External force sensing is essential for enabling robot manipulators to perform force control tasks in industrial applications. However, the use of force sensors increases costs and requires additional calibration tasks, such as gravity compensation. This study proposes a sensor-less force estimation strategy to develop a virtual force estimator for n-link industrial robot manipulators. A model-based identification scheme is used to determine dynamic parameters and construct an inverse dynamics model for external force estimation. Neural learning-based semi-parametric friction compensation model further enhances joint friction estimation, mitigating uncertainties during low-speed motion. The proposed method is validated through experiments on the industrial robot manipulator HIWIN RT605–710. A surface interaction experiment demonstrated a mean absolute error of 4.7388 N of contact force estimation in the Z-direction during a 12-s contact process with an applied force of approximately 40 N using the proposed force estimator.
This article presents the development of a database and an initial analysis aimed at detecting potential correlations between mobile phone usage and frailty levels among elderly individuals. The primary objective is to conduct exhaustive monitoring of all information provided by smartphones, focusing on both the use of common applications and sensor data related to mobility during activities of daily living (ADLs), and to contrast these findings with assessed frailty levels. This work forms part of a broader study involving elderly participants aged between 73 and 96 years, based in the Barcelona area. The database constructed for this study is based on two main strategies. First, to gather comprehensive data on smartphone usage, a group of 10 elderly individuals was observed under normal daily living conditions over a period of up to four weeks. For this initial group, key geriatric indices—Barthel Index, Lawton–Brody Instrumental Activities of Daily Living Scale, and Frail-VIG Index—were previously determined through ambulatory medical evaluations. Second, the database was expanded by generating synthetic data using Generative Adversarial Networks (GANs) and Gaussian Copula models, incorporating an additional group of real patients for whom frailty indices were already known. For those patients, we do not have any data from their own smartphones. The combined real and synthetic dataset was used to identify the most significant features and to support feature selection for classification. Generalized Linear Models (GLMs) were subsequently applied to develop decision trees, with the goal of establishing correlations between smartphone usage patterns and different frailty indices. Finally, the model outputs were analysed to extract meaningful insights that could inform future clinical applications.
In recent years, the development of blockchain has been flourishing.Blockchain provides decentralization, improved security, and increased transparency compared to conventional centralized transaction systems, which lowers costs and increases transaction efficiency. One frequently used tool in blockchain is the Rank-1 Constraint System (R1CS).By providing a more condensed method of achieving zeroknowledge proofs, R1CS improves the blockchain's scalability and security even further. The goal of this work is the standardization of R1CS so that any combination of constraints can be transformed into one and uniform format. Additionally, it makes it possible to combine seemingly different but fundamentally similar sets of R1CS constraints into a single new R1CS set, which eliminates duplicate R1CS sets in the process and increases the efficiency of later steps of the creation of zero-knowledge proofs. This study proposes a concise method for standardizing R1CS constraint sets. Given any R1CS constraint set (comprising three matrices A, B, C, the solution vector and their corresponding parameter vector), regardless of whether the original problem is represented by a single variable or multiple variables, it can be converted into a unique standard representation. The study utilized symbolic and equation operations to determine the actual problem to be solved and merges various potentially equivalent R1CS constraint sets into one. Additionally, we use a flow chart to illustrate our process of standardizing R1CS constraint sets. In the experiment section, each experimental group engaged in nine different equivalent transformations. The results showed that all could be successfully converted into a unique R1CS constraint set representation. Compared to previous related research, the time required for R1CS standardization was significantly reduced.
The use of autonomous robots is rapidly increasing in industries, requiring advanced navigation strategies to ensure safe operations. SWARM robotics, inspired by natural swarm intelligence, faces critical challenges in collision detection and avoidance. Traditional rule-based methods often struggle in dynamic environments, whereas neural network-based approaches provide adaptive solutions. This paper presents a CNN-based collision detection and avoidance system utilizing LiDAR data for SWARM multi-robot systems. The proposed model was trained and tested using simulated data in the Gazebo and ROS 2 environments. A comprehensive description of the data set, a detailed model architecture, training parameters, and evaluation metrics are provided in this paper. The results show a precision of 85.65
Artificial Intelligence (AI) is evolving rapidly, and Knowledge Graphs (KGs) are widely used in tasks such as smart question answering, recommendation, and search engine optimization due to their semantic search, reasoning, and knowledge aggregation capabilities. However, in industrial applications with complex and heterogeneous data, the quality of KGs suffers, leading to data inconsistency and decision-making errors, which poses a challenge to intelligent manufacturing. We propose an adversarial learning-based error detection method for industrial knowledge graphs (GAN-MKED). This method extracts features from multimodal industrial data including documents, images, and sensor data using a fine-tuned BERT model for document encoding, SENet for image feature extraction, and a combined BiLSTM and FFT approach for sensor signals. A multimodal attention mechanism fuses these features, while a relation-aware graph attention network (R-GAT) aggregates neighborhood information. Spectral Normalization GAN (SN-GAN) introduces random noise to generate pseudo-neighborhood features, improving robustness and stability. The difference between real and generated features is computed and used within an adaptive thresholding mechanism to detect errors effectively. Experimental results on a self-constructed industrial KGs dataset demonstrate that GAN-MKED achieves high accuracy and robustness in error detection for industrial knowledge graphs.
A novel strategy to reduce waiting times by seamlessly integrating traffic flow prediction and (deep) reinforcement learning has been proposed in this paper. Employing an innovative reward function combined with a long short-term memory (LSTM) based forecasting module, future road densities, vehicle speeds, and traffic volume can be accurately anticipated, guiding the dynamic control of traffic light duration. The agent’s action selection aligns with environmental conditions in our framework, contributing to a more comprehensive approach to traffic management. Simulation assessments performed within the SUMO, i.e., Simulation of Urban Mobility, platform validate the effectiveness of our proposed strategy. Compared to the most competitive strategies in the literature, our approach can reduce travel time and fuel consumption by up to 18.1
Dengue fever remains a significant public health challenge globally, necessitating effective surveillance of key data such as dates, locations, and case counts. This study evaluates the potential of large language models and com-pares their performance with deep learning methods in predicting epidemiological data from dengue fever case reports. Although initially hypothesized that large language models would outperform due to their advanced capabilities, the empirical results indicated that deep learning approaches were more precise. Specifically, deep learning models, such as TextCNN and LSTM, achieved superior F1-scores of 0.998 and 0.997, respectively, compared to large language models, which achieved 0.882 for GPT-4 Zero-shot and 0.634 for Few-shot. Despite this, the rapid deployment and scalability of LLMs highlight their potential utility in real-time public health emergencies. The findings suggest that while deep learning methods currently outperform large language models in precision, targeted fine-tuning could enhance the utility of large language models in public health surveillance. This research supports the continued development of AI technologies to improve global health responses, leveraging both deep learning and large language model capabilities for effective disease management.
Worldwide, type 2 diabetes mellitus (T2DM) prevalence is rising, and macrovascular complications remain its primary cause of death. Although glycemic control alleviates microvascular damage, risk factors for macrovascular complications remain unclear. Recent findings indicate that models built on realworld data outperform those based solely on clinical trials in forecasting T2DM outcomes. Using electronic medical records, we constructed XGBoost, decision trees, and random forest models to predict four macrovascular complications in T2DM. XGBoost achieved the highest accuracies-0.985 for hypertensive heart disease, 0.960 for ischemic heart disease, 0.956 for cerebrovascular disease, and 0.969 for coronary heart disease. These results can be incorporated into clinical workflows, streamlining information exchange for healthcare providers and patients.
Navigating harsh environments requires innovative approaches to path planning to ensure robustness and accuracy amidst complex obstacles and demanding conditions. This study presents a hybrid algorithm combining the Whale Optimization Algorithm (WOA) and Dynamic Multi-Swarm Global Particle Swarm Optimization (DMSGPSO) to address these challenges effectively. The hybrid algorithm leverages WOA’s wide exploration capabilities, inspired by humpback whales’ bubble-net hunting behavior, and DMSGPSO’s adaptive balancing of global and local search. This integration enhances the algorithm’s ability to avoid local optima and achieve precise path optimization. Experimental results on a 700 × 700 grid environment with various static obstacles demonstrate that the proposed method reduces path cost by 2.11
With the rapid growth of IoT, more devices are connecting to the internet, increasing data exchange and emphasizing the need for robust security. Future IoT devices, such as embedded systems using 5G and NB-IoT, may operate without traditional safeguards like firewalls or VPNs, making Zero Trust Network Access (ZTNA) a crucial solution for secure data transmission. This study implements a ZTNA-compatible security framework for IoT endpoints, preventing unauthorized access and ensuring data integrity. The proposed system integrates Linear Search (LS), Hash Search (HS), and an optimized Hash then Linear Search (HLS) method, comparing their efficiency in generating non-repeating random numbers to identify the best approach for IoT transmission. The results demonstrate improved data security and transmission efficiency, providing a practical ZTNA-compliant solution for IoT device communication.
Missing values in medical data present significant challenges to patient care, particularly in critical settings like the Intensive Care Unit (ICU), where accurate and timely decision-making is vital. Continuous monitoring of vital signs, such as heart rate and blood pressure, is essential for effective ICU patient management. Addressing missing values in these variables is crucial, as it provides clinicians with enhanced information for informed decisions. This paper introduces MediTHIM, a high-performance and efficient approach for medical data imputation, offering fast computation with minimal resource requirements, while relying only on past existing data. After carefully preprocessing the MIMIC-IV dataset, a recent medical dataset containing information from ICU patients including vital signs, demographic information, physical characteristics, and administered medication, we carried out a comprehensive set of experiments. Our results show that MediTHIM has the best performance while exhibiting low computation times and requiring few resources when compared against multiple competitive approaches that do not use future data. Moreover, MediTHIM also shows a stable performance with minimal fluctuations. These results were consistent across different missingness patterns analyzed and for missing percentages ranging from 1
Ransomware attacks are becoming increasingly complex and sophisticated, posing significant challenges to traditional detection methods. This study focuses on Windows Portable Executable (PE) files and leverages dynamic analysis with Cuckoo Sandbox to generate execution reports, identifying critical features for ransomware detection through API call sequences. To address the rapidly evolving ransomware landscape, we propose a deep learning-based detection model that employs two numerical representation methods for processing API calls: direct conversion and a novel quadruple-based semantic representation, which effectively captures deeper relationships between API calls. Our approach integrates embedding and convolutional techniques to extract semantic relationships and structural patterns from API call sequences, enhancing the model’s ability to recognize ransomware behaviors with high precision. Additionally, the Transformer model is employed to analyze sequence dependencies, further improving detection accuracy. The proposed model demonstrates remarkable performance, achieving an accuracy of 99.94 https://github.com/stwater20/RANsomCheck ).
This study addresses the challenge of detecting helmet violations among motorcyclists by proposing a novel approach based on the BF-YOLOv7 model, which enhances the accuracy of video analytics. Additionally, we incorporate the PRB-FPN6-MSP model to further refine detection results. Our experiments, conducted on the 2024 AI City Challenge Track 5 benchmark dataset, demonstrate the effectiveness of our approach. The results indicate that our method performed exceptionally well across 100 test videos, achieving rankings of 11th and 16th on the public leaderboard, outperforming 43 competing teams.
Detecting sarcasm in social media content has been one of the challenging and interesting tasks in social network analytics due to users' growing creativity, humor, playfulness, and the complexity of the input data. Today, social media users can employ various methods to convey their thoughts, information, or content uniquely, combining captions (text) and images in complex and nuanced ways. Most existing sarcasm detection models are limited because they focus solely on single modalities (text or image). However, most social media content includes captions and pictures created by users. This paper proposes three approaches to detecting sarcasm from various perspectives. We leverage the relationship between textual and imaging data and utilize state-of-the-art, pre-trained large language models for extracting relevant features and improving the performance of the main problem with labels categorized into four types: multi-sarcasm, non-sarcasm, text-sarcasm, and imagesarcasm. We also provide a deeper level of analysis and discuss potential challenges for more accurate sarcasm detection. This work is one of the attempts to address the multiclass classification of sarcasm detection on Vietnamese social platforms.
In the digital transformation era, subscription-based digital services have revolutionized how we access and consume services. Digital music platforms offer vast online music libraries and personalized listening experiences. This digital revolution has transformed consumer habits and reshaped the music industry ecosystem. The traditional model relying on physical record sales has shifted to digital distribution, enabling music creators to reach global audiences more effectively. Independent creators can compete equally through data analytics for precise marketing and social network dissemination. Digital platforms break geographical barriers, creating new opportunities for music discovery and cultural exchange, allowing users to explore global music and experience more diverse musical cultures. This study employs DAA (Desire & Accessibility Analysis) to identify the position of various dimensions and uses NRM (Network Relation Map) analysis to determine the interrelationships between these dimensions. The research proposes optimal development paths to address different dimensions (desire and accessibility). It implements a common improvement method to resolve shared improvement paths.
The success of software projects depends on well-defined requirements. However, specifying these requirements is a time-intensive task that requires domain expertise. Poorly defined requirements often lead to project failures, cost overruns, and inefficiencies. Large language models (LLMs) have the potential to address these challenges by assisting in requirements engineering (RE). This paper explores the capabilities of LLMs in RE, focusing on two key aspects: (i) generating coherent and contextually relevant software requirements for a given project scope and (ii) assessing their applicability in real-world projects. By comparing LLM-generated requirements with real-world requirement samples, we evaluate their relevance and utility. Our findings suggest that LLMs can enhance the efficiency of requirements engineering by generating relevant and useful requirements.
In recent years, some score functions (SFs) for ranking interval-valued intuitionistic fuzzy values (IVIFVs) have been developed. However, the drawbacks of the existing SFs of IVIFVs are that they cannot differentiate the ranking orders of IVIFVs in some situations. In this paper, we propose a novel SF of IVIFVs to conquer the shortcomings of the existing SFs of IVIFVs. The proposed SF of IVIFVs is very useful for ranking IVIFVs.
Leukemia, one of the most frequent cancers affecting all age groups, is defined by the rapid development of abnormal blood cells. With the current trend of personalized medicine, predicting the risk of leukemia based on gene expression is a necessary solution for supporting each patient’s treatment plan. This study evaluates the effects of the T-test-based feature selection technique on improving the performance of machine learning in acute leukemia classification tasks on Deoxyribonucleic acid (DNA) microarrays gene expression data. First, a data pre-processing process is deployed by conducting a data cleaning step and performing a data normalization step with StandarScaler. Next, the feature selection process is employed using the T-test-based feature selection technique to reduce the dimensionality of the gene expression data by choosing the most important genes. Finally, various machine learning models are applied, namely Naïve Bayes, Logistic Regression, Support Vector Machine, K-Nearest Neighbors, Decision Tree, and Random Forest, to classify acute leukemia. The experimental results show that the proposed method achieves better performance than the state-of-the-art studies for leukemia classification on Deoxyribonucleic acid (DNA) microarray gene expression data.
Detecting and localizing faults in building systems like heating or cooling is essential to reduce energy consumption. Hence, there is a need for innovative diagnosis applications. In this paper, we report on an initial study to show the potential of model-based diagnosis in the building domain. We present a model for heating systems and discuss the results of the obtained diagnosis, considering the expected behavior of different case studies. The approach provided reasonable diagnoses depending on available observations. In addition, the same component models can be easily re-used without adaptations. Hence, applying model-based diagnosis for building systems is a feasible alternative to other approaches.
The proliferation of multi-robot systems has significantly advanced the capabilities and applications of robotics in various fields. Efficiently managing and as-signing tasks to multiple robots remains a critical challenge, impacting overall system performance and operational efficiency. This research proposes a cloud-based task assignment strategy leveraging an improved Periodic Min-Max Weight (PMW) algorithm. By integrating cloud computing resources, we aim to enhance the collaboration and coordination among robots, optimizing task distribution and completion. The proposed model demonstrates a scalable and robust solution to the Multi-Robot Task Assignment Problem (MRTA), promising significant improvements in efficiency and adaptability.
