Browse Items (16481 total)
Sort by:
-
Real-Time Fabric Defect Detection Using a Lightweight YOLOv8 Model on Edge Devices
The detection of defects in fabric is a critical process for maintaining quality standards and reducing economic losses in the textile industry. Traditional inspection methods, which rely on human operators, are often slow, inconsistent, and susceptible to error. This research introduces an innovative solution that harnesses Edge AI and deep learning to facilitate real-time, on-site defect detection. We developed a highly efficient and lightweight model based on the YOLOv8 architecture, specifically tailored for deployment on resource-constrained edge devices like NVIDIA Jetson Nano or Raspberry Pi. Through a process of comprehensive literature analysis and domain expertise, a compact, high-precision model was trained on diverse fabric defect datasets. To ensure optimal performance on edge hardware, we employed advanced optimization techniques like quantization and pruning. The primary offering of the work are threefold: the making of a streamlined YOLOv8-based model for fabric defect detection, a comparative analysis of various edge inference strategies, and a proposed system architecture for real-time embedded deployment. This study effectively demonstrates the practical application of advanced AI to solve longstanding challenges in textile quality control. Future efforts will be directed towards extensive real-world operational testing and exploring localized Model Training with Federated Learning enhancement. 2025 IEEE. -
The Executive Copilot: Investigating the Impact of Generative AI Assistants on Strategic Decision-Making, Leadership Efficiency, and Managerial Creativity
The generative artificial intelligence (AI) revolution has given rise to a class of technology - executive copilots that are designed to unclog the brains of overworked managers in the 21st century. This paper examines the effects of generative AI assistants on three intertwined facets of executive performance - strategic decision making, leadership effectiveness, and managerial creativity. Early literature suggests that AI systems could automate administrative tasks and offer data-driven insights; however, we still lack an understanding of their comprehensive impact on executive tasks and outcomes. To bridge this gap, the study uses a multimethod research design, drawing on survey data from 350 senior executives from various firms, controlled experimental studies of strategic problem solving, and in-depth case analysis of firms in finance, consulting, and technology sectors. Results suggest that AI copilots contribute significantly to faster decision-making and creative output by helping to synthesize information and sparking divergent thinking rapidly. Consistent with this, results reveal that a potential 'caveat' of overreliance on AI-generated recommendations might lead to less scrutiny and thus the fragility of the high-stakes decision-making in practice. Theoretically, the study extends the leadership augmentation theory to the context of generative AI. It extends the conceptualization of the executive copilot in terms of a complementary agent that adjusts the boundary of human-machine collaboration. Pragmatically, it has implications for executive education, corporate governance, and the design of responsible AI uptake strategies. Through methodical analysis of opportunities and risks, this study offers a nuanced view of how executive copilots change leadership practice in the era of intelligent augmentation. 2025 IEEE. -
Adaptive Communication Protocols for Manager-Worker Small LLM Multi-Agent Systems in Resource-Constrained Environments
The proliferation of Internet of Things (IoT) and edge computing paradigms has necessitated the deployment of intelligent multi-agent systems in resource-constrained environments, where traditional communication protocols fail to optimize bandwidth utilization and energy consumption. This paper presents a novel adaptive communication framework specifically designed for manager-worker small Large Language Models (LLMs) multi-agent systems operating under stringent computational, memory, and energy constraints. Our approach integrates three synergistic innovations: (1) a lightweight semantic filtering module employing knowledge-distilled small LLMs (DistilBERT 66M parameters, TinyBERT 4.4M parameters) for real-time extraction of task-relevant information with minimal computational overhead, (2) a dynamic hierarchical coordination scheme enabling runtime role reassignment based on evolving task complexity and resource availability, and (3) an adaptive topology control mechanism leveraging algebraic connectivity measures to optimize network robustness while minimizing communication overhead. Comprehensive simulation-based evaluation across five distinct IoT deployment scenarios demonstrates substantial improvements in communication efficiency, achieving an average token reduction of 5 7. 4 % (range: 5 4. 1 - 5 9. 4 %) while maintaining task completion rates within 8.3 percentage points of baseline performance. The framework exhibits superior coordination quality improvements of 6.0 %, coupled with significant resource optimization including 5 1. 2 % CPU usage reduction and 50.4 % energy savings, validating its practical suitability for edge computing deployments in resource-critical applications. 2025 IEEE. -
Evaluating Building Damage Classification Accuracy: A Benchmarking Study of UNet
Building damage classification must be done accurately and quickly in order to support disaster response and recovery activities. Deep learning models, particularly U-Net, have demonstrated strong potential in automating damage assessment from satellite and aerial imagery. This study benchmarks the accuracy of U-Net in classifying building damage across multiple datasets, evaluating its performance against ground truth labels. Key factors such as data preprocessing, augmentation techniques, and model variations are analyzed to determine their impact on classification accuracy. The results provide insights into the strengths and limitations of variations in U-Net for damage assessment, highlighting areas for improvement and future research directions 2025 IEEE. -
Lalasa Quantum Computing Method: A Unique Quantum Convolutional Neural Network Architecture
A novel Lalasa Quantum Computing Method architecture is presented in this paper for classification of medical images, aiming to enable efficient early detection of cancer. The proposed framework integrates a custom preprocessing pipeline that removes specular reflections using a SWIN Transformer and segments regions of interest via an Enhanced Gaussian Mixture Model. The quantum classification module employs amplitude encoding to map classical image data into quantum states, enabling structured feature extraction through a sequence of quantum convolutional layers with trainable variational circuits. The model was implemented using IBM Qiskit and trained on the publicly available Intel & Mobile ODT Cervical Cancer Screening dataset. Experimental results show a high overall classification accuracy of 98.58%, with moderate performance on class-specific F1-scores, recall, and precision. These results demonstrate the feasibility and effectiveness of quantum-classical hybrid models for medical image analysis, particularly in high-dimensional, low-sample scenarios. The study sets stage for future advancements in quantum machine learning applications in healthcare, with potential extensions involving real quantum hardware deployment and multiclass classification improvements. 2025 IEEE. -
Detecting Student Depression Using Non-Clinical Measures with Explainable Predictive Modeling
Depression among students is a serious global mental health concern, affecting academic performance, emotional wellbeing, and long-term development. While traditional diagnostic tools like self-reported questionnaires and clinical interviews are useful, they often suffer from subjectivity, recall bias, and limited scalability. This study introduces a data driven, interpretable machine learning approach to predict student depression using both academic and nonacademic factors, without relying on clinical indicators. The dataset comprises student information on demographics, academic workload, lifestyle habits, social interactions, financial stress, and emotional state. Following thorough preprocessing including handling missing values, encoding variables, correlation-based feature reduction, and SMOTE to address class imbalance, ten supervised machine learning models were trained and assessed. Among them, a SMOTE enhanced XGBoost model achieved the highest test ROC AUC score of 0.95. To maintain transparency, SHAP (Shapley Additive Explanations) was employed to interpret the model's predictions, highlighting key risk factors such as academic pressure, poor sleep quality, financial difficulties, and low social support. These findings can help guide early interventions and build trust with stakeholders. Future work may involve incorporating longitudinal and multimodal data, deploying real time solutions in educational settings, and addressing ethical considerations around privacy, fairness, and consent in AI based mental health systems. 2025 IEEE. -
Artificial Intelligence and Machine Learning in Financial Fraud Research: A Bibliometric Analysis of Trends and Collaborations
This study conducts a bibliometric analysis of the financial fraud research domain, synthesizing publication metadata to uncover collaboration structures, thematic hotspots, and temporal trends. The research utilizes the bibliometric tool Vos viewer to retrieve the insights on the various aspects of financial fraud detection using the emerging technologies. The study analyses the integration of AI and machine learning techniques on the detection of financial fraud. Using co-authorship networks, we identify influential authors and collaboration clusters, while keyword frequency and co-occurrence patterns reveal core topics and emerging fronts such as anomaly detection, machine learning, insider trading, and regulatory analytics. Temporal profiling of publications highlights growth phases and shifts in emphasis across years and venues. The analysis provides an integrated view of the field's intellectual structure, enabling scholars and practitioners to locate key contributors, benchmark thematic coverage, and identify gaps for future inquiry. Results offer actionable insights for forming research collaborations, prioritizing topics, and designing evidence- based strategies against financial fraud. 2025 IEEE. -
Dynamic Financial Portfolio Optimization Using Temporal Convolutional Networks and Real-Time Data Analysis
This paper presents an integrated framework for AI-driven portfolio optimization combining temporal convolutional networks (TCNs) with conditional value-at-risk (CVaR) minimization. Our system processes real-time market data through an automated pipeline implementing volatility-adjusted feature engineering and walk-forward validation. The architecture employs dilated causal convolutions for temporal pattern extraction combined with Ledoit-Wolf shrinkage covariance estimation for robust portfolio optimization. Experimental results demonstrate an 18.7% annualized return with 22.3% volatility, outperforming traditional mean-variance optimization by 14.2% in risk-adjusted returns. The implementation addresses key challenges in numerical stability and computational efficiency through eigenvalue clamping and gradient checkpointing. 2025 IEEE. -
Graph Convolutional Networks for Predicting Postpartum Depression: A Symptom-Based Analysis
Postpartum Depression (PPD) is a serious mental health condition affecting new mothers and aligns with the United Nations Sustainable Development Goal (SDG) 3: Good Health and Well-being, which stresses early detection and intervention. This research investigates the use of Graph Neural Networks (GNNs)specifically, Graph Convolutional Networks (GCNs)to predict PPD by modeling the interdependencies among symptoms. A preprocessed dataset of 1,503 records was utilized, involving categorical encoding, missing value imputation, and feature standardization to enhance model reliability. The GCN model was built using a K-Nearest Neighbors (KNN)-based graph structure, enabling the network to learn intricate relationships between symptoms. Experimental results showed that the GCN achieved an accuracy of 89%, identifying key predictors such as trouble sleeping, guilt, irritability, difficulty concentrating, and anxiety. The use of SHAP explainability tools further validated these predictors, enhancing interpretability and revealing the models decision-making process. While traditional models like Random Forest achieved higher classification accuracy (95%), GCN offered valuable insights into the underlying structure and relationships among symptoms, supporting its potential in mental health diagnostics. Future work may explore hybrid architectures and larger datasets to further improve the models predictive performance and contribute to AI-driven early screening strategies for PPD. 2025 IEEE. -
Deep Learning Approaches for Detection and Classification of Microplastics in Water for Clean Water Management
Microplastic pollution is a growing environmental concern, threatening aquatic ecosystems and human health. This study presents a dual deep learning approach for microplastic detection and classification using two datasets. For water microplastics, YOLOv8 and YOLOv11 were employed for object detection. InceptionV3, VGG19, ResNet50, ResNet152, DenseNet121, EfficientNetB0, and a custom CNN were applied for classification, classifying three distinct microplastic types in non-aquatic environments. Experimental findings display high accuracy, and indicate the potential of AI-enabled solutions for environmental monitoring. This research contributes to SDG 6 Clean Water and Sanitation, promoting sustainable management of water. 2025 IEEE. -
Attention-Enhanced Vision Transformer Model for Precise Skin Cancer Detection
Skin cancer is one of the most prevalent and potentially fatal diseases, requiring early and accurate detection for effective treatment. Recent advances in deep learning have significantly improved automated skin lesion classification, but traditional Convolutional Neural Networks (CNNs) struggle with capturing long-range dependencies in dermoscopic images. To address this limitation, we propose a Preprocessing-Optimized Vision Transformer (ViT) Model that enhances lesion detection using attention-based feature fusion. Our methodology includes contrast enhancement (CLAHE), hair removal (DullRazor), lesion segmentation (K-Means + Otsus Thresholding), and data augmentation, ensuring robust model training. The proposed Attention-Enhanced ViT Model effectively learns global contextual features from dermoscopic images through self-attention mechanisms. The proposed model is evaluated our model on the ISIC Skin Cancer Dataset, achieving an accuracy of 94.6%, precision of 92.8%, recall of 93.5%, and an AUC-ROC score of 0.97, outperforming traditional CNN-based models such as ResNet50 (92.1% accuracy) and EfficientNet-B0 (93.3% accuracy). Our results demonstrate that integrating preprocessing techniques with Vision Transformers significantly enhances classification performance, making this approach a viable solution for real-world computer-aided dermatology. 2025 IEEE. -
Intelligent Research Summarization: Enhancing Academic Productivity through AI for a More Sustainable Future
The efficiency of researchers is often hindered by the enormous volume of academic research, which hinders the extraction of useful insights in a timely manner. This paper proposes an extension for summarizing research papers using artificial intelligence to ease the process through automated concise summarization. Our system leverages cutting-edge natural language processing tools, including PyPDF2 for text extraction, LangChains text chunking method, FAISS for similarity-based information retrieval, and Googles Gemini AI for high-quality summarization. We evaluated the performance of the model using a variety of metrics, including Rouge scores in conjunction with human judgments. Experimental results show that the suggested approach significantly improves research efficiency, reducing reading time by 60% without compromising high accuracy rates. 2025 IEEE. -
Advancing Astronomical Science - Machine Learning-Based Classification of Variable Stars for Scientific Innovation and Research
Variable star classification is an important part of astrophysics and gives astrophysicists a way of studying stellar evolution, structure and dynamics. Due to the availability of large scale surveys such as Gaia DR3, machine learning techniques are used in automation of the classification process. In this study, RF, SVM, MLP and XGBoost (XGBClassifier) models are evaluated for classification of variable stars. The data set used in this work was collected from Gaia DR3 using Astroquery and the ability of these models is evaluated for different star classes. The result shows that the XGBoost had the best accuracy of 91% compared to RF (89.98%), MLP (88%) and SVM (83%). A comparison of various metrics such as precision, recall and F1-score of each method is also provided to address their strengths and weaknesses. This work further emphasises the need of sophisticated machine learning techniques in astrophysical data analysis and discusses problems of certain kinds of variable star classification. KeywordsGaia, variable stars, Classification, Machine Learning, Random Forest, XGBoost, Multi-Layer Perceptron, Support Vector Machine. 2025 IEEE. -
Fuzzy Logic-AHP Hybrid Model for Faculty Performance Evaluation to Enhance Educational Quality in Higher Education
Guaranteeing equitable and precise evaluation of teacher performance is a continual challenge in higher education, as subjective discrimination, uneven metrics, and absence of cohesive frameworks frequently obstruct informed decision-making. A hybrid Fuzzy Logic-Analytic Hierarchy Process (AHP) model is developed to integrate systematic requirement weighting through AHP with the uncertainty management features of fuzzy logic. The method assesses faculty performance across various dimensions, including classroom effectiveness, research output, service involvement, and professional advancement. The integration guarantees impartiality in criterion weighing and adaptability in managing qualitative assessments, resulting in a balanced and thorough evaluation method. The proposed hybrid framework, in contrast to standard models, reduces subjectivity, improves interpretability, and provides greater accuracy in prediction. Experimental findings indicate that the model attains an Accuracy of 96.8%, Precision of 97.2%, Recall of 96.5%, F1score of 96.8%, and AUC of 0.98, surpassing baseline methods like Decision Trees, Logistic Regression, and Support Vector Machines. These findings confirm the resilience and flexibility of the proposed methodology in practical teacher evaluation contexts. The research enhances educational quality and facilitates the integration of hybrid decision-support systems into institutional policy-making and future academic performance evaluations. 2025 IEEE. -
Financial Behaviour Analysis for Payment Bank Adoption Using Random Forest and PCA: An Indian Perspective
The gradual acceptance of payment banks in India constitutes an important challenge to equitable financial development, especially for underbanked rural communities. The proposed study handles the difficulty by examining financial behaviour through a hybrid machine learning methodology that integrates Principal Component Analysis (PCA) for dimensionality reduction with a Random Forest classifier for predictive modelling. The study utilises datasets obtained from Kaggle that reflect demographic, behavioural, and digital engagement variables. PCA preserves approximately 9 0% of variance while reducing feature complexity, allowing the Random Forest to effectively characterise adoption behaviour. In comparison to conventional classifiers such as Logistic Regression, SVM, and Decision Trees, the suggested model improved performance, attaining 96.7% accuracy, 95.8% precision, 97.1% recall, 96.4% F 1-score, and an AUC-ROC of 0.982. The findings exceed all chosen baselines, demonstrating the system's resilience and reliability. The approach provides behavioural insights essential for policy formulation and strategic engagement by pinpointing the most significant adoption determinants. This research greatly advances the digital banking sector by integrating data science with social impact, providing a clear, high-performing solution to inform financial inclusion policies. It establishes a basis for the development of future real-time and personalised adoption prediction systems utilising advanced AI methodologies. 2025 IEEE. -
Microgrid Energy Management- An Optimization Approach for Operational Cost Reduction
Optimization-based energy management and the related economic viability analysis of replacing a diesel generator alone-based microgrid with a PV, battery, fuel cell and diesel generator-based hybrid microgrid to minimize the operations cost, considering battery degradation and emissions, are presented in this work. The fuel cell considered here is the hydroplus fuel cell, which has minimum emissions, and eliminates the need for hydrogen storage. Mathematical modeling is based on experimentally obtained parameters for fuel consumption coefficients and emissions. The study considers an islanded mode of operation with two different scenarios of microgrid configuration, the first case with the hybrid microgrid working under optimal power conditions, and the second case with only diesel generator available to meet the load demand, for four representative months of the year. Comparing the operations cost under the two scenarios, replacement of a diesel generator-based microgrid, with a PV/battery/hydroplus fuel cell/diesel generator based microgrid resulted in reduction of daily operations cost by 54.06%, 35.25%, 34.38% and 32.71% for the months of January, April, July and October respectively. A sensitivity analysis with varied costs of fuel, operation and maintenance, and battery degradation proved that irrespective of these costs, a considerable amount of reduction in the operations cost is achieved. The results presented here are highly beneficial for application in microgrids worldwide. 2025 IEEE. -
Advance Data Ingestion Framework - Integration, Processing, Transformation, and Loading
The research introduces a new concept known as the Advanced Data Ingestion Framework, which is aimed at enhancing the process of getting into stored information through some intelligent methods like data preprocessing, transformation and loading. By making use of Azure services, the platform considers distributed computing and parallel processing so that structured as well as unstructured data can be incorporated from various origins without any difficulty. To begin with, the proposed framework starts with setting up scalable Azure infrastructure and integrating SAP S4/HANA for secure and efficient data transfer purposes. Within Azure Data Factory the ingestion occurs while Delta Lake ensures proper housekeeping & integrity within the system. It includes creating Power BI dashboards which allow users to see patterns easily and make better decisions based on what they know or can learn. The study brings out the flaws of current data input solutions and emphasizes the urgent requirement for a highly scalable low latency system that can support real time data processing efficiently. It tests the framework under different performance environments showing that it can effectively manage modern data within it. Finally, there is discussion about future improvements such as incorporating more sophisticated analytics or ML models thereby strengthening the decisionmaking process based on available facts. 2025 IEEE. -
An Efficent Deep Learning Framework for Cyberbullying Detection Using DistilBERT and Sentiment Analysis
Particularly because of the complex and changing character of online communication, which hampers conventional detection strategies, the frequency of cyberbullying presents a significant threat to mental health and well-being in the digital age. This article presents a fast deep learning approach to improve cyberbullying detection by combining sentiment analysis with a lightweight transformer model, DistilBERT. This work intends to increase classification performance by using sentiment-based features and using DistilBERT's language and contextual awareness. Unlike conventional approaches and simpler machine learning methods, which can depend on feature extraction techniques like Bag of Words (BOW) or TF-IDF, the proposed model directly leverages contextual embeddings. Moreover, DistilBERT provides a balance between speed and performance unlike deep learning models like CNN, BLSTM, and LSTM, which could suffer with computational efficiency. Experimental results demonstrating remarkable accuracy and recall on many different datasets indicate the effectiveness of our hybrid approach. demonstrating a significant rise in cyberbullying detection over conventional methods, to evaluate performance criteria including computational efficiency, accuracy, and F1-score. With an outstanding 93.7 % accuracy, the proposed model exceeded earlier evaluated methods on this dataset. 2025 IEEE. -
Scribble and Learn-A Leaner Centric Approach
This innovative practice full paper describes a creative learning method that support collaborative, creative visual leaning by improving the happiness index in classroom environment. Teamwork and creativity are limited in traditional classrooms that emphasize individual learning and technology dependence. This study offers 'Scribble and Learn,' a novel instructional tool for Cognitive Psychology and artificial intelligence courses to fill this need. 'Scribble and Learn' uses unique scribbling to encourage active participation and collaboration. Team research with limited technology is done by students. This constraint promotes teamwork and communication as they uncover and study memory performance determinants. Teams use keywords and images to communicate their findings on a huge writing surface after this research phase. This 'scribbling' approach accommodates varied learning styles and stimulates creative knowledge expression, creating a dynamic learning environment beyond lectures. Teams present and discuss their findings in a moderated class discussion. This conversation improves Cognitive Psychology and AI memory performance theory knowledge. 'Scribble and Learn' improves teamwork, creativity, and memory. Positive student feedback shows that the exercise promotes active learning and participation. In a class discussion that is moderated, teams present and discuss the discoveries that they have uncovered. The knowledge of cognitive psychology and artificial intelligence memory performance theory is improved by this discourse. The fact that the activity gets positive feedback from students demonstrates that it encourages active learning and involvement. 2025 IEEE. -
Balancing the Cart: Evaluating Imbalance-Aware Machine-Learning Pipelines for Predicting E-Commerce Purchases
We present a comprehensive investigation into predicting purchase conversions in e-commerce sessions, addressing the challenges of severe class imbalance and complex user behavior signals. Using a real-world dataset of 12,330 user sessions described by 24 features (interaction counts, durations, bounce/exit rates, page values, temporal and device metadata), we first conduct exploratory analysis to reveal seasonal peaks in conversion and strong correlations between page value metrics and purchase likelihood. To mitigate the low positive-class rate (10.8%), we embed SMOTE oversampling within our training pipelines, ensuring balanced learning for all classifiers. We then perform a head-to-head comparison of twelve algorithmsranging from linear and generative methods (Logistic Regression, LDA, Gaussian NB), instance-based learners (KNN, SVM), bagging ensembles (Random Forest, Extra Trees, AdaBoost), gradient boosters (XGBoost, LightGBM, CatBoost), to a feed-forward neural network (MLP). Evaluation on a stratified 80/20 holdout set uses overall accuracy plus precision, recall, and F1-score for the purchase class, alongside ROC AUC. Our results demonstrate that ensemble tree methods dramatically outperform simpler models: LightGBM achieves the highest F1 (0.694) and ROC AUC (0.924), with Extra Trees closely following (F1 0.678, AUC 0.926). Simpler classifiers, despite SMOTE, lag markedly in recall and F1, underscoring the importance of powerful nonlinear learners. These findings establish a new benchmark for imbalance-aware conversion prediction and recommend SMOTE-augmented gradient boosting and randomized tree ensembles as the methods of choice for future research and practical deployments. 2025 IEEE.