The rapid advancement of connected vehicle technologies has intensified the need for efficient spectrum utilization. Cognitive radio (CR) enables dynamic access to underutilized spectrum, with spectrum sensing playing a key role in detecting primary users (PU). This study introduces a novel spectrum sensing approach that integrates filter bank (FB) signal decomposition with convolutional neural networks (CNNs)—referred to as filter bank decomposition and convolutional neural network (FB-CNN)—to enhance detection performance compared to conventional methods. Unlike traditional techniques such as energy detection (ED), the proposed FB-CNN leverages the frequency components extracted by the FB as CNN inputs, enabling robust identification of PU signals across multiple modulation schemes, including BPSK, QAM, FSK, and GMSK. Simulation results demonstrate substantial gains, particularly in low-SNR scenarios: for example, at an SNR of 5 dB, FB-CNN achieves a detection probability of 89% for 2-FSK, compared to only 22% with ED—representing a fourfold improvement. These findings highlight the novelty and effectiveness of FB-CNN in significantly improving spectrum sensing reliability for connected vehicle networks operating in challenging signal environments.
Existing arbitrage detection techniques rely on exhaustive search or linear programming, which are computationally expensive and often miss profitable cycles in dynamic markets. Triangular arbitrage is a profitable trading strategy that exploits discrepancies in currency exchange rates, but common algorithms detect only a limited number of loops and cannot find non-loop opportunities. To address these gaps, this study presents a realtime, graph-based framework for identifying triangular arbitrage opportunities in cryptocurrency markets using an optimized implementation of the Bellman–Ford algorithm. By modeling currency exchange rates as a directed graph and detecting negative-weight cycles, the framework efficiently identifies profitable arbitrage opportunities under realistic trading conditions. The proposed framework achieves an average detection latency of 0.002 milliseconds, providing empirical performance benchmarks for single-exchange cryptocurrency trading systems. Experiments on a six-month historical dataset yielded a detection accuracy of 92%, while additional validation on live cryptocurrency market data streams confirmed the framework’s real-time performance and low latency. This high-speed detection is crucial in high-frequency trading (HFT), where brief pricing inefficiencies can yield significant profits before being corrected. The experimental pipeline is designed to support reproducibility and comparative evaluation in applied FinTech research.
Despite growing interest in automated waste detection, existing surveys either focus on a narrow set of models or lack systematic comparisons across object detection paradigms. This review addresses that gap by examining recent advances in deep learning for waste management, spanning two-stage detectors (Faster region-based convolutional neural network (Faster R-CNN) and Mask region-based convolutional neural network (Mask R-CNN)), single-shot frameworks (you only look once version 1 (YOLO)v1 to YOLOv11), and emerging Transformer-based models (ViT-WM and AL-DETR). Faster R-CNN achieved category-level accuracy of 91.68% and overall accuracy of 89.68%, while Mask R-CNN reported AP values between 26.2% and 34.5% across varied datasets. YOLO models demonstrated strong real-time capability, with YOLOv5 reaching a mAP@0.5 of 92.96% and YOLOv8 achieving 97.63% accuracy with precision and recall above 93%. Transformer-based approaches are especially promising: ViT-WM achieved 98.17% accuracy, the highest among reviewed models, and AL-DETR reported a mAP of 58.9% while integrating active learning (AL) strategies to reduce reliance on extensive labeled data. These results emphasize YOLO’s efficiency for real-time waste sorting and the potential of Transformer architectures for handling complex, cluttered environments. Remaining challenges include dataset variability, computational demand, and limited standardized benchmarks. Future research should prioritize developing comprehensive datasets, optimizing Transformers for real-time use, and leveraging AL to enhance generalizability with reduced annotation effort.
Brain tumors are life-threatening conditions requiring accurate and timely diagnosis for effective treatment. This paper proposes a novel hybrid model combining U-Net for tumor segmentation and residual network 50 (ResNet50) architecture for classification to achieve performance in brain tumor classification from magnetic resonance imaging (MRI) images. This paper proposes a novel hybrid model that integrates U-Net for tumor segmentation with ResNet50 architecture for classification, enabling robust multi-class classification across glioma, meningioma, pituitary tumor, and no tumor classes. Utilizing a diverse dataset of 7,023 MRI images, the model achieves a remarkable accuracy of 99.78±0.05%, outperforming existing methods. Compared to related works, the proposed model demonstrates superior accuracy and scalability. This hybrid approach addresses key challenges in medical imaging, providing a robust and interpretable solution for real-world clinical applications.
This study introduces SignVerse, a novel bi-directional sign language translation (SLT) system, to enhance communication between the hearing-impaired community and the general public. SignVerse makes real-time, two-way conversations easy for Indian Sign Language (ISL) users—no special hardware needed. The system uses smart artificial intelligence (AI) tech: computer vision, deep learning, and natural language processing (NLP). When someone types or speaks, the text/speech-to-sign module runs the input through NLP-based syntactic reordering and shows the ISL translation using a lively 3D avatar. On the flip side, the sign-to-text/speech module leverages MediaPipe to spot hand landmarks in real time, and the convolutional neural network-long short-term memory (CNN-LSTM) model accurately recognizes each gesture. Everything works together to help ISL users connect smoothly with others 94.8% recognition accuracy, less than 1.8-second translation latency, and more than 90% gesture clarity in user studies are all demonstrated by experimental evaluations. The lightweight model, which is optimized through knowledge distillation, guarantees excellent performance even on common consumer devices. With significant potential for societal impact, SignVerse is a significant step toward real-time, AI-driven ISL translation. When everything is taken into account, it is a dependable, scalable, and reasonably priced choice for inclusive communication.
This paper provides a scoping review of gamified learning in virtual reality (VR) from 2010-2025, which combines trends from various disciplines, game types, and learning approaches. This paper differs from previous ones, which mainly focus on bibliometric results, by critically examining 108 publications to determine the dominant types of VR games, their correlation with learning outcomes, and the impact on the learning process. This paper follows the preferred reporting items for systematic reviews and meta-analyses extension for scoping reviews (PRISMA-ScR) protocol. The findings suggest that simulation, role-playing, and problem-solving VR games are the most common, particularly in healthcare and science, technology, engineering, and mathematics (STEM) fields, improving engagement, retention, and skill development. The use of VR games in other fields and the needs of neurodiverse or physically disabled learners have yet to be explored. The review discusses the use of big data and cloud computing for VR deployment and adaption, and the significance of low and no-code technologies for educators developing VR without programming knowledge. Through the synthesis of patterns, research gaps, and cross-disciplinary challenges, this study provides a roadmap for VR-based gamification research, with emphasis on inclusivity, ethical considerations, long-term learning effectiveness beyond novelty effects, and sustainable educational integration.
The increasing demand for high-speed, low-latency connectivity has driven the rapid deployment of fifth-generation (5G) networks. However, the enhanced performance of 5G systems comes at the cost of higher energy consumption, posing a significant challenge to sustainability goals. This study explores energy-efficient coordination strategies for macro and pico cells to optimize power usage while maintaining network performance. By employing a systematic mapping study (SMS) and a systematic literature review (SLR), we analyze current research trends, challenges, and emerging solutions in energy-efficient 5G network design. Key strategies, including AI-driven resource allocation, dynamic spectrum management, and interference mitigation techniques, are examined to assess their effectiveness in reducing energy consumption. The findings highlight the critical role of intelligent coordination mechanisms in achieving a balance between energy efficiency and service quality. This research contributes to the development of sustainable 5G architectures by identifying optimal methodologies for macro- and pico-cell integration, paving the way for greener and more adaptive next-generation networks.
The integration of solar and wind energy has increased electricity generation but also introduced power quality disturbances (PQDs) that threaten grid stability. This study examines the detection and classification of five PQD types—voltage sag, swell, interruption, harmonics, and normal conditions—across noisy environments (0, 10, 20, and 30 dB) signal-to-noise ratio (SNR). Traditional methods— support vector machine (SVM), random forest (RF), artificial neural networks (ANN), and 1D convolutional neural networks (1D CNN)—are evaluated on raw signal data, while advanced models—2D CNN and fused 2D CNN-LSTM—utilize time-frequency representations (continuous wavelet transform (CWT) and short-time Fourier transform (STFT)). Results show that deep learning (DL) models achieve high accuracy even in noisy environments, with the fused 2D CNNLSTM using CWT outperforming all other methods. Noise adversely affects feature extraction, with CWT consistently outperforming STFT under low SNR conditions. These findings demonstrate that combining DL models with robust time-frequency analysis and temporal modeling enhances PQD classification and supports dependable monitoring in smart grid environments.
This study proposes a face recognition pipeline that integrates scale-invariant feature transform (SIFT) descriptors, the bag of visual words (BoVW) model, Kernel principal component analysis (KPCA), and support vector machine (SVM) classification. It starts by extracting local keypoint descriptors from preprocessed face images using SIFT. These descriptors are subsequently vector-quantized into a visual vocabulary through MiniBatch K-Means clustering, yielding fixed-length BoVW histograms for each image. Nonlinear dimensionality reduction is achieved by applying KPCA with a radial basis function, addressing the complexity of the feature space. The resulting compact feature representations are subsequently classified using a linear SVM. The proposed method is evaluated on labelled faces in the wild (LFW) dataset with filtered 100 classes, demonstrating notable classification accuracy and reliable generalization across training, validation, and testing splits. Our experimental evaluation confirms that integrating local invariant features, nonlinear feature reduction, and discriminative classification allows the proposed method to exceed state-of-the-art face recognition performance. In addition, this proposed method is particularly suitable for scenarios with limited training data and computational resources, providing a lightweight but robust alternative to deep learning-based models.
Heart sound, or phonocardiogram (PCG), signals are often distorted by noise from respiration, movement, and the surrounding environment, which complicates accurate cardiac feature extraction in portable monitoring systems. This study aims to design and evaluate an effective digital filtering method to enhance PCG signal quality obtained from a low-cost acquisition system based on a condenser microphone sensor and an ESP32 microcontroller. The main contribution of this work is the implementation and comparison of two noise-reduction approaches: Butterworth band-pass filters of various orders and the Kalman filter applied to PCG signals acquired from mannequin-based simulations. Data were recorded for 10 seconds at a 1000 Hz sampling rate, processed in MATLAB, and analyzed using the fast Fourier transform (FFT) to determine the optimal frequency ranges. Experimental results demonstrate that the 8th-order Butterworth band-pass filter achieved the highest signal-to-noise ratio (SNR) improvement, averaging 25.659 dB, outperforming other configurations. These findings indicate that an appropriately tuned Butterworth filter provides a simpler yet robust solution for real-time PCG denoising in embedded systems. Future work will integrate the filtering process directly into the ESP32 firmware and evaluate its performance on human subjects to enhance clinical applicability.
Waste has become a serious environmental issue that requires effective and efficient management systems. This study compares three residual network (ResNet) variants (ResNet-34, ResNet-50, and ResNet-101) within the single shot detector (SSD) framework for visual waste detection. The dataset consists of 800 images in four categories—food, plastic, paper, and wood—with a 70:20:10 split for training, validation, and testing. The backbone architecture, optimizer (stochastic gradient descent (SGD) and Adam), and learning rate are varied to evaluate fifteen experimental configurations. Model performance is assessed using precision, recall, F1-score, and mean average precision (mAP). The results show that SSD–ResNet-34 with SGD and a learning rate of 0.0005 works best, with a mAP of 91.02%, which is better than deeper backbones. Deeper backbone architectures do not consistently improve accuracy; instead, they increase the risk of overfitting on small datasets. These findings highlight that lightweight architecture, when used with the right hyperparameter settings, strikes a better balance between accuracy, computational efficiency, and generalization for small-scale waste detection tasks.
Silk production depends heavily on accurate cocoon grading, yet manual inspection is slow, inconsistent, and varies between operators. This creates problems in quality control and affects the final yield of raw silk. To address this, we present an automated system that uses computer vision to detect, separate, and grade silk cocoons without human involvement. The system combines a you only look once version 8 (YOLOv8) model for segmenting individual cocoons from tray images and an EfficientNetB0 classifier for identifying defects across six categories, including one qualified class and five defect types. After detection and grading, the pipeline also estimates the percentage of good cocoons and predicts silk yield based on standard industry measures. The model was trained on 3,068 cocoon samples and achieved 96.1% mean average precision (mAP) for segmentation and 97% accuracy for classification. The system can count cocoons, assess quality distribution, and provide batch-level yield estimates. This automated approach improves reliability, reduces manual effort, and offers consistent grading suitable for both farm-level and industrial environments. With low operating cost and simple deployment, the system supports modern, scalable, and data-driven sericulture.
Sustainable agriculture faces increasing challenges due to climate variability, which affects crop productivity and resource efficiency. This study proposes a sustainable greenhouse system that integrates internet of things (IoT) sensors and machine learning models to optimize the microclimate for lettuce cultivation. Environmental data, including temperature, humidity, and light intensity, were collected through IoT sensors and processed using machine learning algorithms, specifically neural networks and support vector machines (SVM), implemented on the Orange data mining platform. The results indicate that the neural network model achieved superior performance, reaching an accuracy of 99.99% in predicting optimal greenhouse climate conditions, outperforming the SVM model. The best-performing model was subsequently implemented on an Arduino-based IoT system to automatically regulate greenhouse conditions. The proposed system improved resource efficiency and supported optimal lettuce growth while promoting sustainable agricultural practices. These findings demonstrate that integrating IoT and machine learning can enhance greenhouse management, contributing to climate-resilient agriculture and improved food production systems.
This paper proposes a modified approximating function (MAF)-based analytical method for broadband impedance matching in radio-electronic systems. Unlike traditional Chebyshev and Butterworth approaches, which rely on fixed pole distributions and predefined amplitude responses, the proposed method analytically embeds load-specific transmission zeros directly into the approximation function. This modification enables more accurate reconstruction of frequency-dependent impedance behavior without increasing the network order or circuit complexity. The method establishes a unified analytical synthesis framework linking impedance modeling, ladder-network realization, and constrained optimization. Validation was performed over the 1–10 GHz band using numerical simulations, Monte Carlo tolerance analysis, and prototype measurements. Compared with classical Chebyshev and Butterworth designs, the MAF-based approach achieves a 15–25% reduction in maximum reflection coefficient, a 30–40% decrease in optimization iterations, and improved robustness, with reflection variations remaining within 2% under ±10% parameter deviations. The results confirm that the proposed method provides superior analytical flexibility, improved matching accuracy, and reduced computational effort, making it suitable for automated broadband radio frequency (RF) design applications.
To enhance user security awareness in response to the numerous cyberattacks targeting users, we developed an artificial intelligence (AI)based prototype utilizing the naive retrieval augmented generation (RAG) method. In pursuit of this objective, we conducted user testing employing the usability testing method to evaluate users' comprehension of the developed prototype. We integrated moderate and guerrilla techniques in usability testing by engaging 20 randomly selected respondents from the government, private sectors, and industries. The majority of participants were male, aged 26-35 years, holding a bachelor's degree, and possessing 510 years of computer experience. The test data were analyzed using the USE assessment matrix, which includes four assessment parameters: usefulness, satisfaction, ease of use, and ease of learning (USE). The data were presented in tabulated form, with total and average values. The test results indicate that usefulness, satisfaction, ease of USE achieved a total value exceeding 4.00 and an average value of 4.29, within an interval range of 4.20-5.00, categorized as very good. The findings of this study have implications for enhancing user security awareness and provide feedback for refining the framework and prototype in subsequent research.
Eczema, also known as dermatitis, is a chronic skin condition that causes recurring episodes of dry and itchy skin. It can be managed through medication and by avoiding triggers like stress and certain foods. To help patients avoid food-related triggers, researchers conducted a study to detect allergenic food compositions in packaged products using optical character recognition (OCR) techniques, specifically open computer vision (OpenCV) and Tesseract. The study involved analyzing 120 images of food labels. The process included several steps: preprocessing the images by converting them to a text-friendly format (gray scaling, denoising, and thresholding), using Tesseract for text detection, followed by case folding and tokenization. The results showed that the system achieved an average text detection accuracy of 61.88% and an average allergen detection accuracy of 83.06%. The highest accuracy for text detection was 78.52%, and the highest accuracy for allergen detection was 100%. These findings suggest that OCR techniques can be a useful tool for helping eczema patients manage their diet and minimize flare-ups.
As the world suffers from intrusions and malware extensively nowadays, intru-sion detection systems (IDS) play a critical role in protecting cyberspace from attacks. However, attacks become more complex every day, leading to the neces-sity of developing new techniques that can protect our digital infrastructure from cyber-attacks. Deep learning (DL) is one of the techniques that are investigated to fight against cyber-attacks. However, due to the nature of traffic data, most of the techniques focus on the deep neural network (DNN) as the performance of the DNN dependsonthetraining data. In this paper, we investigate the effective-ness of using convolutional neural networks (CNN) to detect malware apps and network intrusions. The cybersecurity datasets are converted from tabular data into images using the DeepInsight technique. Experiments are conducted using two datasets, NSL-KDD and CICMaldroid20 datasets. The proposed method demonstrates that converting cybersecurity datasets from tabular data into im-ages may decrease the model’s accuracy. Furthermore, this approach introduces additional challenges in the detection of network intrusions and malware. More-over, the added architectural complexity may cause a dilution or distortion of feature representations, making it harder for the model to preserve the original semantic meaning of critical features.
Electric vehicle charging infrastructure (EVCI) has become essential. However, these infrastructures are increasingly vulnerable to cyber threats, particularly through spoofing and adversarial attacks on charging ports. This paper introduces a robust anomaly detection framework leveraging long short-term memory (LSTM) based autoencoders to identify anomalies in electric vehicle (EV) charging port current magnitudes. A simulated EVCI setup is developed in MATLAB/Simulink to capture charging behaviors under normal and adversarial scenarios. To generate adversarial data, the fast gradient sign method (FGSM) is employed. The reconstructed outputs from the LSTM-autoencoder (LSTM-AE) are statistically compared to real-time observations using the Kolmogorov–Smirnov (KS) test to detect anomalies. The framework achieves a high detection accuracy of 98.5%, demonstrating strong resilience against cyber-injected data anomalies and setting a foundation for enhanced EVCI cybersecurity.
Autonomous mobile robots (AMRs) are becoming integral to applications ranging from industrial automation to urban mobility. A core challenge in deploying AMRs effectively is the path planning problem determining an optimal and collision-free path from a start to a goal location within a given environment. This paper proposes a novel method, Dhouib-Matrix-SPP (DM-SPP), that enhances path planning efficiency and adaptability for AMRs operating in different statistical environment. Basically, DM-SPP is developed to unravel the shortest path in a graph and based on columns-rows structure with polynomial computational time. Here, the DM-SPP method is adapted to plan the shortest feasible path between two positions while avoiding obstacles. In order to prove the validity of the proposed DM-SPP method, it is applied to different environments and compared to different case studies taken from the literature. The simulation results show that the DM-SPP method was able to find, with a significantly lower number of iterations, the optimal solutions in comparison with other results obtained by the genetic algorithm (GA) method. DM-SPP presents an overall average improvement in computation time of (37882.55%) compared to the GA, which can reduce search and execution time.
The government ensures educational quality in universities through a quality assurance (QA) system implemented via accreditation, which evaluates both study programs and institutions. A key concern in accreditation is the decline in new student enrollment, making accurate predictions of enrollment numbers essential for quality assessment. This study proposes a linear regression (LR) model to forecast future university student enrollments based on enrollment figures from the previous year as input feature. Using a dataset from one of Indonesia’s leading university spanning 2013 to 2023, the experimental results demonstrate that the LR model outperforms other regression techniques, including multi-layer perceptron (MLP), K-nearest neighbors (KNN), decision tree (DT), and random forest (RF). The LR model achieves R² values between 0.87 and 0.95, reflecting a strong linear relationship between current and future student numbers. It also delivers high accuracy, with root mean square error (RMSE) values ranging from 11.72 to 41.21 per year. The trained LR model has been integrated into a web-based system, offering data visualization and enrollment predictions to support university management in monitoring quality, addressing enrollment challenges, and facilitating informed decision-making.
