of 36.2%), the prominence of quantum error correction and quantum machine learning as subdomains experiencing growth, the advent of fault-tolerant logical qubits in 2025-2026, and decoherence and fault-tolerance overhead as the main bottlenecks preventing quantum computing from achieving practical results.

ABSTRACT. Organic food credibility depends on production claims that consumers cannot directly verify at the point of purchase. Existing digital traceability systems frequently admit cultivation-stage records without screening for suspicious upstream signals, and even when product histories are available, final representations are often too technical for non-expert shoppers to interpret quickly at the point of purchase. This paper presents an evidence-mediation framework that addresses both gaps through two cooperating components. At the cultivation stage, soil readings from a FastAPI-based simulation pipeline are assessed using an unsupervised Isolation Forest model trained on an eight-dimensional feature vector capturing both absolute nutrient concentrations and inter-reading change rates for Nitrogen, Phosphorus, Potassium, and Electrical Conductivity. Comparative evaluation against Local Outlier Factor and One-Class SVM confirmed that all three methods achieved zero false negatives on the anomaly test set, with Isolation Forest selected for deployment based on scalability and interpretability advantages. When unusual behavior is detected, the system cross-references the event against farmer-declared input records through a crop-batch-linked mobile application, issuing one of three evidence-based verdicts COMPLIANT, UNEXPLAINED SPIKE, or DATA MISMATCH rather than treating sensor flags as immediate non-compliance. At the consumer stage, validated batch evidence is condensed by a FastAPI confidence-synthesis service into freshness, organic grade, trust score, and warning cues using transparent deterministic rules, then rendered through a QR-triggered Unity-Vuforia AR overlay. The study argues that organic traceability becomes practically effective only when backend evidence is screened, contextualized, and reformulated into a concise trust language before consumer presentation.

ABSTRACT. Papaya cultivation is an important part of Sri Lanka’s agricultural sector however farmers continue to face challenges related to disease management, inconsistent fruit grading, improper harvest timing, and unstable market pricing. Most existing farming practices rely heavily on manual observation and experience-based decision-making, which often affect productivity and profitability. This research presents PapayaPulse, an integrated mobile-based smart farming framework developed to support data-driven papaya cultivation using artificial intelligence and machine learning techniques. The proposed system combines four major functionalities: papaya quality grading using image analysis, growth stage detection with first harvest prediction, disease and pest detection with severity-based treatment recommendations, and market price prediction for informed selling decisions. Deep learning models, including Convolutional Neural Networks and Vision Transformer architectures, were utilized for image-based recognition tasks, while Random Forest Regression and Classification, Gradient Boosting Regression and Classification models were applied for predictive analysis and decision support. Experimental evaluation was conducted using an 80:20 training and testing data split, where the image-based classification models achieved an average accuracy of approximately 87% under guided image-capturing conditions. In addition, a user evaluation survey conducted with farmers and agricultural users reported an overall user acceptance rate of 91%, highlighting the practical usability and understandability of the system. The mobile application supports both Sinhala and English languages, making the framework more accessible to local farming communities. The findings demonstrate that integrating multiple AI-driven agricultural services into a single platform can improve decision-making, reduce crop losses, and enhance transparency in papaya cultivation and marketing.

ABSTRACT. This research provides a comprehensive review of different biometric modalities with their advantages and disadvantages Biometric sensing-based systems are utilized to scan and identify individuals based on their behavioural or physiological traits. A single recognition device does not meet the requirements of recognition systems. The security of the system can be improved by multi-factor authentication where two or more security types (e.g. password and card) are used, but this is not the best security solution. Users may forget passwords, mistype them, or lose their card. For identification and categorization of people by their biometric features biometric devices are applied. These technologies fall under two categories, behavioural and physiological biometrics. The former has many limitations to the system accuracy, including noisy data, between class similarity, within class variability, universality and spoofing. Success rate of recognition and verification is however greatly enhanced with multimodal biometric sensing and processing systems, which employ the sensing or processing of more than one physiological and/or behavioural characteristic.

ABSTRACT. Standard Instrument Departure (SID) procedures define the post-takeoff climb and route integration that pilots must execute with precision, yet current training relies on two-dimensional charts that force pilots to mentally reconstruct three-dimensional flight profiles—a cognitive transformation that increases error rates during complex departures and high-traffic operations. Existing full-motion flight simulators address this gap at prohibitive cost, while low-fidelity desktop trainers lack geospatially accurate terrain and procedural fidelity. This paper presents SID-Viz3D, an interactive 3D visualization and simulation framework built on Unity 3D with Cesium geospatial integration that transforms ICAO-compliant SID chart data—including waypoint coordinates, altitude constraints (AT, AT_OR_ABOVE, AT_OR_BELOW, BETWEEN), and waypoint types (FLY_BY/FLY_OVER)—into a real-coordinate briefing and flight environment. The system provides real-time three-axis deviation tracking (lateral, vertical, and course) against planned procedure profiles, creating a basis for future quantitative pilot performance measurement without full-simulator infrastructure. Preliminary validation on Narita International Airport (RJAA) Runway 34L/34R SID procedures demonstrates that the framework preserves ICAO PANS-OPS procedural semantics while supporting real-time interaction in the 3D simulation environment.

ABSTRACT. Aircraft telemetry systems generate large volumes of rapidly changing flight data that are commonly presented through numerical dashboards and two dimensional monitoring interfaces. Although these systems provide access to raw telemetry information, operators may experience difficulty interpreting complex parameter relationships during real time missions, potentially reducing situational awareness and delaying abnormal condition recognition. This paper presents a real time 3D aircraft telemetry visualization platform that transforms live telemetry streams into an interactive spatial monitoring environment. The system visualizes aircraft 6DoF state, trajectory history, and control surface movements within a geospatial 3D environment. Abnormal or critical telemetry conditions are presented through HUD style interface overlays, presenting real time mission relevant information to support operator awareness during live operation. The proposed system is designed as a modular and extensible architecture that enables easy integration of different telemetry sources and supports adaptation to various mission specific requirements. To improve accessibility and reproducibility, the implementation uses openly available geospatial datasets, reducing dependence on commercial or restricted data sources.

ABSTRACT. Runtime fault diagnosis in Avionics Full-Duplex Switched Ethernet (AFDX), standardized under ARINC 664 Part 7, remains fundamentally limited in operational deployments. Faults such as Bandwidth Allocation Gap (BAG) violations, sequence number anomalies, babbling-idiot failures, and asymmetric redundancy path degradation manifest as transient, correlated perturbations across multiple switches and Virtual Links, rendering them invisible to single-stream packet inspection. Existing diagnostic approaches — including dedicated hardware analyzers and offline simulation frameworks — operate at the packet-trace level in isolation from the underlying network topology, and none provides interactive, topology-aware observation of fault propagation during live operation. This paper presents a software-based real-time framework for AFDX network analysis and fault diagnosis implemented in Unity. A Python-based traffic generator emits binary-encoded AFDX frames over UDP, preserving Virtual Link semantics and BAG constraints specified in ARINC 664 Part 7. The Unity runtime engine parses incoming frames, enforces protocol compliance — including BAG enforcement, sequence number integrity verification, and dual-redundancy management across Network A and Network B — and renders the results on a topology-aware interactive display. Six fault injection scenarios executed within the Systems Integration Laboratory on a five-end-system, five-Virtual-Link topology confirm detection of all fault types within one render cycle, providing fault observability that offline simulation frameworks do not support by design and that hardware analyzers achieve only through dedicated avionics interfaces. The framework architecture extends directly to full-scale avionics topologies and richer fault vocabularies without structural modification.

ABSTRACT. Multi-sensor imaging systems are widely used in defense, robotics, autonomous systems, and advanced perception applications. In such systems, maintaining accurate geometric alignment between sensors is critical for ensuring reliable system performance. However, mechanical tolerances, assembly imperfections, and environmental effects may introduce angular misalignments between sensors, leading to increasing Line-of-Sight (LOS) deviations and error propagation at the system level. Although existing studies mainly focus on calibration parameter estimation and alignment optimization, limited work has been conducted on intuitive and real-time visualization of angular alignment error propagation under varying scene conditions. In this study, a Unity-based interactive simulation framework is proposed for analyzing and visualizing angular alignment error propagation in multi-sensor imaging systems. The proposed environment models a master-slave camera configuration mounted on a shared mechanical structure, where angular offsets can be dynamically adjusted while the target scene distance is varied. Real-time LOS separation and propagated error metrics are calculated and visualized simultaneously, allowing observation of how small angular deviations evolve into significant scene-level divergences at long distances. Experimental analyses demonstrate that small angular offsets produce critical LOS separations as scene distance increases, with the propagation rate scaling significantly with the applied offset magnitude. The proposed framework also provides a preliminary decision-support mechanism for evaluating whether observed deviations can be corrected electronically through boresight compensation or require mechanical realignment. The presented approach offers an intuitive visualization-based analysis environment for understanding system-level error propagation in multi-sensor systems and provides a scalable foundation for future real-time augmented reality-assisted calibration and diagnostic applications.

ABSTRACT. The goal of this study is to empirically investigate whether and in what way hybrid quantum-classical neural networks (QNN) generalize differently from classical neural networks under data scarcity. Rather than claiming a quantum advantage, the study focuses on providing a rigorous and reproducible framework to evaluate neural network performance under different data scarcity regimes. Two classical feedforward neural networks and two QNNs are evaluated across six datasets. This evaluation includes both fractional and absolute scarcity regimes to capture different low-data conditions. Evaluated characteristics include generalization, overfitting, stability, and degradation. Results show that while classical models do achieve higher predictive accuracy, QNNs more frequently exhibit smaller generalization gaps and reduced overfitting signatures. Statistical analysis confirmed that classical models significantly outperform WNNs in accuracy across most evaluated settings. Additionally, QNNs display increased seed sensitivity. Overall, these findings suggest that classical and QNNs exhibit different generalization and degradation dynamics under scarcity, extending beyond predictive accuracy alone.

ABSTRACT. The worldwide expansion of DC fast-charging infrastructure is accelerating to meet the continuously rising demand from the electric vehicle sector. This rapid growth imposes considerable pressure on existing electrical distribution networks, especially due to the increasing deployment of high-power charging stations and their strict requirements for maintaining power quality and grid stability. In this context, power electronic converters play a central role in enabling efficient and reliable energy conversion between the grid and electric vehicles. To address these issues, this paper proposes a nonlinear PI (N-PI) based control strategy for a Vienna rectifier–based EV charging system. The proposed control method is designed to enhance dynamic response, improve disturbance rejection capability, and maintain stable DC-link voltage regulation under varying operating scenarios, including reference changes, load variations, and imbalance conditions. Furthermore, an N-PI–based neutral-point balancing controller is applied in the Vienna rectifier system to reduce DC-link capacitor voltage imbalance and maintain stable midpoint voltage regulation. The effectiveness of the proposed approach is verified through comparative analysis with conventional PI control, demonstrating improved overall system performance and robustness for next-generation fast-charging applications.

ABSTRACT. This paper investigates a single-phase dual-active-bridge (DAB) DC–DC converter employing a sliding mode control (SMC) strategy for electric vehicle (EV) applications. To enhance dynamic performance and ensure stable operation, an SMC-based strategy is designed and implemented, and its effectiveness is evaluated under different operating conditions. The proposed approach integrates a solid-state transformer (SST)-based DAB converter, enabling high-efficiency bidirectional DC–DC power conversion. A mathematical model of the system is developed, and the corresponding control algorithm is designed and validated. Both transient and steady-state performances are analysed under various input voltage and load conditions. In addition, the SMC strategy improves system robustness against disturbances such as parameter variations, external perturbations, and unmodeled dynamics in the voltage regulation loop. All modelling, control design, and simulations are performed in the MATLAB/Simulink environment, widely used for power electronics and power system studies.

ABSTRACT. The integration of technology into educational paradigms has become imperative in modern pedagogy. This paper presents “Turathy Quest,” an innovative educational game designed to address the challenges associated with traditional methods of teaching cultural and historical heritage, specifically within the context of the United Arab Emirates (UAE). Tradi- tional pedagogical approaches often rely on rote memorization, leading to student disengagement and poor long-term knowledge retention. Turathy Quest leverages immersive 3D environments and interactive gameplay to transform learning into an engaging, exploratory experience. Players navigate a virtual character through significant UAE landmarks, solving contextually relevant quizzes and collecting cultural artifacts like poetry. This work details the project’s objectives, its alignment with UAE’s strategic priorities for cultural preservation, and a comprehensive analysis of its technical and operational feasibility. A comparative analysis with existing educational tools highlights the unique proposition of a dedicated 3D exploration game for UAE heritage. The expected outcomes, analyzed through hypothetical yet realistic data models, indicate a significant potential for increased student engagement, improved knowledge retention, and the provision of a valuable tool for both formal education and cultural tourism. The project demonstrates the viability of using serious games to make learning about national heritage more effective and enjoyable.

ABSTRACT. The reliable prediction of Electric Vehicle (EV) battery State of Health (SOH) is one of the key factors for guaranteeing the dependable operation of vehicles, maximizing performance, and evaluating residual value. In practice, the main limitation is usually the lack of comprehensive operational data, since direct sensor data are generally not accessible to end, users and third, party systems. In this paper, we offer a thorough review and a comparative study of machine learning (ML) techniques that aim to tackle this data, gap issue. We explore various regression, based methods, starting from the classical Linear Regression and even SVR to ensemble methods like Random Forest and Gradient Boosting, as well as neural networks. A major aspect of this article is the ”multi, stage prediction pipeline”, an architectural pattern aimed at extracting high, value SOH features from scarce, user, provided inputs. This pipeline is structured in two phases: (1) Initially, it utilizes Charging Duration, Total KM Traveled, and Battery Type, the three most accessible data points, to model and predict a bunch of unobserved, intermediate operational parameters, namely SOC, Battery Temp (C), Ambient Temp (C), and Charging Cycles.(2) These derived features are subsequently mixed with the original user inputs to create a complete feature set, which is then used to train a second group of models for predicting final SOH indicators: Efficiency and Degradation Rate. By combining the results of various algorithms (based on R and MAE) performance within this two, stage framework, this manuscript recognizes and confirms this ”feature, inference” structure as a strong and practical resolution. It successfully closes the gap between highly data, demanding laboratory models and the stringent requirements of the real, world battery diagnostics, thus leading to scalable and easily accessible SOH estimation tools.

ABSTRACT. The existing LoRaWAN networks face problems because they can only operate in short distances, and they use one single transmission path which prevents their use in areas with remote locations or physical obstacles. This paper offers a complete assessment of how LoRaWAN technology developed from its initial single-hop system to its current multi-hop and mesh network structure. The study evaluates the routing performance and energy usage, modulation settings, and network expansion capabilities of LoRaWAN and LoRa Mesh systems. The study looks at how Spreading Factor (SF) and Chirp Spread Spectrum (CSS) modulation settings impact the distance transmission and data transmission capacity of LoRa devices. The review examines recent research on LoRa mesh networks that were published between 2022 and 2026, focusing specifically on the comparative findings that other studies reported. The surveyed systems utilize multi-hop communication as their primary method to increase coverage across different areas which includes 3 to 10 hops and beyond 10 hops. The reviewed studies demonstrate that adaptive spreading factor selection, commonly between SF7 and SF12, plays a key role in balancing transmission range and energy consumption. The majority of research studies utilize either energy-aware routing or lightweight routing systems, which enable them to communicate over distances of a few kilometers up to more than ten kilometers while using minimal power. Through a detailed review of existing routing protocols including reactive, proactive, and hybrid approaches. From the analysis of previous studies, we conclude that the spreading factor is directly proportional to the communication distance; specifically, the transmission range extends with each increasing spreading factor, while the higher spreading factor values enhance the ability to detect signals in very noisy conditions.

ABSTRACT. The high prevalence of mental health issues across the world has created a high demand for cost-effective and efficient technological solutions that can be used alongside existing clinical pathways. In this paper, the development of AIMind, a novel multi-service application created as an academic prototype at AISSMS Institute of Information Technology in Pune, will be discussed. The application’s architecture integrates four individually deployable services: a user interface created with React 19/Vite, a core server created with Node.js/Express/Socket.io, a Python/Flask inference backend, and a Mindspace community module. Two unique machine intelligence pathways power the application. The first pathway uses a journaling affect recognition model wherein user-written text is sent from the browser through a Node.js gateway to a Python child process running the jhartmann/emotion-english-distilroberta-base model. The model generates a seven-class emotional score vector stored in Supabase. The second pathway uses a dialogue model wherein a locally resident SmolLM2-360M model, managed by a stateful Flask service, keeps track of ongoing conversation history with each user. Running both models locally rather than through cloud services keeps user data private and reduces costs. In this paper, the authors describe the development of the application’s architecture, the choice of model used in each pathway, a comparison of the working prototype with the conceptual design at each layer of abstraction, a quantitative comparison with existing AI mental health systems, and the authors’ assessment of the ethical dimensions that remain unimplemented in the current build. Index Terms—mental health AI, affective computing, emotion recognition, DistilRoBERTa, SmolLM2, HuggingFace Transformers, Flask, Socket.io, Supabase, full-stack AI, privacy-preserving inference, digital well-being

ABSTRACT. This study presents a mathematical analysis of an integrated wind energy conversion system. The system consists of several integrated components: a wind turbine, a permanent magnet synchronous generator, an electronic power inverter, and a controller. A dynamic mathematical model was developed to represent the electromechanical behaviour of the entire system. A Maximum Power Point Tracking (MPPT) control strategy was modeled and developed using proportional integration (PI) controllers to optimize energy extraction efficiency. Real-world operating data from well-known international brands of 5 kW permanent magnet synchronous generators, as well as 6 kW wind turbines were used for simulation and validation. A comprehensive evaluation and study of the system's performance under varying wind speed conditions was conducted. Furthermore, the proposed system was tested using wind speed data from different regions in Iraq (specifically, Diyala Governorate) to assess its practical feasibility for the country. The results demonstrate improved efficiency and confirm the effectiveness of the proposed control strategy under varying operating conditions.

ABSTRACT. Supraharmonics arise as a result of the widespread use of inverters in renewable energy sources, and these propagate from the PCC to other components of the grid. Supraharmonic components emitted from nonlinear loads can also interact with supraharmonic components from such sources, leading to a highly complex harmonic spectrum spread across the high-frequency range. In this study, the interaction of harmonic components originating from an inverter in a grid-connected PV inverter system with a rated power of 120 kW and a switching frequency of 40 kHz, with power converter loads connected to the PCC at various switching frequencies, and their effects on the voltage harmonic spectrum are analyzed. Then, the effects of these supraharmonic components on the magnetization current of the LV/MV transformer are investigated.

ABSTRACT. This paper provides a detailed comparative study of conventional Proportional-Integral (PI), Genetic Algorithm (GA) tuned PI, and Sliding Mode Control (SMC) controllers for Maximum Power Point Tracker (MPPT). In our previous work, we established that the GA tuned PI controller demonstrated a 38% greater power tracking efficiency and an 80% faster response time compared to conventional PI controllers. For this current work, we have studied and compared these findings with the SMC controller to enhance Maximum Power Point Tracking (MPPT) performance in wind system applications. The SMC controller exhibits superior robustness with 45% better power tracking efficiency and 85% faster response time under varying wind conditions compared to conventional PI. Yet, when compared to our previously developed GA-tuned PI controller, SMC demonstrated competitive performance but was still surpassed by the GA-tuned approach in terms of overall efficiency and higher computational requirements. This comparative study concludes that although both advanced control algorithms significantly improve wind turbine control performance, the GA-tuned PI controller is still the optimal choice for wind energy systems, considering the balance between performance, robustness, and implementation complexity.

ABSTRACT. The growing energy crisis has driven research toward renewable energy systems for electricity generation. Integrating multiple sources can enhance performance, and while recent studies have combined two renewable sources, this study integrates three—wind, solar, and hydro—to maximize power output. The sources were modeled using transfer function techniques and integrated, with a PID controller included to improve system stability. From the MATLAB/Simulink simulation results, it is evident that the PID controller demonstrated a remarkable performance by completely eliminating the overshoot from 59% to 0%, albeit at the cost of a slightly slower rise time, which increased from 0.15 s to 0.40 s, while also effectively limiting the transient oscillations to fewer than four cycles. The stability analysis performed on the system showed that the use of the PID controller reduces the system gain from 45 dB to 0 dB. These results demonstrate the effectiveness of the proposed approach for developing a robust hybrid renewable power generation system.

ABSTRACT. This paper proposes an improved Direct Torque Control (DTC) method based on fuzzy logic for an induction motor drive system using a three-level Neutral Point Clamped (3L-NPC) multilevel inverter. The proposed control structure integrates a fuzzy PI controller in the outer speed control loop to generate the reference torque, while a fuzzy-based DTC scheme is implemented in the inner loop to replace conventional hysteresis comparators and switching tables. The proposed method reduces electromagnetic torque ripple and stator flux ripple, improves switching frequency stability, and enhances overall control performance. In addition, the utilization of the 3L-NPC inverter with a rich set of voltage vectors contributes to improving the output voltage and current quality. Simulation results obtained from the Matlab/Simulink environment demonstrate that the proposed method improves dynamic response, reduces torque ripple, enhances speed tracking capability, and decreases current harmonic distortion compared with the conventional DTC method. Therefore, the proposed control structure can be considered an effective and practical solution for high-performance induction motor drive applications.

ABSTRACT. This paper presents a comparative evaluation of continuous-control-set model predictive control (CCS–MPC) and finite-control-set model predictive control (FCS–MPC) for a three-phase grid-connected voltage-source inverter. Both controllers are developed using the same synchronous referenceframe current model, system parameters, phase-locked loop, and reference-current profile to ensure a fair comparison. In the continuous-control-set scheme, a continuous modulation vector is optimized and implemented through space-vector modulation, which enables fixed-switching-frequency operation. In contrast, the finite-control-set scheme directly selects the optimal switching state from a finite set of inverter voltage vectors, thereby eliminating the need for an intermediate modulation stage. The two methods are evaluated under three active-current reference levels of 0.5 A, 5 A, and 10 A, while the reactive-current reference is kept at zero to achieve near-unity-power-factor operation. Simulation results show that both controllers successfully track the current references and keep the reactive-current component close to zero. Under the considered parameter setting, finitecontrol-set model predictive control achieves lower currenttracking error, smaller current ripple, and lower phase-current total harmonic distortion. At the 10 A operating point, the total harmonic distortion of the phase current is reduced from 2.42% with continuous-control-set model predictive control to 1.76% with finite-control-set model predictive control. However, continuous-control-set model predictive control preserves a fixed switching frequency of 20 kHz, whereas finite-control-set model predictive control exhibits a slightly varying average switching frequency in the range of 20.8–21.05 kHz. The results highlight the practical tradeoff between current quality and switchingfrequency regularity in predictive control of grid-connected inverters.

ABSTRACT. Industrial motor fault diagnosis is significantly affected by environmental noise and sensor degradation, which reduce the reliability of conventional deep learning models. This paper proposes a multi-sensor fault diagnosis framework based on Multi-Head Attention and LSTM networks enhanced with a Reliability-Gated fusion mechanism. The proposed framework dynamically evaluates the reliability of current, vibration, and stray-flux signals before feature fusion. Experimental results demonstrate superior diagnostic performance compared with conventional LSTM-Attention models, with an accuracy of 97.67% when the motor is operating under full load. Furthermore, under severe noise conditions (20 dB SNR), the proposed model maintains 91.5% accuracy. The reliability-gating strategy preserves robust diagnostic performance during total sensor failure, achieving an accuracy of 92.4%, while maintaining sensitivity to incipient winding faults with only 2% severity. The proposed framework provides a fault-tolerant solution for predictive maintenance in industrial environments.

ABSTRACT. The modern power systems are developing quickly due to the higher integration of renewable energy resources, electric vehicles, distributed generation systems and intelligent sensing devices. In these aspects, the developments have considerably enhanced online associated appliance tracking and operational control of power grids. The ability to maintain internal situational awareness has thus become critical for sustaining system stability, operational dependability, and effective power flow management under complex and dynamically evolving operating conditions. Typical state estimation approaches are vulnerable to performance degeneration resulting from nonlinear environments, noise, sparsely populated coverage of sensors, communication time delays, and large-scale network structures. To overcome such limitations, this work proposes an interpretable TFT-based hybrid deep learning framework for real-time wide-area power system state monitoring/forecasting applications. In the proposed framework, a physics-inspired state estimation module is combined with a transformer-based temporal forecasting architecture This hybrid approach can catch both short-term fluctuations (voltage profiles and power flow trajectories) as well as long-term temporal dependencies of renewable generation variability and load demand patterns . In contrast to traditional recurrent learning methods, the TFT architecture uses multi-head attention, gated residual learning and feature selection mechanisms to achieve both state-of-the-art forecasting performance alongside an interpretable model that is operationally transparent. The proposed framework is validated on IEEE 57-bus and IEEE 118-bus benchmark systems through noisy and partly missing measurements under dynamically varying operating conditions. Experimental results are provided to show that estimation based on the developed framework improves both estimation accuracy, forecasting stability and reduces computational cost compared with classical Gauss-Newton methods, feedback on neural network methods, recurrent neural networks (RNNs) and long short-term memory (LSTM)-based approaches. Additionally, the attention-based interpretability mechanism accurately identifies key operating intervals and dominant system variables affecting grid dynamics. The framework itself also showed high scalability and applicability for real-world smart grid monitoring and energy management systems in the coming ages.

ABSTRACT. More renewables, scattered power sources, plus modern converter tech have turned old-style electricity systems into fast-changing smart setups. Even though advanced controls using deep learning adjust better and run more smoothly, their lack of clarity makes them hard to trust when stability matters most. Because of this gap, a new approach called XDLF is introduced here - built to make deep learning decisions clearer in managing power electronics across smart networks. Inside, it brings together time-based convolution models, two-way memory networks that track past and future patterns, along with focus layers that highlight key inputs during shifting, unpredictable operations. On top of that, methods from Explainable AI using SHAP help uncover how models make choices, pointing out which electrical traits most affect converter behavior. Instead of fixed conditions, testing happens across shifting situations - like harmonics, uneven renewables, shifts in voltage, and changing loads - using open smart grid data.

ABSTRACT. The KisanBandhu Project introduces a decision support system based on artificial intelligence to help smallholder farmers make informed choices about the best crops for their soil and the climate around them. The system uses 7 machine learning algorithms, specifically Random Forest, CatBoost, XGBoost, Gradient Boosted, Decision Trees, KNearest Neighbors, and Support Vector Machine). These algorithms were trained using a total of 2,200 observations of 22 types of crops and with 7 input features while attempting to compensate for class imbalance through use of synthetic samples. For the models that were evaluated, CatBoost outperformed all the others to provide the best classification accuracy at 80.0%, and was also able to provide fairly equal precision, recall and F1-scores, all of which are close to 0.80

ABSTRACT. Efficient waste classification is critical for automation in modern waste management systems, especially in places where it is hard to use manual classification. In this work, a model for waste and non-waste classification based on the YOLOv8n framework, known for high efficiency and applicability for embedded systems, is designed. For this purpose, a customized dataset was collected using real photos under different lighting conditions, different backgrounds, and with varying objects to reflect a more realistic environment. The training process included augmentation, normalization, and hyperparameters tuning to improve the efficiency of the model. Evaluation metrics such as accuracy, precision, recall, F1 score, and confusion matrix were used. According to the experimental results, the developed model provides an accuracy of 92.5% and F1 score of 0.94 with maintaining fast inference performance for real-time applications. The designed framework aims to become a module in an autonomous waste collecting robot to validate objects before manipulating them by robots, ensuring reliable and efficient real-world system performance.

ABSTRACT. Body Mass Index (BMI) cannot differentiate muscle mass from fat distribution, while manual somatotyping offers comprehensive assessment but lacks scalability. This study developed and validated an AI-driven web application for automated somatotype classification from 2D camera images with personalized diet recommendations. The system employed a five stage hybrid pipeline: CNN-based silhouette feature extraction, SVR/Random Forest prediction of anthropometric measurements (trained on CAESAR dataset, N=4,431), Bayesian Ridge calibration to Filipino anthropometry, deterministic somatotype computation, and multivariate bias correction for seven-category classification. Validation on 30 Filipino males using nested Leave One-Out Cross-Validation achieved 76.7 percent accuracy (Cohen’s kappa of 0.670, substantial agreement), improving from 36.7 percent baseline with measurement MAE of 0.21 to 3.49 mm. The rule-based diet engine received Scale-Level Content Validity Index of 1.00 from certified Nutritionist-Dietitians. Automated camera-based somatotype classification approaches human inter rater reliability and offers scalable health profiling, though low Mesomorph recall (17 percent) and pilot sample size require further validation.

ABSTRACT. Abstract— Polypharmacy substantially increases the risk of drug-drug interactions (DDIs), which threaten patient safety and clinical outcomes. Graph neural networks have emerged as powerful tools for DDI prediction; however, most current approaches treat binary classification and multi-class interaction type prediction as separate tasks, reducing their practical clinical value. These models also demand significant computational resources and large training datasets, creating obstacles for both development and real-world deployment. We propose a unified hierarchical hypergraph neural network that jointly predicts DDI occurrence and interaction types through two stages. HGNN-Chemical (hypergraph neural network Chemical) learns drug embeddings from SMILES substructures, achieving 98.46% ROC-AUC and 93.2% accuracy for binary prediction. HGNN-Interaction type (hypergraph neural network Interaction type) leverages these embeddings to predict 86 interaction types, achieving 85.73% Top-1 and 98.12% Top-3 accuracy. Systematic ablation studies confirm that hierarchical structure and attention flow direction significantly improve performance. Our model outperforms multiple baselines using only SMILES input, enabling predictions for novel compounds without feature engineering. Trained on DrugBank with 1,709 drugs and 191.877 interactions with CPU-only processing.

ABSTRACT. This paper proposes an Offline Multimodal Retrieval-Augmented Generation (RAG) System that is fully processed on local hardware, enriching the conventional RAG system with support for text, image, and audio modalities. The suggested architecture combines features of multimodal retrieval and generation, with a focus on privacy, efficiency, and accessibility. In contrast to cloud-based solutions, the system uses on-device encoding, retrieval, and generative functions, which allows organizations and individuals to implement intelligent assistants, knowledge agents, and multimodal analysis tools without sharing sensitive information with external servers. It builds on recent progress in the open-source LLMs and multimodal embedding models to form a unified local inference pipeline. The architecture has three key components: (1) Multimodal Encoders transform heterogeneous inputs (text, images, audio) into homogeneous vector representations using trained embedding models; (2) a Local Vector Store, implemented using FAISS or ChromaDB, which facilitates the efficient semantic search and retrieval of relevant documents or media (via cosine similarity); and (3) an Offline Generative Model, which synthesizes coherent, contextually informed responses by conditioning on the retrieved information. The whole system, including encoding, retrieval, and generation, runs in a safe, offline environment and ensures data confidentiality and lower latency than cloud-based APIs.The system is tested against benchmarks for retrieval accuracy, generation quality, response latency, and computational efficiency. Accuracy in text retrieval will be measured by Recall at K, Mean Reciprocal Rank (MRR), whereas the performance of the generators will be evaluated by BLEU, ROUGEL, and BERTScore..

ABSTRACT. The rapid proliferation of AI systems across safety-critical domains has exposed a fundamental flaw in current evaluation methodologies: automated benchmarks, however sophisticated, are insufficient proxies for real-world trustworthiness. We introduce AISA --- the AI System Assessment Protocol --- a formal, reproducible hybrid evaluation framework that systematically integrates automated AI scoring with structured human expert judgment. AISA defines a four-layer architecture comprising Task Specification, Automated Evaluation, Human Expert Review, and Hybrid Arbitration, governed by an arbitration function that dynamically weights human and machine contributions based on confidence, domain sensitivity, and historical accuracy. We formalize the protocol mathematically, specify inter-rater reliability guarantees, and demonstrate its superiority over both purely automated and purely human evaluation baselines across three benchmark domains. AISA is designed to become an international reference standard for AI tools evaluation, with an open specification that scales from individual research groups to regulatory bodies. This work establishes both an academic standard and the methodological foundation for the Aisetapp.com evaluation platform.

ABSTRACT. Digital supply chains are evolving to include autonomous software agents. These agents support the execution of actions in areas such as planning, logistics, inventory, and manufacturing. The roles of AI agents and their impact vary across the respective domains. This study reviews recent literature sources after 2021 and presents real-world case studies to: i) distinguish agent-based from conventional AI; ii) define a two-dimensional typology that classifies agents by functional area and level of autonomy (monitoring, forecasting/consulting, and prescription/execution); and iii) indicate the effects of increasing the performance of AI agents on resilience, efficiency, and sustainability. The results of the literature review show that agents with high autonomy for inventory/logistics and smart manufacturing are associated with operational potential, such as faster replanning, lower waste and energy consumption, and better service levels. In contrast, procurement and compliance agents are emerging and understudied. The study discusses the need for governance and a human in the loop for safe deployment. It highlights gaps in research on inter-organizational orchestration and ethical oversight. This study outlines the benefits and challenges of deploying AI agents in digital supply chains, as well as opportunities to assess their strategic impact.

ABSTRACT. Digital sovereignty has emerged as a critical concern in cloud computing, driven by increasing regulatory requirements and data localization mandates. This systematic literature review examines architectural models, key technologies, and implementation challenges in sovereign cloud ecosystems, with particular emphasis on artificial intelligence integration. Following a prisma protocole, we analyzed 46 papers published between 2020 and 2026 from Scopus, Web of Science, and ScienceDirect databases. Our findings reveal that decentralized architectures, particularly federated learning and blockchain-enabled systems, dominate sovereign cloud implementations. Key enabling technologies include trusted execution environments, edge computing, and privacy-preserving machine learning techniques. The integration of AI technologies presents both opportunities for enhanced automation and challenges related to computational overhead and model governance. Major technical challenges include interoperability across heterogeneous systems, balancing privacy with performance, and ensuring real-time processing in distributed environments. This review provides a comprehensive taxonomy of digital sovereignty approaches and identifies critical research gaps in standardization, cross-border data governance, and scalable AI deployment in sovereign clouds.

ABSTRACT. Rare-event prediction in non-stationary environments remains a central challenge in applied machine learning, particularly in operational settings where class imbalance, temporal dependence, and distributional shift co-occur. This study presents a drift-aware evaluation framework for temporally faithful assessment of rare-event prediction models and demonstrates the framework on the publicly available MIMIC-IV intensive care dataset. The analysis compares three major predictive modeling paradigms—gradient-boosted decision trees, Temporal Convolutional Networks, and Transformer-based temporal models—under chronologically ordered training, validation, and test partitions that reflect progressively increasing temporal drift. Rather than relying on pooled evaluation, the proposed framework examines model behavior across regime-specific test windows and jointly evaluates discrimination, constrained-retrieval quality, calibration, robustness, and computational efficiency. The results show that gradient boosting remains highly competitive under mild drift and strong efficiency constraints, Temporal Convolutional Networks provide the most stable balance between retrieval performance and robustness under progressive drift, and Transformer models achieve the strongest retrieval under severe drift, albeit with higher calibration error and computational cost. The study contributes a reusable evaluation framework for drift-sensitive rare-event learning and provides practical guidance for model selection in temporally evolving decision environments.

ABSTRACT. This research explores the application of big data analytics and machine learning in forecasting movements in the Indian stock market. The objective is to design and evaluate predictive models that estimate stock price fluctuations using historical financial data, focusing on four major companies—HDFC Bank, Infosys, State Bank of India (SBI), and Tata Motors. Multiple approaches are assessed, including traditional statistical techniques such as ARIMA, machine learning algorithms like Support Vector Machines (SVM), and deep learning architectures including LSTM, BiLSTM, CNN-LSTM, and ResNLS. To improve prediction accuracy, the study employs advanced data preprocessing, feature engineering, and time-series modeling. Model performance is measured using evaluation metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the coefficient of determination (R²). Findings indicate that the integration of big data and AI-driven models can enhance decision-making in investment strategy and risk management by uncovering complex patterns within financial datasets and providing actionable insights for the Indian market.

ABSTRACT. The widespread adoption of digital technologies for knowledge access has exposed the limitations of static presentation methods in enabling immersive and interactive historical learning, a gap that is particularly evident in the context of Sri Lankan history. To address this, a multi-component AI-driven museum platform is introduced to enhance cultural heritage engagement through integrated interaction, exploration, and reasoning capabilities. The platform integrates four complementary modules, including a graph-based reasoning framework, Global Context Explorer for Sri Lanka (GCESL), which operates on a spatio-temporal representation of historical events to identify plausible global influences on local contexts, a conversational museum guide enabling real-time and context-aware interaction with artifacts and historical personas, an adaptive cultural exploration system supporting cross-cultural understanding and immersive engagement, and an AI-driven narration and artifact reconstruction module for multimodal storytelling and visualization. This integrated approach supports a transition from static information access toward inquiry-driven learning environments and provides a scalable framework for enhancing digital cultural heritage experiences.

ABSTRACT. Incremental Neural Networks (INNs) serve as fundamental instruments for attaining highly accurate approximation outcomes. Specifically, convexity constitutes a crucial property in the context of approximating neural networks within incremental frameworks. This characteristic facilitates the derivation of precise approximations of the target function. Moreover, convexity is applicable to both constrained (i.e., convex) and unconstrained approximation scenarios. Furthermore, leveraging convexity in both network architectures and approximation processes enables the determination of optimal function approximations through incremental neural networks. In the present study, we perform an approximation of neural networks by driving the difference sequence between the target function and the Modulus of Convexity in Incremental Neural Networks (MCINNs) toward zero. Consequently, we also derive the degree of approximation via neural network approximation, yielding a rate of O(1/n^(p+1) ). These theoretical findings are corroborated by experimental results obtained using MATLAB, wherein an algorithm initialized with the neural network was implemented. Additionally, the algorithm demonstrates, for several test functions, an improved approximation quality compared to the standalone neural network. The convex incremental iteration algorithm progressively approximates the objective function by means of a neural network, relying on a sequence of functions to achieve enhanced approximation performance. The residual error was compared against the Mean Squared Error (MSE), resulting in a lower residual error value.

ABSTRACT. Ceylon cinnamon is one of Sri Lanka’s most important agricultural exports, yet smallholder farmers face challenges in yield prediction, nutrient management, pest and disease identification, and market price uncertainty. This paper presents CinoGrow, an integrated mobile-based decision support system designed for precision cinnamon farming. The system incorporates four machine learning modules: yield prediction using YOLOv5 and XGBoost, nutrient deficiency detection using a vision transformer-based model, pest and disease detection using a Roboflow object detection model, and price forecasting using an ARIMA time-series model. The application supports multilingual interfaces and limited offline functionality through edge-deployable models. Field evaluation involving 90 farmers in Galle and Matara districts demonstrated a 93.3% task completion rate and an average user satisfaction score of 4.4/5. Results indicate that CinoGrow can improve decisionmaking, resource efficiency, and productivity in cinnamon farming systems.

ABSTRACT. The immutable nature of smart contracts means that they can suffer permanent and catastrophic financial losses due to just one line-of-code vulnerability, as in decentralized finance (DeFi) assets, hundreds of billions of dollars worth are being managed on public blockchains. Current security solutions are severely limited: static analyzers have high false positive rates, supervised deep learning models are extremely expensive to train, and existing active learning setups are extremely susceptible to pseudo-label poisoning. Moreover, there is no solution that can be integrated that is able to combine high level of accuracy for vulnerability detection with the ability to explain why the vulnerabilities exist and to automatically remediate them. To fill these gaps, this paper proposes an integrated, four-phase modular framework that addresses the full life cycle of vulnerability management including detection, forensic auditing, explainability and automatic remediation. The first phase is a model known as ASSBert, which is a CodeBERT-based model improved with a new multi-head attention pooling mechanism and a tri-modal active learning strategy. ASSBert achieves 98.95% classification accuracy over five important classes (reentrancy, integer overflow, timestamp dependency, dangerous delegatecall, safe), using a special "defer zone" to capture the predictions at the border of the classification space, avoiding contamination of the training set, and reduces the manual labeling time of the experts by 72%. The second step is a modular forensic framework, which connects Random Forest detection, semantic reasoning with LLM, and symbolic execution verification in a single discern-understand-prove pipeline that can greatly reduce false positives. The third phase brings in SCXAI – an explainable AI system – which combines model-intrinsic feature importances with a fine-tuned Flan-T5 model, enabling complex code signals to be converted into structured explanations, legally defensible, and presented in natural language for forensic purposes. Last but not least, there is a post-detection remediation module that includes a rule-based patch engine with a 96% syntactically correct code fix rate, a financial attack simulator and an ML-based risk predictor that predicts the probability of financial attacks with 91% accuracy. As a whole, these co-operative pieces move smart contract auditing from a disconnected, black-box hunt for bugs to a comprehensive, end-to-end, and practically-instantly-deployable security system.

ABSTRACT. Studying approximation spaces for neural networks (NNs) is of great importance, not only theoretically, but also for practical applications, such as designing, generalizing, and improving the performance of the NNs. In this paper, NNs were built based on multi-channel convolution and the concept of a tensor product of matrices. The concept of generalized and strict NNs was introduced. Also, the basic properties of the parameters in artificial NNs were studied, and present the classes of NNs to identify the properties that facilitate the study of NNs. Moreover, the definition of approximation spaces for generalized and strict NNs is presented, and it is proven that the classes of approximation spaces are quasi-Banach spaces and achieve continuous embedding. Therefore, the main goal of studying NNs is to determine the range of functions from quasi-Banach spaces, where approximation spaces provide an overview of the goal by identifying its properties and relationships. Finally, a multi-convolutional neural network model was built and trained on a representative dataset using a reduced loss function, with the mean squared error (MSE) used as a performance evaluation metric. The results showed a significant decrease in error, reflecting the model's ability to learn and achieve high approximation accuracy, generating good agreement between actual and expected values. This highlights the effectiveness of neural networks as a powerful tool for addressing mathematical modeling and numerical approximation problems.

ABSTRACT. Underground parking lots present challenges for semantic segmentation due to low-texture surfaces, weak illumination, and severe class imbalance, which often lead to blurred boundaries, missed fine structures, and limited real-time performance in general-purpose models. To address these issues, this paper proposes ParkSegNet, an engineering-oriented adaptation of the DeepLabV3+ framework for parking lot scenarios. A lightweight MobileNetV2 backbone is adopted to reduce computational cost and improve inference efficiency. To mitigate extreme class imbalance and enhance the segmentation of elongated structures such as parking lines, a combined loss function integrating Focal Loss and Dice Loss is employed. In addition, an efficient channel attention mechanism is incorporated into the atrous spatial pyramid pooling module to suppress pseudo-feature amplification under low-light conditions. The activation function is further replaced with h-swish to improve gradient continuity and feature representation capability. Experiments on the SUPS underground parking lot dataset demonstrate that ParkSegNet achieves a balanced improvement in segmentation accuracy and real-time performance while maintaining model lightweightness. The proposed approach provides a practical solution for deploying semantic segmentation models in resource-constrained parking lot environments.

ABSTRACT. For millions of deaf and hard-of-hearing people around the world ,sign language serves as their primary form of communication. However, the hearing population’s limited comprehension of sign language means that communication barriers still exist. Even though sign language recognition research has advanced significantly in recent years, many current methods only use isolated signs or heavily annotated datasets and computationally demanding models, which restricts their use in real-time and resource-constrained settings. this paper presents a vision-based lightweight framework for phrase-level American Sign Language recognition from video sequences. In order to capture the dynamic evolution of hand gestures, the suggested method uses MediaPipe for hand skeletal landmark extraction and Long Short-Term Memory (LSTM) networks for temporal modeling. The system preserves the necessary motion information for precise recognition while reducing visual complexity by representing gestures as series of hand landmarks . Robustness to signer variability, environmental changes, and dynamic gesture transitions are key components of the methodological design which is based on review of recent literature. The suggested framework is appropriate for real-time assistive communication and human–computer interaction applications since as analysis of recent literature and system behavior suggests that landmarkbased temporal modeling provides a good trade-off between recognition accuracy and computational efficiency.

ABSTRACT. Intelligent monitoring technologies are being increasingly incorporated into smart agricultural practices to improve crop yields and resource utilization efficiency. In this paper, we propose an intelligent monitoring framework for smart greenhouses using multimodal monitoring approaches, including Internet of Things-based environmental monitoring, camera-based visual monitoring, and machine learning-based decision support systems. The system continuously monitors vital microclimatic factors, including temperature, humidity, and luminosity, using an IoT-based distributed sensor network. The environmental monitoring module employs an unsupervised anomaly detection algorithm, known as the Isolation Forest algorithm, to analyze the environmental data, allowing the system to identify abnormal climate patterns without requiring prior knowledge of normal patterns. At the same time, an intelligent camera-based monitoring module performs visual monitoring of the plant conditions using an image classification algorithm, providing additional information about the plant conditions beyond the environmental monitoring module. The results of both monitoring approaches are fused using a multimodal fusion algorithm to improve the accuracy of anomaly detection and reduce false alarms. The decision support system uses rules to provide recommendations based on the detected anomalies, displayed using an interactive dashboard interface. The proposed system design is efficient, scalable, and interpretable, making it appropriate for deployment in resource-constrained agricultural environments. The system’s integration of IoT, artificial intelligence, and machine learning technologies makes it an innovative solution for advancing precision agriculture, thereby promoting sustainable food production globally.

ABSTRACT. This project aims to create an automated Vision-Based Attendance Marking System that is expected to replace the inefficient and inaccurate manual attendance systems currently used in educational and professional settings. The system is developed using Python and OpenCV for video processing and the Face Recognition library for identity verification using 128-dimensional facial embeddings. After the identity is confirmed, the attendance information of the individual (name, date, and time) is recorded automatically in a Microsoft Excel or Google Spreadsheet. Keywords: Computer Vision, Face Recognition, Python, OpenCV, Arduino, NumPy, Automated Attendance, IoT, Real-time Database.

ABSTRACT. Effective communication between hearing individuals and those who are deaf or hard of hearing (DHH) remains a significant challenge in education, healthcare, and public services. This paper presents SunoSign, a full-stack bidirectional assistive communication system for Indian Sign Language (ISL). The system operates through two pipelines: (1) speech or text input in five Indian languages is translated into animated ISL gesture sequences, and (2) real-time webcam-based hand gestures are recognized and converted to text using a trained deep learning model. The browser-based frontend is implemented using HTML, CSS, JavaScript, the Web Speech API, MediaPipe Hands, and the Canvas API. The backend is a Python FastAPI server hosting a TensorFlow/Keras feedforward neural network for sign recognition and a scikit-learn TF-IDF intent classifier. A multilingual pipeline using deep-translator supports English, Hindi, Marathi, Tamil, and Gujarati. The recognition model processes 63-dimensional MediaPipe landmark vectors (21 landmarks × 3 coordinates) and classifies 36 ISL characters (A–Z and 0–9) with an inference latency below 50 ms on a standard CPU. The model was trained using the Adam optimizer (learning rate = 0.001), categorical cross-entropy loss, a batch size of 32, and 50 epochs over an 80/20 train-validation split, achieving 92.8% validation accuracy. An intent-aware chatbot provides contextual responses, further enriching user interaction. SunoSign is designed as a deployable, end-to-end assistive platform to support ISL accessibility across education, medical, and government services.

ABSTRACT. Distracted driving caused by smartphone usage remains a significant contributor to road traffic accidents worldwide. Studies have consistently shown that mobile phone interaction while driving increases cognitive workload, visual distraction, and crash risk. This study suggests creating a prototype Driving Mode AppLocker that uses smartphone accelerometer data to identify vehicle motion and blocks access to distracting mobile apps in order to address this issue. For real-time motion categorization, the system combines an Android smartphone with a Raspberry Pi 4 platform and makes use of a one-dimensional Convolutional Neural Network (1D CNN) model that was trained using Edge Impulse. To ascertain whether the user is in a moving vehicle or stationary, the accelerometer data is processed as time-series inputs. Once driving motion is detected, predefined applications such as messaging and social media apps are automatically disabled using Android accessibility services. The system aims to reduce distracted driving behavior through automated enforcement rather than voluntary compliance. Performance evaluation will be conducted using a confusion matrix, targeting at least 85% classification accuracy. The proposed prototype contributes to intelligent transportation safety systems and supports Sustainable Development Goal 3.6 by promoting road safety through embedded artificial intelligence and edge computing technologies.

ABSTRACT. The human shoulder’s biomechanical complexity and wide range of motion make it prone to maladaptive movement patterns during exercise. Laboratory-based motion capture systems provide precise kinematic data but are costly and restrictive, limiting the study of naturalistic movement. This research explores shoulder kinematics and associated upper-body movements during gym exercises using a non-invasive and practical video-based approach. Monocular RGB videos at 60 fps were recorded in a gym environment. MediaPipe BlazePose 3D was employed to track upper-limb key points, while cervical alignment was indirectly estimated from the geometric relationship between the nose landmark and the midpoint of the bilateral heel landmarks, providing a measure of head-neck deviation relative to the vertical axis. Upper-limb kinematics and cervical posture were analyzed to evaluate movement coordination and postural control during loaded exercises. During shoulder exercises, cervical deviations were observed that, in some cases, reflected natural postural adjustments or reflexive responses. Involuntary movements, such as brief shifts in head position or gaze, occasionally amplified these deviations, particularly during sustained loaded tasks. The importance of postural control during resistance exercises is emphasized by these findings, which demonstrate how subtle, involuntary behaviors can influence cervical-shoulder stability and upper-limb coordination. Monocular pose estimation combined with geometric inference successfully captured meaningful kinematic patterns without the necessity of a specialized laboratory. The proposed framework provides a practical, non-invasive, and economical method to study shoulder and associated upper-body kinematics outside the laboratory. It enables assessment of movement patterns, detection of maladaptive behaviors, and evaluation of exercise technique in real-world settings.

ABSTRACT. Malware is a significant cybersecurity risk, particularly when it is embedded in executable formats like Portable Executable (PE) files. This research aims to deliver a whole machine learning pipeline for malware detection using the PE Files Malwares dataset. The dataset was thoroughly preprocessed, followed by experiments using ten different feature selection techniques, of which SelectKBest using ANOVA F-score performed optimally, reducing the dataset to the top ten most informative features. To establish comprehensive baselines, a number of machine learning and deep learning algorithms were trained and tested using 5-fold stratified cross-validation. The Random Forest classifier performed optimally, achieving an average accuracy of 95.95, closely followed by Decision Tree at 95.94 and XGBoost at 95.69, all of which reported consistent precision, recall, and F1-score metrics of about 0.96. On the other hand, deep learning models like MLP, Conv1D, and LSTM performed poorly at 54.90, 55.08, and 45.46 respectively. This significant performance gap is fundamentally attributed to the structured, tabular nature of the ten selected features, which limits the representational advantage of deep neural networks compared to high-dimensional or sequential domains. This research suggests that traditional ensemble-based machine learning models are more suitable for malware detection than deep learning models, and feature selection is critical for better model performance.

ABSTRACT. Outlier detection in highly imbalanced datasets remains a critical challenge in machine learning, particularly in applications such as fraud detection, anomaly monitoring, and cybersecurity. Traditional approaches often suffer from performance degradation due to skewed class distributions, noise, and high-dimensional feature spaces. In this paper, we propose a hybrid outlier detection framework that integrates Isolation Forest with a multi-stage data enhancement pipeline combining statistical resampling, dimensionality reduction, and generative modeling. Specifically, the proposed approach leverages SMOTE and SMOTE-ENN to address class imbalance and noise, Principal Component Analysis (PCA) for feature space transformation, and Generative Adversarial Networks (GAN) for realistic synthetic data generation. Furthermore, a data refinement module, referred to as TriEnhance, is introduced to improve feature quality and reduce noise effects. Experiments conducted on multiple benchmark datasets demonstrate that the proposed framework consistently outperforms classical and state-of-the-art methods, achieving superior performance in terms of Recall, F1-score, and PR-AUC. The results highlight the effectiveness of combining complementary data enhancement strategies for robust anomaly detection.

ABSTRACT. Feature extraction reduces high-dimensional image data into compact and informative descriptors, improving classification efficiency and performance. In this study, moisture content classification is performed using color image analysis. RGB channels are first separated, and 26 features are extracted from each channel, including statistical measures (mean, standard deviation, variance, median, minimum, maximum, skewness, kurtosis, Shannon entropy, and Renyi entropy) and texture features derived from Gray-Level Co-occurrence Matrix (GLCM). GLCM features contrast, correlation, energy, and homogeneity—are computed across four orientations (0°, 45°, 90°, and 135°). To enhance performance, Chi-square-based feature selection is applied to rank and select the most discriminative features. A Support Vector Machine (SVM) with a radial basis function (RBF) kernel is used for classification. Feature-level fusion of RGB channel combinations is also investigated. The model is evaluated using 4-fold and 10-fold cross-validation. The proposed method achieved a highest classification accuracy of 91.54 and 90.42 for 3-class and 5-class classification of moisture content, respectively. Results demonstrate that optimal feature selection and channel combinations significantly improve classification accuracy and robustness.

ABSTRACT. In the current study, an integrated machine learning (ML) framework has been proposed for the automatic detection and classification of the four main hematologic diseases, i.e., leukemia, thalassemia, anemia, and malaria, from the microscopic blood smear images. The traditional approaches in the diagnosis of blood-related diseases are based on microscopic examination, which is a tedious, subjective, and skill-dependent approach, sometimes resulting in delays in the detection of these diseases, particularly in underdeveloped countries. To overcome these limitations, the current study proposes an integrated detection system using disease-specific deep learning models, including a DenseNet-MobileNet hybrid network with a Convolutional Block Attention Module (CBAM) for the detection of leukemia, a MobileNetV2-ViT two-stage hybrid network for the detection of thalassemia, a ResNet50-EfficientNet hybrid network for the detection of anemia, and a CNN-ViT network for the detection of malaria. In addition, the proposed system will include the integration of Explainable Artificial Intelligence (XAI) techniques, i.e., Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP), to improve the interpretability of the results. The proposed system will ensure the accuracy of the results to be above 80%, the inference time below 5 seconds, and will be made available through a cost-effective web platform to ensure the applicability of the system in both urban and rural settings.

ABSTRACT. Insider threats represent a critical challenge to enterprise cybersecurity, as the misuse of authorized credentials can lead to devastating data breaches. While centralized User and Entity Behavior Analytics (UEBA) systems are effective at modeling normal behavioral patterns to detect these threats, they often face significant privacy and regulatory constraints. This paper presents a Federated User and Entity Behavior Analytics (F-UEBA) framework designed to detect insider threats through a decentralized, privacy-preserving Bidirectional LSTM Autoencoder. Concurrently, we investigate the impact of parameter compression techniques, specifically FP16 Quantization and Top-K Sparsification, to address the communication bottlenecks inherent in federated insider threat detection. Our experiments on the CERT r4.2 dataset reveal that a hybrid compression approach reduces communication overhead by approximately 85% (from 4,295.7 MB to 642.3 MB) while maintaining a high detection fidelity, as indicated by a PR-AUC of 0.718. Bootstrap confidence intervals confirm that this compression level does not statistically significantly degrade detection performance relative to the uncompressed federated baseline. These findings establish an optimal operational balance for bandwidth-constrained, privacy-preserving behavioral analytics in real-world enterprise environments.

ABSTRACT. Compressive sensing (CS) is a recent approach to minimize energy consumption for multi-target detection and localization in Wireless Sensor Networks (WSNs). However, it is complicated by real-time constraints because traditional sparse recovery approaches, such as Basis Pursuit (BP), are computationally costly and cannot be applied as the target density increases. This paper presents the Alternating Direction Method of Multipliers-Net (ADMM-Net) deep unfolding for an adaptive model. This model reformulates a task into a multi-label detection and builds a fixed network structure. Simulation results show that the proposed model is adaptive to both sparse and dense targets and reduces execution time by 99.95 percent (2000 times faster) compared to BP, making it suitable for real-time WSN applications.

ABSTRACT. This study develops a hybrid AI-driven framework combining Random Forest, Social Network Analysis (SNA), and Graph Convolutional Networks (GCNs) for credit risk classification. Centrality metrics—PageRank, eigenvector, betweenness, closeness, and degree—quantify borrower influence, brokerage roles, and detect potential default cascades. The best-performing centrality-driven random forest model achieved 91.23% accuracy (precision and recall ¿89% for both default and non-default), while the GCN-stacking hybrid attained 90.28% accuracy with balanced precision, recall, and F1-score. SNA visualizations are provided to explicitly illustrate borrower influence, network positioning, and risk propagation, enhancing interpretability and scientific rigor.

ABSTRACT. The rise of social media platforms has facilitated rapid communication but also led to the widespread dissemination of hate speech, particularly in low-resource languages such as Hindi. This study presents a deep learning-based approach for detecting hate speech in Hindi using a Bidirectional Long Short-Term Memory (BiLSTM) architecture. A dataset of 15,000 annotated posts—sourced from Twitter, newspapers, and televised news—was curated, capturing both formal and informal language, including code-mixed Hindi-English content. To enhance robustness and generalization, the dataset was combined and split into three randomized train-test configurations (10k-5k), with the model trained and evaluated independently on each. Preprocessing steps included tokenization, padding, and label encoding, with text sequences passed through an embedding layer followed by stacked BiLSTM and dense layers. The model achieved consistent accuracy across all splits (72.67%–74.10%), demonstrating its stability under varied data distributions. The findings underscore the linguistic challenges of hate speech detection in Hindi and propose a multi-split evaluation framework as a reliable alternative to single-split benchmarks. This work contributes to the growing body of research on inclusive and context-aware content moderation systems for underrepresented languages, and lays the groundwork for future advancements involving transformer-based models and multi-label classification.