The growing intricacy of medical data presents a significant challenge for electronic health record (EHR) systems, which are required to integrate multimodal data in order to enhance diagnostic precision and treatment efficacy. This paper proposes a novel GPT-4-based approach to enhance medical data processing capabilities in EHR systems. In particular, the GPT-4 was initially employed to encode clinical text data, and a pretraining and fine-tuning strategy was devised based on domain knowledge with the objective of enhancing its comprehension of medical terminology and context in accordance with the specific requirements of the medical field. Secondly, in order to address the issue of incomplete text data in EHRs, we propose a contextbased adaptive filling strategy to facilitate the dynamic completion of missing information and enhance the integrity of text data through the analysis of analogous historical records. In order to enhance the model's capacity to comprehend lengthy texts, we have employed a hierarchical attention mechanism to partition lengthy texts into multiple sub-blocks. Through the adjustment of hierarchical attention weights, GPT-4 is capable of effectively capturing cross-paragraph and cross-topic relationships. Furthermore, data augmentation technology has been employed to generate diverse training data through semantic extensions and data augmentation models, thereby enhancing GPT-4's adaptability to diverse inputs. The experimental results demonstrate that the proposed method has led to notable enhancements in prediction accuracy, text generation capability, and data processing efficiency.

This work proposes an ethical artificial intelligence optimization framework for cybersecurity tasks aiming at maximizing robustness, detection efficiency, minimizing computational complexity, and being compliant with ethics principles (fairness, transparency, etc.). Focusing on fast convergence on the one hand and a diverse solution search on the other, the framework employs metaheuristic optimization approaches like PSO and GA. We test this framework with real-world cyberattacks, for example, CICIDS and UNSW-NB15 datasets, to get results signifying adaptability to various attacks, with average F1-scores of 0.92 due to PSO and a high R(w) = 0.18 robustness metric via GA. Its explainability mechanism with SHAP and LIME explains the model predictions showing that features like packet size and connection durations were pivotal in impacting the model output. The learned adaptive configurations of optimization parameters (α, β, γ) make the framework customizable for diverse cybersecurity applications and accommodate tradeoffs between detection performance and adversarial robustness. Additionally, the framework is lightweight, allowing scalability to large datasets and high-dimensional use cases. Privacy-preserving approaches such as Federated Learning can also be incorporated into this framework, enabling its application to real-time cybersecurity attacks alongside hybrid optimization techniques. Such a framework will enable the development of secure, efficient, and ethically robust AI systems in sectors where AI must address the 21st-century security challenges.

Thermophotonic devices are radiative heat engines in which the exchange of electroluminescent radiation between a heated light-emitting diode and a cool photovoltaic cell allows for the conversion of heat into electrical power. Here, we introduce CRESCENT-1D, the solver we have developed to simulate the performance of one-dimensional thermophotonic systems, which is made publicly available on GitHub. It couples photon transport in the far or near field, based on the fluctuational electrodynamics framework, and charge transport in heterostructures, modelled with the drift-diffusion and Poisson equations. We include both thermionic emission and charge carrier tunnelling to precisely model charge transport at heterointerfaces, while the photon chemical potential is computed in a self-consistent manner between the radiative and electrical sections of the solver. Compared to simpler formulations, these models provide accurate results at high voltages, which is essential to achieve high power output. The capabilities of CRESCENT-1D are illustrated with an optimised InGaP/InGaAs thermophotonic heterostructure, whose maximum power reaches 1.6 W.cm −2 for an efficiency of 19.7% considering a 300 K temperature difference between the light-emitting diode and the photovoltaic cell. This solver makes it possible for anyone to design various categories of optoelectronic structures (thermophotonics, light-emitting diode, thermophotovoltaics, thermoradiative, etc.), and represent an important step in the development of near-field radiative heat engines.

This paper presents the application of Infrared Thermography to measure distribution transformer efficiency. The method is very intuitive, practical, and easy to use. The results obtained with the proposed method are compared to those obtained by the standard procedure, showing excellent agreement with them. The paper presents the nature of the losses in energy transformation, the theoretical basis of the proposed method, and the comparison of its results with those obtained from the standard procedure.

With the increasing integration of distributed energy resources (DER) into the distributed power system, the security of the power system from cyber-attacks is paramount. Cyber attacks make unintentional islanding detection challenging within the networked DER-integrated microgrid system. This study proposes a physics-informed approach for feature extraction and dimension reduction, leveraging principles from physics and domain-specific knowledge to analyze three-phase voltage and current signals. Moreover, a recurrent neural network (RNN) based gated recurrent unit (GRU), is introduced to fortify networked DER-integrated microgrids against cyber-physical threats, particularly focusing on unintentional islanding triggered by cyber-attacks at the point of common coupling (PCC). The most essential and difficult stage in an intelligent islanding detection system (IIDM) is feature extraction and selection, for which a physics-informed wavelet scattering network (WSN) and minimum redundancy maximum relevance (MRMR) algorithm are proposed. The WSN facilitates enhanced signal representation by capturing low and high-frequency information simultaneously, ensuring translation-invariant and deformationstable signal representations. The MRMR algorithm is applied for dimension reduction, ensuring that the reduced feature space retains the most informative and physically relevant features while minimizing redundancy. Finally, GRU network is proposed for islanding detection. We developed an API integrated with RT-Lab to generate a diverse dataset for islanding, faults, and nonislanding scenarios to gather PCC voltage and current signals. The proposed method is assessed under various islanding, faults, and non-islanding scenarios, also considering the non-detection zone (NDZ), a critical factor affecting islanding conditions using an OPAL-RT real-time digital simulator. The proposed method is also validated using accuracy, selectivity, and sensitivity performance indices, demonstrating the effectiveness of physicsinformed feature extraction/selection and GRU-based islanding detection (PI-IIDM), ensuring accurate islanding detection.

Edge computing has emerged as a pivotal solution to meet the growing demand for real-time data processing and reduced latency in modern distributed systems. However, its decentralized nature increases the attack surface, making it vulnerable to Distributed Denial of Service (DDoS) attacks. Traditional security models, which rely on perimeter-based defenses, struggle to address these threats due to their inability to effectively manage insider threats, dynamic traffic, and sophisticated attack vectors. This paper explores the integration of Zero Trust Security Models (ZTSM) with behavioral analytics to enhance DDoS detection and mitigation in edge computing environments. Zero Trust principles, such as "never trust, always verify," enforce strict access control, continuous authentication, and micro-segmentation, reducing unauthorized access to critical resources. Behavioral DDoS detection complements this by analyzing patterns in network traffic, user behavior, and resource utilization to identify anomalies indicative of potential DDoS attacks. The proposed framework leverages machine learning algorithms to create adaptive and context-aware detection mechanisms that are scalable across edge nodes. By combining ZTSM with behavioral detection, the model offers proactive mitigation strategies, ensuring minimal disruption to edge services. This integration also addresses challenges like scalability, real-time decision-making, and false positive reduction. The study demonstrates the effectiveness of this approach through simulated DDoS scenarios in edge environments. Results indicate significant improvements in detection accuracy, response time, and overall resilience of the edge infrastructure. These findings underscore the importance of adopting Zero Trust principles and advanced behavioral analytics to safeguard edge computing systems against evolving DDoS threats. Future work will focus on enhancing model interoperability, extending the solution to heterogeneous edge environments, and incorporating predictive capabilities to anticipate and mitigate threats proactively.

This paper presents a wireless power transfer (WPT) for a mid-sized inspection mobile robot. The objective is to transmit 100 W of power over 1 meter of distance, achieved through lightweight Litz wire coils weighing 320 g held together with a coil structure of 3.54 kg. The Wireless Power Transfer System (WPTS) is mounted onto an unmanned ground vehicle (UGV). The study addresses an investigation of coil design, accounting for misalignment and tolerance issues in resonancecoupled coils. In experimental validation, the system effectively transmits 109.7 W of power over a 1-meter distance, with obstacles present. This achievement yields a system efficiency of 47.14%, a value that is remarkably close to the maximum power transfer point (50%) when the WPTS utilises the full voltage allowance of the capacitor. The paper shows the WPTS charging speed of 5 minutes for 12 V, 0.8 Ah lead acid batteries.

Multi-Link Operation (MLO) in Wi-Fi 7 is expected to tangibly boost throughput while lowering transmission latency at the same time. This is very relevant in industrial scenarios and makes MLO suitable, e.g., to support seamless device mobility. Benefits depend on the ability of multi-link devices to select at run-time the best link, among the available ones, in order to maximize both communication performance and reliability. In this paper an experimental platform is proposed, with the aim of leveraging commercial hardware and open source software, and easing prototyping and evaluation of MLO techniques. The platform has been employed to analyze the transmission quality of two pairs of non-overlapping channels, and in particular to assess whether or not adequate diversity is provided, so that those channels can be exploited to improve reliability. Results point out that correlation between different links is, in most cases, limited, which makes MLO a valuable approach.

Current trajectory datasets of traffic participants often lack detailed environmental information, which is crucial for developing effective data-driven methods for future mobility solutions. To address this gap, we introduce the comprehensive DLR Urban Traffic dataset version 1.2.0. The dataset includes 32,296 trajectories of traffic participants, along with traffic light data, local weather data, air quality data, and road condition data collected at the Application Platform for Intelligent Mobility Research Intersection during a single day of recording. A comparison with other publicly available datasets reveals that our dataset offers more comprehensive information about the traffic environment than existing alternatives. An analysis of our dataset shows that trajectories are available for all possible 16 routes at the intersection, with the number of trajectories per route varying significantly, ranging from 11 to 3,344. Most interactions between motorized road users occur during unprotected left turns with oncoming traffic. However, there are also interactions between motorized and vulnerable road users, particularly during right turns. All in all, the dataset provides researchers with the resources needed to improve urban mobility solutions.

Federated learning has emerged as a promising approach for distributed machine learning, particularly in the edge-cloud infrastructures, where clients (i.e., edge devices) collaboratively train models while maintaining data privacy. However, the dynamic heterogeneity induced by the FEderated Edge Learning (FEEL) environment poses many performance challenges. Indeed, real-world edge devices are considered heterogeneous, where both system resources and data distributions fluctuate significantly over time across edge clients. Thus, selecting the most contributing and resource-efficient clients is crucial in determining the computational efficiency and performance of the overall FEEL system. We propose FedCDRP, an efficient and adaptive Class Diversity and Resource Performance-aware client selection strategy, founded upon our FedCD-CS and FedRP-CS methodologies. Our approach prioritizes clients based on two factors: their resource capabilities and the diversity of their local data. Specifically, FedCD-CS tackles the imbalanced non-I.I.D. data by considering the richness of the local data of each edge device, to enhance the overall learning system by selecting the most contributing clients. Additionally, based on a real-time cross-dimensional system performance profiling, FedRP-CS efficiently assesses the system resources of each edge device, to guarantee the selection of the most resource-efficient clients to effectively manage the computational needs. By combining these factors, our FedCDRP strategy identifies clients that can contribute significantly to the global model update while maintaining computational efficiency in dynamic heterogeneous environments. Through real-world circumstances, the conducted experiments demonstrated the computational efficiency of our FedCDRP strategy, compared to the-state-of-the-art protocols, achieving a 2.9-fold speedup in reaching the target accuracy of 99%.  

Traffic flow prediction is a critical area of research within Intelligent Transportation Systems (ITS), aiding in traffic planning, control, and more. Real-world traffic flow exhibits inherent non-stationarity, presenting a significant challenge for deep forecasting models. Current non-stationary traffic flow prediction models often decompose the sequence into multiple frequency components, process them in groups, and then recombine to get the output. This approach can lead to a loss of information from the original sequence. In this paper, we propose a novel framework for non-stationary traffic flow prediction called Trend-Variation Separated Learning Framework (TVSLF). TVSLF addresses these limitations by decomposing the raw traffic flow input into trend and variation terms. These terms are processed separately through NET1 and NET2, respectively, before being combined to generate the final output. TVSLF effectively mitigates two key challenges in traffic flow forecasting: underfitting the detailed variations of true values in multi-horizon predictions and lagging in single-horizon predictions. Our method's effectiveness and efficiency are validated against existing methods using four real-world datasets.

Efficient resource allocation is essential for optimizing various tasks in wireless networks, which are usually formulated as generalized assignment problems (GAP). GAP, as a generalized version of the linear sum assignment problem, involves both equality and inequality constraints that add computational challenges. In this work, we present a novel Conditional Value at Risk (CVaR)-based Variational Quantum Eigensolver (VQE) framework to address GAP in vehicular networks (VNets). Our approach leverages a hybrid quantum-classical structure, integrating a tailored cost function that balances both objective and constraint-specific penalties to improve solution quality and stability. Using the CVaR-VQE model, we handle the GAP efficiently by focusing optimization on the lower tail of the solution space, enhancing both convergence and resilience on noisy intermediate-scale quantum (NISQ) devices. We apply this framework to a user-association problem in VNets, where our method achieves 23.5% improvement compared to the deep neural network (DNN) approach.

This research enhances and evaluates the performance of a Decentralized Nash Bargaining (DNB) adaptive traffic signal controller that operates a flexible National Electrical Manufacturers Association (NEMA) phasing and timing scheme responding dynamically to the fluctuating traffic demand. The DNB controller is enhanced to 1) use traffic density estimates instead of queues to optimize signal timings; 2) to consider the eight-phase two-ring NEMA controller configuration within the game-theoretic approach; and 3) to consider dynamically adaptable control time steps. The enhanced DNB controller is benchmarked against 1) fixed-time traffic signal control using the state-of-the-practice Webster's method and an emerging Laguna-Du-Rakha (LDR) method for computing the optimum cycle length; 2) state-of-practice actuated traffic signal control; and 3) state-of-art Reinforcement Learning (RL) traffic signal control. The controller is tested on two isolated signalized intersections demonstrating enhanced overall intersection performance compared to the baseline pre-timed and actuated controllers at various demand levels and offers comparable performance to a previously developed RL controller. Specifically, the DNB controller results in a decrease in the average vehicle delay and queue size by up to 54% and 63%, respectively compared to Webster's state-of-the-practice pre-timed control. Unlike the RL controller, the DNB controller requires no pre-training while adapting to fluctuating traffic conditions, thereby providing a flexible framework for reducing traffic congestion at signalized intersections. As such, this research contributes to the development of smarter and more responsive urban traffic control systems.

Modern systems face increasingly sophisticated threats, driving the need for detection frameworks capable of identifying complex behavioral anomalies. Through integrating neural architectures with flowgraph analysis, a versatile framework was developed to classify malicious activities with high accuracy and adaptability. Temporal dependencies and structural patterns in system behavior are dynamically analyzed to differentiate between benign and malicious operations, even in the presence of obfuscation and polymorphic strategies. Experimental results demonstrated consistent outperformance over traditional methods, particularly in terms of reducing false positive rates while maintaining robust classification precision. The modular design facilitates deployment across a variety of operational environments, from distributed systems to real-time detection scenarios. Scalability testing revealed that detection performance remained stable even with increasing data complexity, affirming its suitability for large-scale applications. Adversarial testing highlighted the framework's resilience against perturbation attacks, ensuring reliable performance under evolving threat landscapes. Efficiency metrics, including energy consumption and latency, further validated its practicality for enterpriselevel integration. Adaptability to diverse ransomware families, such as LockBit, Hive, and Conti, was demonstrated through thorough experiments, illustrating its versatility in addressing different attack vectors. Challenges posed by imbalanced datasets were effectively mitigated, showcasing the system's capacity to maintain accuracy in skewed conditions. Real-world applications extend to securing critical infrastructure, IoT ecosystems, and cloud platforms, where rapid and accurate threat detection is crucial. Neural-Mimetic Flowgraph Analysis emerges as a substantial advancement in cybersecurity, offering a robust solution to evolving digital threats.

In this paper, a tunable piezoresistive pressure sensor is designed and developed using junctionless nanowire (JL-NW) gate-all-around (GAA) field-effect transistors (FETs) integrated into a circular diaphragm. The measured results show that the piezoresistive sensitivity is improved by ~4 times in the subthreshold (SS) regime (depletion regime) compared to the ON-state (ION) condition (partial depletion regime). This improvement is achieved by reducing the channel conductivity with a low gate bias below the threshold voltage (VTH). Piezoresistivity variability is observed to be ~20% in the SS regime and ~10% in the ION conditions due to a diameter variation of ~±4 nm in the nanowires during a single fabrication run. The symmetric thin diameter of nanowires also shows a higher gauge factor compared to the asymmetric thickness of nanowires. This variability estimation suggests that controlling fabrication process variations is crucial for developing stable sensing elements for nanoelectromechanical (NEMS) sensors in advanced technology nodes. Additionally, low-frequency noise (LFN) is measured to estimate the minimum detectable strain (MDS) and signal-to-noise ratio (SNR). These measurements indicate that the JL-NW GAA FET offers higher resolution and better performance as a sensing element compared to the inversion mode NW GAA FET in advanced technology nodes. In addition to offering better reliability and sensing resolution, it is also much easier to integrate with advanced CMOS technologies.

Continual learning, also known as lifelong learning, aims to enable artificial intelligence systems to learn from a continuous stream of data while retaining previously acquired knowledge. Unlike traditional machine learning approaches that assume static datasets and offline training, continual learning faces unique challenges such as catastrophic forgetting, knowledge transfer, and scalability. Over the years, a variety of strategies have been developed, including regularization-based methods, memory replay techniques, and dynamic architectural adaptations. This survey provides a comprehensive review of continual learning, covering fundamental concepts, major learning paradigms, and state-of-the-art methodologies. We discuss the primary challenges that hinder the deployment of continual learning models in real-world applications and examine existing solutions, highlighting their strengths and limitations. Furthermore, we explore diverse application domains, including computer vision, natural language processing, and robotics, where continual learning plays a pivotal role in enabling adaptive intelligence. Despite significant progress, several open problems remain unresolved, necessitating further research. We outline future directions such as neuroscience-inspired learning mechanisms, meta-learning for rapid adaptation, scalable and efficient model architectures, and task-agnostic continual learning. Additionally, we emphasize the importance of ethical considerations, fairness, and security in developing responsible lifelong learning systems. By addressing these challenges, continual learning has the potential to revolutionize artificial intelligence, enabling autonomous systems to learn in an open-ended and evolving manner. This survey serves as a reference for researchers and practitioners interested in advancing the field and building more robust and adaptable AI models.

In modern automotive technology, the majority of energy produced by internal combustion engines is lost as waste heat, primarily through heat dissipation from exhaust gasses. This inefficiency has significant implications for fuel consumption and environmental impact, especially in an era of increasing energy demand and concerns about pollution. Thermoelectric generators (TEGs) offer a promising solution by converting waste heat directly into electrical energy, thereby improving fuel efficiency and reducing emissions with the benefit of not introducing moving parts or complex systems. First, this paper provides a comprehensive review of TEGs in automotive waste heat recovery, exploring their underlying principles, material advancements, and most optimal conditions for the best power output and efficiency. Next, key design challenges are addressed, including heat exchanger optimization, thermoelectric material efficiency, leg geometry, and system integration, all to determine the method in which TEGs can be used to convert the most possible heat to electric power as efficiently as possible. Finally, the paper discusses potential enhancements, research gaps, and future directions in TEG development.

The expanding capabilities of language models have intensified the need for adaptive mechanisms that can manage dynamic contextual shifts in extended interactions. Current approaches primarily rely on static embeddings, which limit the model's responsiveness to real-time contextual variations, thereby hindering coherence in complex dialogues. The proposed Dynamic Language Interaction Mapping (DLIM) framework addresses this gap by introducing a novel methodology for continuous context-sensitive adaptation, enabling more fluid, relevant responses. Through a series of quantitative analyses, the DLIM-enhanced model exhibited substantial improvements across metrics such as contextual relevance, coherence over longer sequences, response diversity, and robustness to input noise, significantly outperforming baseline models. The architecture of DLIM not only advances the adaptability of language models but also broadens their functional applications, providing a scalable solution suitable for deployment in multiuser environments. These findings demonstrate the potential of DLIM to transform contextually adaptive language technologies, supporting a range of applications that require both high accuracy and flexibility in response generation.