Large-scale models for language tasks have demonstrated remarkable capabilities in generating and understanding text across multiple domains, yet challenges remain in efficiently processing complex syntactic structures and handling long-range dependencies. The novel concept of Dynamic Token Hierarchies (DTH) offers a significant advancement by introducing a multitiered token processing framework that dynamically adjusts token representations based on contextual and semantic relevance, optimizing both performance and computational resource allocation. Through a series of experiments, the DTH-enhanced model outperformed traditional models, demonstrating substantial improvements in accuracy, efficiency, and the ability to process linguistically intricate tasks. The implementation of DTH reduced memory usage and processing time while improving task-specific outcomes such as classification and named entity recognition. This innovative framework provides a robust solution to the scalability challenges associated with complex language models, proving its efficacy in balancing computational efficiency with enhanced linguistic comprehension.

A novel three-phase grid inverter topology with a split dc link and LC filter is proposed. It allows for a full parallel connection of multiple inverters simultaneously on both the ac and dc sides, offering exceptional modularity, redundancy, expandability, and overall system reliability. A holistic dynamical system model is derived, considering the coupling effects between parallelized inverters, physical constraints, and varying grid impedance. A decentralized proportional-integral state feedback control (PI-SFC) with an extended Luenberger state observer is developed and compared with a conventional PI controller to control the inverterside currents. A stability analysis shows that both closed-loop control systems remain stable even if an arbitrarily large number of grid inverters are connected in parallel. Simulations and experiments confirm the functionality and robustness of the closed-loop system under varying grid impedances and during grid faults. For the experimental results, the controllers were implemented on commercially available hardware of the proposed topology. In particular, the PI-SFC allows for better exploitation of the full power of the inverters due to its better controller performance and damping ability. Finally, the exceptional match between simulation and measurement results proves the accuracy of the proposed system model as well.

A comparison between the use of PAM and QAM constellations in Rate-Splitting Multiple Access (RSMA) is presented. Previous studies of RSMA with discrete or finite input signaling have considered either QAM or PSK. We propose the use of PAM constellations for every stream, with a phase rotation of 90 degrees applied to the common stream to achieve orthogonality with private streams. Numerical results based on data rate and Symbol Error Rate (SER) analysis indicate that PAM is superior to QAM when there is a significant disparity between channels for a two-user scenario. This finding indicates that PAM constellations can extend the operational region of RSMA over other Multiple Access (MA) strategies.

Nowadays, cybersecurity has grown into a more significant and difficult scientific issue. The recognition of threats and attacks meant for knowledge and safety on the internet is growing harder to detect. Since cybersecurity guarantees the privacy and security of data sent via the Internet, it is essential, while also providing protection against malicious attacks. Encrypt has grown into an answer that has become an essential element of information security systems. To ensure the security of shared data, including text, images, or videos, it is essential to employ various methods and strategies. This study delves into the prevalent cryptographic methods and algorithms utilized for prevention and stream encryption, examining their encoding techniques such as advanced encryption standard (AES), Blowfish, Twofish, Salsa20, and ChaCha20. The primary objective of this research is to identify the optimal times and throughputs (speeds) for data encryption and decryption processes. The methodology of this study involved selecting five distinct types of images to compare the outcomes of the techniques evaluated in this research. The assessment focused on processing time and speed parameters, examining visual encoding and decoding using Java as the primary platform. A comparative analysis of several symmetric key ciphers was performed, focusing on handling large datasets. Despite this limitation, comparing different images helped evaluate the techniques' novelty. The results showed that ChaCha20 had the best average time for both encryption and decryption, being over 50% faster than some other algorithms. However, the Twofish algorithm had lower throughput during testing. The paper concludes with findings and suggestions for future improvements.

The lifetime of a battery pack consisting of many cells in series is determined by the weakest cell. Heterogeneous degradation of the battery cell, as a result of inherent discrepancy between cells, accelerates the aging of the weakest cell, which cannot be overcome by conventional passive and active balancing techniques. Therefore, this paper presents a methodology for charging series-reconfigurable Lithium-ion battery packs. To mitigate the negative effects of unregulated temperature increases, thermal gradients, state-of-charge imbalances, and other cell-tocell variations, we formulate and evaluate a charging strategy that addresses temperature and charge homogenization and temperature tracking. Model Predictive Control (MPC) was used to optimally homogenize the charge and temperature in the reconfigurable battery. Comparative simulations were carried out on the passive and the proposed reconfigurable battery, showing the superior perormance of reconfigurable battery dealing with battery cells with different aging conditions. Moreover, the scalability of the proposed method is verified on a string of 100 serial cells with low computational time. Finally experiments were conducted for four different case studies to verify the proposed MPC-based charge and temperature homogenization and control.

Stability of multi-vendor power electronics-dominated grids requires proper specifications for converters interoperability, e.g. tuning boundaries for the converters' control loops to be respected by all vendors. The confidentiality and variety of implementation of the converters' control system by different vendors and the nonlinear dynamics of power converters make this task arduous. This paper formulates this research question as a robust stability problem, embedding frequency-dependent uncertainties in the converter control systems and analytically quantifying the amount of uncertainty able to warranty the power grid stability. Grid following converters are considered and modeled with a nonlinear grey-box approach, in contrast with conventional small-signal black-box models which suffer from inaccuracy when their operating point deviates from the nominal one. This paper also demonstrates both analytically and with Hardware-In-the-Loop (HIL) experiments that different converters operating points (e.g. active power injections, ac bus voltages) radically affect their stability, empathizing the necessity of nonlinear converters models to realize accurate power grid stability analyses.

A document by Chibuike Onuoha. Click on the document to view its contents.

In today's Agile and DevOps-driven software development landscape, the need for rapid and accurate bug reporting is more critical than ever. This paper presents a nextgeneration automation tool powered by large language models and machine learning, aimed at innovating the bug reporting process. The tool automates every phase of bug reporting, from failure detection to severity assessment, duplicate detection, and report generation. By addressing the limitations of manual bug reporting such as inconsistency, scalability challenges, and time inefficiencies, the proposed solution enhances the software testing workflow. Initial findings demonstrate significant time savings, reduced manual errors, and improved collaboration between testers and developers. This work establishes a foundation for fully automated bug reporting, poised to accelerate software development cycles while maintaining high-quality standards.

This paper deals with a three-dimensional multiple input, multiple output (MIMO) passive coherent location (PCL) system in a scenario with multiple transmitters and receivers. The goal is to accurately estimate the location of a single observed target despite challenges such as undetected measurements and outliers due to non-line-of-sight effects. A volumetric search method is proposed to overcome outlier presence. This method, which divides the search region into cuboids, offers a fresh perspective on identifying and discarding outliers while selecting consistent measurements. The centroid of the selected cuboid provides an alternative accurate location estimate, which is particularly valuable when many outliers are present. We also introduce an iterative approach to outlier removal based on model-mismatch cost functions for benchmarking. The selected consistent measurements are used to feed a nonlinearly constrained least squares (NLCLS) algorithm explicitly designed for MIMO PCL systems. A unified mathematical framework is established for this NLCLS approach, which accounts for the nonlinear relationships among the optimization variables via constraints. Additionally, a simplified and efficient version of the NLCLS algorithm is proposed. This streamlined approach iteratively minimizes the model mismatch by using a single constraint, ensuring accurate estimation with reduced computational complexity. The proposed NLCLS method and its simplified version are compared against unconstrained counterparts, namely spherical interpolation and spherical intersection. The Cramer-Rao lower bound is derived for this application, and numerical experiments highlight the robustness of the cuboid-based volumetric search and the nonlinearly constrained algorithms.

Automotive energy management systems (EMS) are advancing towards comprehensive lifecycle-intelligence, spanning from product development phase to customer usage phase. This paper proposes a bi-level transfer approach with model-agnostic meta-learning (MAML) to realize cross-platform transferable and online-adaptive EMS. During the development level, MAML is adopted to calibrate the heuristic control maps of an instantaneous optimization-based EMS, with a per-unit state and action space design facilitating knowledge transfer. Leveraging user data, online adaptation is carried out with reliance on digital twin-based cloud computing, updating onboard controller parameters. The effectiveness of the proposed lifecycle intelligent EMS is validated through real vehicle experiments. Firstly, the entire MAML-assisted V-cycle development process is finished to validate the optimality and knowledge transfer of the EMS, resulting in zero-shot transfer for EMS calibration on new vehicle products. Additionally, real vehicle experimental tests show that a correction of 8.0%~9.5% fuel economy is improved against the convention reinforcement learning-based EMS during usage via online-adaptation, effectively bridging the optimality gap between the control policy learned in development and the global optimal control in actual driving scenarios.

IntroductionEffective data curation is crucial for ensuring findability, accessibility, interoperability and reusability (FAIR) of data. Despite Data Curation Network’s CURATED framework integrating FAIR principles, there remains a lack of a standardized method for evaluating and enhancing data curation activities across organizations, hindering the ability to benchmark practices, identify areas for improvement, and ultimately, elevate the quality of data management.MethodsThe Data Curation Maturity Model (DCMM) was developed using a focus area maturity model framework, applying a mathematical formalization to define a structured matrix of maturity levels and capabilities. The model divides data curation into five focus areas, each characterized by a set of capabilities structured in a matrix format. Capabilities within each focus area are mapped onto a maturity scale using an ordered set, allowing for a systematic assessment and progression tracking.ResultsThe DCMM categorizes data curation activities into five progressive maturity levels: Initial (0-13 points), Managed (14-26 points), Defined (27-39 points), Quantitatively Managed (40-52 points), and Optimizing (53-66 points). To operationalize the model, a Data Curator Log Template employing a three-point Likert scale (0 = Not Done, 1 = Partially Done, 2 = Fully Done) is provided. This template is an essential tool for evaluating the maturity of data curation activities, enabling organizations to track and quantify improvements in a structured and consistent manner.DiscussionThe Data Curation Maturity Model (DCMM) enhances the effectiveness of data curation activities with its innovative, mathematically quantifiable matrix approach, specifically tailored to meet the dynamic needs of data curation. By integrating focus area maturity models with specific curation requirements, the DCMM provides a structured progression framework that allows organizations to visualize and measure their advancements in data management. This model not only elevates data quality and reproducibility across various research domains but also establishes a new standard for strategic, targeted improvement in data curation practices.

As autonomous vehicles (AVs) advance, the integration of Large (Vision) Language Models (L(V)LMs) has emerged as a promising approach to enhance AV capabilities in perception, planning, decision-making, and data generation. However, the practical challenges of incorporating L(V)LMs into AV systems, including computational efficiency, real-time processing, and ethical considerations, remain underexplored. This survey aims to provide a comprehensive review of the current research on L(V)LM applications in AVs, focusing on four key areas: modular integration, end-to-end integration, data generation, and evaluation platforms. We systematically analyse 62 recent papers published before June 2024, detailing methodologies, models, datasets, and benchmarks used, while identifying common tradeoffs and limitations. Our findings highlight the potential of L(V)LMs to improve AV system performance but emphasise the need for further research in real-world integration, regulatory challenges, and V2X communication. This survey offers valuable insights and guidance for researchers and practitioners aiming to optimise L(V)LMs in autonomous vehicles.

Magnetic soft robots offer unique advantages in biomedical applications, particularly in drug delivery and minimally invasive procedures, due to their untethered and fast actuation, and physical adaptability. However, real-time control of magnetic soft robots is challenging because of non-linear dynamics and complex interactions with external magnetic fields. We thus propose a novel, model-free approach to controlling magnetic soft robots based on Proximal Policy Optimization (PPO) combined with Random Network Distillation (RND), integrated with a Convolutional Encoder-Decoder Network (CEDN). The control system autonomously generates action-state pairs to train the CEDN, which predicts control inputs for desired state transitions by abstracting low-level dynamic behaviors. This combination resulted in a real-time control of the robot without necessitating prior knowledge of the robot's geometric design, magnetic or dynamic profile. Experimental validation proves the control system's effectiveness in complex tasks, including successful multi-target navigation and obstacle avoidance. The control system was evaluated using an H-shaped magnetic soft robot. This method addresses the limitations of traditional control methods for magnetic soft robots, offering a practical and rapid solution for controlling magnetic soft robots in real-world biomedical applications potentially enabling practical applications of these robots. This work has significant implications for healthcare industry such as targeted drug delivery, minimally invasive procedures, and on-chip tissue manipulation.

In the computer world, ransomware continues to evolve as a significant threat to cybersecurity, employing increasingly sophisticated techniques to evade detection and cause widespread disruption across industries. The novel approach presented in this research leverages opcode sequences combined with machine learning to address the limitations of traditional static analysis methods, offering a more robust and accurate detection mechanism. Through the use of t-SNE for dimensionality reduction and SVM for classification, this method efficiently handles high-dimensional opcode data, allowing for precise identification of ransomware samples. The experimental results show substantial improvements in detection accuracy, precision, and F1-score compared to baseline models, while demonstrating the importance of feature reduction in optimizing performance. Despite the computational overhead introduced through t-SNE, the overall findings indicate that the proposed framework enhances the static detection of ransomware, making it a valuable tool for cybersecurity efforts aimed at mitigating evolving threats.

Coastal regions are extremely vulnerable to storms and tropical cyclones, which have caused significant economic costs and numerous fatalities. This underscores the urgent need for action to protect the sustainability and resilience of coastal communities. In this paper, we present an AI-driven geospatial analysis pipeline for automating coastal disaster assessment by detecting building damage from satellite imagery. Firstly, we propose an effective and scalable pipeline to train an artificial intelligence (AI) model for damaged building detection using a limited dataset. Specifically, we pretrain our AI model with Microsoft's Building Footprint dataset to adapt to Puerto Rico's landscape, then fine-tune it using a mix of manually and self-annotated data. Upon training, we used our AI model to generate geospatial heatmaps of damaged building counts and damage ratio, which are useful to assess the storm damage and coastal vulnerability. Our approach enhances disaster response efficiency and supports better planning and resilience strategies. Additionally, it can significantly reduce labour costs associated with manual ground surveys for insurance claims and post-disaster socioeconomic impact evaluation. Our solution is one of the Top 10 Global Semi-Finalists in 2024 EY Open Science Data Challenge.

Large Language Models (LLMs), despite their impressive performance in general domains, face challenges in precise domain-specific question answering due to the absence of a built-in training set for niche knowledge and issues related to "hallucinations." This makes it essential to incorporate domain-specific knowledge bases as reliable sources of information. This paper introduces a novel method for domain-specific knowledge-based question answering called KGRA, which enhances Retrieval-Augmented Generation (RAG) by integrating knowledge graphs (KGs) to address cold-start issues and improve accuracy in domain-specific queries. The approach starts by using a closed-source LLM for data distillation, where data augmentation is employed to create a dataset focused on domainspecific entity extraction. The LLM is then fine-tuned and combined with an entity enhancement strategy to extract key terms effectively within the context of a query. Next, a heuristic classifier is used to derive keyword-driven subgraph query and re-ranking strategies, enabling effective interaction with the knowledge graph through query language. Finally, prompt engineering is utilized to generate the final response. The research is contextualized within the scenario of the graduate admissions office at Xi'an Jiaotong University, and an intelligent customer service system was developed as a proof of concept. Our model achieved the best performance on a meticulously crafted domainspecific test set, attaining an F1 score of 95% in keyword extraction, surpassing even GPT-4. Compared to traditional RAG, KGRA showed improvements across all metrics in endto-end question answering, with a notable 23.9% enhancement in its ability to reject inappropriate or incorrect question.

A document by Muhammad Wisal. Click on the document to view its contents.

Power quality phenomena, especially voltage sags, characterized by short-duration reductions in RMS voltage, pose significant challenges to power systems. This study investigates and validates different faults-induced balanced and imbalanced voltage sags at comparative magnitudes of 50% and 30% with a frequency of 50Hz in a multilevel industrial distribution system. The sag magnitudes and phase angle jumps are calculated and further validated through simulations based on phase and RMS voltage drops, sag-type transformations, and sequence component propagation. Fast Fourier transform (FFT) analysis is employed to investigate the magnitudes and propagation of sequence components across different distribution levels. Based on fault induction, the results reveal the transformation of different types of sags or the conversion to a new distorted version of the original type of sag with a unique matrix. The zero-sequence component for each type of sag is removed after passing through a ∆Y configured transformer with a Y-connected load. Moreover, a comparison shows the critical role of different sag magnitudes on phase angle jumps. At both 50% and 30% magnitudes, the system becomes more vulnerable during a phase-to-phase fault. Based solely on phase angle jumps, the sag types D and F are more susceptible at 30% magnitude, while the types C and G are more susceptible at 50%.