Accurate histopathological classification is essential for early cancer diagnosis, yet deep learning models often require extensive computational resources, limiting their deployment in real-world clinical settings. Parameter-Efficient Fine-Tuning (PEFT) strategies offer a promising alternative, but existing methods such as LoRA and adapter-based approaches still update a substantial number of parameters. This study introduces Adaptive BitFit, a sparsity-driven fine-tuning framework that selectively updates only the most informative bias parameters using gradient-based importance profiling. Integrated with MobileViT and Vision Transformer backbones, the proposed method significantly reduces computational overhead while preserving high diagnostic accuracy. On the LC25000 dataset, standard BitFit alone achieved up to 99% accuracy with around 99% F1-score, demonstrating the potential of bias-only tuning even before applying the proposed framework. Extensive experiments on LC25000, NCT-CRC-HE-100K, and the CRC-VAL-HE-7K held-out validation set (no patient overlap with the training NCT-CRC-HE-100K but from the same source study) demonstrate that Adaptive BitFit achieves up to 60% reduction in trainable bias parameters with minimal accuracy loss, attaining 97.86% accuracy on LC25000 and 96.99% on CRC-VAL-HE-7K. Also the model with 20% sparsity level, achieved around 92% accuracy and F1 score on a brain cancer MRI dataset, which proves the generalizability of this study. Comparative analyses against Adapter, BitFit, LoCon, and SSF confirm the superior balance between efficiency and performance. Furthermore, Grad-CAM visualizations enhance interpretability by highlighting clinically relevant tissue structures. These results position Adaptive BitFit as an effective, lightweight, and generalizable solution for deploying deep learning models in resource-constrained medical environments. Code Availability: https://github.com/N-Kibria/PEFT-CancerX.

With an increased frequency of accidents and birth defects issues, such as congenital amputation, the need for improved prosthetic devices increases. In this work, we present a rigid prosthetic system fabricated by 3D printing and a deep learning network to classify hand movements with surface electromyography (sEMG) signals of forearm muscles. The model, called NeuroAttend-EMG, comprises a CNN, a BiLSTM, and a self-attention module, which runs on edge devices like the Jetson Nano. With a parameter count of 1.2 million and a computational complexity of 0.0015 GFLOPs, the model achieved 98.58% and 97.34 % classification accuracies on two distinct sEMG-based tabular signal datasets. Four time-domain features were selected and evaluated for feature extraction on both datasets, ensuring a robust representation of the EMG signals. Furthermore, interpretable AI methods like Shapley Additive Explanations (SHAP) were used to recognize the most important features. A 3D-printed hand prosthetic was created, which performed some of the classification movements based on the unseen sensor data provided within the Jetson Nano board. The study proves that it is possible to apply the light deep learning to the classification of hand movements in prosthetics for real-time control, striking a balance between efficiency and accuracy in edge computing. The study focuses on the field of assistive robotics from optimized deep learning-based prosthetics systems, enabling affordable, accurate movement classification on low-powered devices.

Automatic Identification System (AIS) data are vital for maritime domain awareness, yet they often suffer from domain shifts, data sparsity, and class imbalance, which hinder the performance of predictive models. In this paper, we propose a robust data augmentation method, AISCycleGen, based on Cycle-Consistent Generative Adversarial Networks (CycleGAN), which is tailored for AIS datasets. Unlike traditional methods, AISCycleGen leverages unpaired domain translation to generate high-fidelity synthetic AIS data sequences without requiring paired source-target data. The framework employs a 1D convolutional generator with adaptive noise injection to preserve the spatiotemporal structure of AIS trajectories, enhancing the diversity and realism of the generated data. To demonstrate its efficacy, we apply AISCycleGen to several baseline regression models, showing improvements in performance across various maritime domains. The results indicate that AISCycleGen outperforms contemporary GAN-based augmentation techniques, achieving a PSNR value of 30.5 and an FID score of 38.9. These findings underscore AISCycleGen's potential as an effective and generalizable solution for augmenting AIS datasets, improving downstream model performance in real-world maritime intelligence applications.

This study evaluates the electrical performance, preliminary surface-temperature behaviour, and outdoor operability of a reflector-assisted spherical photovoltaic (PV) prototype using short-term outdoor measurements and geometry-aware numerical simulation. The spherical prototype was fabricated using 450 monofacial silicon solar cells mounted on an opaque 0.70 m diameter spherical substrate and integrated with a reflective base. The simulation framework was used to examine geometry-driven irradiance collection, temperature-corrected electrical output, and selected-period and annual energy-yield trends under Doha, Qatar, conditions. A nine-day outdoor experiment was conducted as an initial proof-of-concept evaluation, during which voltage, current, power, surface temperature, and environmental variables were monitored using an ESP32-based data acquisition system. The prototype generated a cumulative measured energy output of 183.17 Wh, confirming successful outdoor operation and a clear irradiance-dependent electrical response. The measured profiles followed the expected daily solar-availability trend and therefore provide useful experimental support for the qualitative behaviour predicted by the geometry-aware model. Preliminary infrared surface-temperature observations were also recorded. however, because wind speed and IR measurement parameters were not fully characterized, these measurements are reported only as indicative thermal trends and not as evidence of a confirmed thermal advantage. For the selected simulation period, the reflector-assisted spherical configuration generated 113.0 kWh/module, compared with 105.3 kWh/module for the fixed-tilt flat reference, corresponding to a 7.3% increase. The present work establishes a practical prototype-scale foundation for spherical PV development and identifies the key design improvements needed for future optimization, including protective encapsulation, calibrated irradiance measurement, quantitative soiling assessment, durability testing, and longer-term outdoor validation.

Infrared (IR)-induced heating in photovoltaic (PV) systems is a critical challenge that lowers efficiency and accelerates module degradation. In this work, we propose a multilayer thin-film IR filter (TiO2 (50 nm)/NiO x (100 nm)/Ag (various thickness)) as a reduced heating solution integrated with PV modules. The filter is designed to transmit visible light while reflecting IR radiation, thereby reducing thermal load without sacrificing photovoltaic current. We incorporate machine learning models specifically Gaussian Process Regression (GPR) to optimize the Ag layer thickness for maximum performance. The AI models are trained on a combination of experimental optical data, enabling efficient exploration of the thickness-performance space. Key findings demonstrate that an optimized Ag thickness (∼10 nm) yields high IR reflectance (over 50% in the 750-1200 nm range) while maintaining sufficient visible transmittance (>50%). Silicon solar cells with this filter showed improved performance: the short-circuit current and power output increased due to reduced thermalization losses, translating to a ∼2-3 °C drop in operating temperature and corresponding efficiency gains. These improvements can extend PV module lifespan and energy yield. Our results indicate that the TiO2/NiO x /Ag filter can be manufactured via scalable e-beam evaporation, and the integration of AI optimization accelerates the design of such photonic coatings. The demonstrated performance enhancements and the low-cost, scalable nature of the solution highlight strong potential for commercial deployment in extending PV module longevity. We adopt Gaussian Process Regression (GPR) as a surrogate tailored to small, high-fidelity experimental data sets. GPR provides calibrated predictive uncertainty and smooth, physics-consistent interpolation across film thickness and wavelength, enabling uncertainty-aware selection of robust thicknesses with few experiments. This contrasts with polynomial/linear fits (insufficiently expressive) and black-box models without calibrated uncertainty.

Sodium-ion cells suit low-cost stationary and cold-climate energy storage, but gas evolution during cycling remains poorly understood and can trigger sudden failure. This study re-analyses a multi-condition ageing campaign on a commercial sodium-ion cell with a documented severe-gassing history, combining electrochemical diagnostics with physics-informed machine learning. Cells were cycled across depths of discharge, current rates and temperatures, with periodic check-ups separating temporary capacity loss from permanent degradation. Depth of discharge, not current rate, emerged as the dominant driver of capacity fade under normal operating temperatures. Below freezing the mechanism changes: a charge-balance analysis shows charge consumed by side reactions far exceeds sodium loss alone, pointing to a self-sustaining, regenerative process consistent with the cell’s known gassing behaviour, though gas evolution was not directly measured. A discharge-voltage plateau that grows with cycling, alongside rising resistance and hysteresis, gives a practical, non-invasive electrical signature of this degradation, though it tracks rather than precedes collapse in the fastest-failing cells. Building on these findings, we introduce NaPINN, a sodium-ion physics-informed neural network that predicts a cell’s next capacity reading under a strict, leakage-free validation scheme, matching a well-tuned baseline while keeping every prediction non-negative and physically consistent by design. An ablation study confirms that this structural consistency, not added model complexity, drives the network’s accuracy. Together, these results give sodium-ion battery developers a candidate non-invasive signature for this cell’s documented gassing-related degradation and a physics-constrained tool for forecasting cell health within the tested conditions

It's important to monitor road issues such as bumps and potholes to enhance safety and improve road conditions. Smartphones are equipped with various built-in sensors that offer a cost-effective and straightforward way to assess road quality. However, progress in this area has been slow due to the lack of high-quality, standardized datasets. This paper discusses a new dataset created by a mobile app that collects sensor data from devices like GPS, accelerometers, gyroscopes, magnetometers, gravity sensors, and orientation sensors. This dataset is one of the few that integrates Geographic Information System (GIS) data with weather information and video footage of road conditions, providing a comprehensive understanding of road issues with geographic context. The dataset allows for a clearer analysis of road conditions by compiling essential data, including vehicle speed, acceleration, rotation rates, and magnetic field intensity, along with the visual and spatial context provided by GIS, weather, and video data. Its goal is to provide funding for initiatives that enhance traffic management, infrastructure development, road safety, and urban planning. Additionally, the dataset will be publicly accessible to promote further research and innovation in smart transportation systems.

Cognitive workload recognition from EEG signals remains challenging due to real-world artifacts and missing data. To address this, we propose a unified reconstruction-classification framework that integrates EEG denoising and workload inference. Firstly, a Conditionally-Guided Denoising Diffusion Probabilistic Model (CG-DDPM) is introduced, which combines Gaussian noise modeling, a conditional encoder, and a Conditional Variational Autoencoder (CVAE) to guide a U-Net in removing diverse artifacts such as EMG, EOG, ECG, respiratory motion, powerline interference, and masked regions, while preserving essential neural activity. Secondly, an advanced classification network, EEG Graph Fusion Network (EEGGX-Net), is designed with a Hybrid Multi-Branch Encoder, a Bidirectional Multi-Head Cross Attention Fusion (MHCAF) module, and a Hierarchical Capsule Classifier (HCC) to jointly capture spatial, topological, and nonlinear dynamics of EEG signals. Both quantitative metrics (SNR: 16.50 dB, CC: 0.86, SC: 0.79) and topographic visualizations confirm CG-DDPM's efficacy in restoring meaningful neural activity. Using a strict subject-independent 5-fold cross-validation protocol on the STEW dataset, along with external validation on the iNCog-EEG dataset, the framework achieves state-of-the-art performance in both binary and ternary settings across raw, noisy, and reconstructed conditions, exceeding 98 % and 95 % accuracy, with narrow 95 % confidence intervals confirming statistical reliability. Comparative analyses also showed statistically significant performance gains, supported by p-value evaluations across models. Ablation studies and t-SNE visualizations reaffirm robustness and generalization. These results highlight the significant potential of this unified framework for real-time cognitive workload assessment in noise-prone environments such as neuroergonomics and human-automation systems.

Floating waste in inland water bodies poses severe threats to aquatic ecosystems, water quality, and public health. The accurate and timely detection of such waste is essential for enabling autonomous cleanup sys-tems like unmanned surface vehicles (USVs). However, detecting floating waste remains challenging due to the small size of debris, water surface reflections, glare, and complex backgrounds. This study presents a comparative evaluation of state-of-the-art deep learning-based object detection models—YOLO (v8–v10), Faster R-CNN, and Real-Time Detection Transformer (RT-DETR)—using the FloW-Img dataset, which is specifically designed for floating waste detection from USV perspectives. To enhance detection performance, we also explored four ensemble strategies: Weighted Box Fusion (WBF), Non-Maximum Suppression (NMS), Soft-NMS, and Non-Maximum Weighted (NMW). Our experiments show that the ensemble of RT-DETR-X and Faster R-CNN using WBF achieves the best results, with a mean Average Precision (mAP50) of 89.081

Rapid and reliable disaster severity assessment from social media is difficult due to noisy content, modality imbalance, and limited labeled data. This paper introduces ReliefNet, a real-time multimodal framework for disaster severity classification that jointly analyzes images and text. The system integrates DualEmbedNet, a dual-encoder transformer for textual classification, and DisasterNet, a CNN with channel-wise attention for visual analysis, combined through an accuracy-weighted decision-level fusion mechanism that dynamically adapts to modality reliability at inference time. To mitigate labeling scarcity, we construct a unified dataset of 14,996 image-text pairs by merging CrisisMMD and TSEqD. Unlabeled samples are annotated using unsupervised multimodal K-Means clustering, followed by Grad-CAM guided ROI refinement and manual auditing to reduce label noise. Models are evaluated using an 80/10/10 train-validation-test split with metrics including accuracy, precision, recall, F1-score, calibration error, and robustness under missing-modality conditions. The image-only model achieves an F1-score of 97.66%, while ReliefNet attains 97.00% F1 with superior robustness and stability under noisy or incomplete inputs. Accuracy-weighted fusion improves F1-score by 0.7% over equal-weight fusion and outperforms standard CNN baselines such as ResNet and EfficientNet. All models operate in real time on a single NVIDIA GPU, demonstrating ReliefNet's practical deployability.

Artificial intelligence (AI) is rapidly reshaping engineering education, yet evidence on its impact and best practices remains fragmented across disciplines and approaches. To consolidate current knowledge, this review applies a PRISMA-ScR scoping methodology, surveying over 3,000 records from Scopus, Web of Science, IEEE Xplore, ERIC, and Google Scholar (2000–2024) and synthesizing 200 relevant studies. The analysis highlights dominant applications such as intelligent tutoring, adaptive assessment, and VR/AR simulation, with emerging interest in large language model–based feedback and generative AI. Reported outcomes suggest moderate improvements in student performance and engagement, though gaps remain in areas such as ethics, equity, and long-term evaluation. Expert perspectives further underscore opportunities in micro-credentialing and AI-driven design studios, alongside challenges of data privacy, transparency, and faculty readiness. Overall, AI-enhanced strategies show strong potential to support personalization and innovation in engineering education, but sustainable adoption will require longitudinal studies, ethical safeguards, and targeted professional development.

Electric power generation in an islanded microgrid depends largely on a variety of renewable energy sources (RESs), and the load frequency control of the islanded microgrid has become significantly challenging due to increased frequency fluctuations induced by the intermittent RESs and low system inertia. Although energy storage systems are crucial contributors in the frequency regulation of microgrids, their high costs and significant power density demands the use of alternative approaches, like using electric vehicle (EV) batteries to address frequency fluctuations in microgrids. As a way to address the issues of frequency fluctuations due to intermittent inverter-equipped power source penetration, an EV energy storage-based virtual inertia (VI)/auxiliary control strategy using brain emotional learning (BEL)-based auxiliary controller is proposed in this research paper. The proposed control strategy’s effectiveness is demonstrated using MATLAB/SIMULINK simulations and comparative analysis against other control methods under various scenarios featuring diverse load disturbance profiles and under intermittent RES. Additionally, the proposed BEL-based VI control strategy could provide reduced frequency deviation compared to the proportional integral and linear quadratic Gaussian-based VI controllers under various operating conditions and system inertia parameter variation.

Accurate prediction of pressure variations during CO2 injection is crucial for the safe and efficient operation of carbon capture and storage (CCS) in subsurface geological formations. Excessive pressure can compromise caprock integrity or create leakage pathways, posing significant operational risks. Traditional numerical simulations, while precise, are computationally intensive and often unsuitable for real-time applications. This study presents PDP-PINN, a Fourier-augmented physics-informed neural network designed for pressure drop prediction across heterogeneous porous media during CO2 injection. The model integrates spatio-temporal flow features with physical constraints derived from Darcy’s law and employs Fourier feature embeddings to mitigate spectral bias. Evaluated on a high-resolution two-phase CO2-water simulation dataset, PDP-PINN achieves strong predictive performance across various heterogeneity levels, with a coefficient of determination (R2) up to 0.9941 and mean squared error (MSE) of 5.56, while maintaining physical consistency. Furthermore, explainable AI techniques, including SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), are applied to provide global and local interpretability, revealing the influence of individual features on model predictions. The proposed architecture demonstrates potential for real-time injectivity assessment, flow resistance estimation, and subsurface risk mitigation in CCS operations.

The identification of spoofed AIS data is crucial for ensuring secure and dependable maritime navigation, vessel tracking, and surveillance systems. Recent advances in generative modeling, particularly CycleGANs, have made identifying falsified AIS trajectories significantly more difficult as synthetic data can closely replicate the spatiotemporal characteristics of genuine vessel movements. Conventional anomaly detection and supervised learning approaches often fail to generalize to such sophisticated spoofing strategies. This paper presents an LSTM-based fingerprinting framework for robust detection of CycleGAN-generated AIS spoofing. The proposed method learns intrinsic spatiotemporal fingerprints of authentic AIS trajectories by modeling long-term temporal dependencies and vessel motion dynamics, without requiring labeled synthetic data during training. Incoming AIS sequences are classified as real or spoofed by comparing their learned fingerprint representations using a similarity-based decision mechanism. Experimental evaluations conducted on a large dataset comprising real AIS data and CycleGAN-generated synthetic trajectories demonstrate that the proposed approach achieves a precision of 0.95, recall of 0.96, F1-score of 0.955, outperforming state-of-the-art anomaly detection methods including kinematic interpolation, GeoTrackNet, and Transformer-based classifiers. Furthermore, the method exhibits strong generalization to unseen GAN variants, with only a marginal recall degradation from 0.96 to 0.93 and a consistently low false positive rate of 0.02–0.03. These findings demonstrate the effectiveness and robustness of the proposed fingerprinting approach in real-world maritime security scenarios, offering a dependable mechanism for mitigating sophisticated AIS spoofing attacks.

The Metaverse integrates immersive virtual environments, digital interaction, and intelligent computational systems to support new forms of social, educational, industrial, and economic activity. Artificial intelligence is one of the key enabler of this ecosystem because it supports perception, reconstruction, interaction, content generation, personalization, security, and scalable computation. This review presents an AI-centric synthesis of the Metaverse by examining how AI contribute to immersive virtual systems. The paper first outlines the theoretical foundations and enabling technologies of the Metaverse, then reviews major application domains, including intelligent avatars, virtual content creation, recommendation systems, blockchain-enabled digital economies, accessibility, and smart-city digital twins. It further analyzes core AI functions in perception, computation, reconstruction, interaction, and cooperation, followed by a comparative discussion of AI methodologies and their suitability for different Metaverse tasks. The review also discusses public datasets, reproducibility issues, ethical risks, privacy concerns, scalability barriers, and future research directions. Through reducing fragmented discussion and synthesizing methodological trends, this paper provides a structured roadmap for developing trustworthy, scalable, and human-centered AI-enabled Metaverse systems.

Crowdsourced smartphone-based road condition monitoring offers a cost-effective alternative to traditional infrastructure surveys; however, widespread adoption is hindered by serious privacy concerns. Raw sensor data reveals precise locations, speed violations, and behavioral patterns, deterring user participation due to risks of surveillance, re-identification, and misuse by authorities and insurers. Existing federated learning approaches protect only model gradients while leaving raw data exposed locally, and standalone privacy methods such as k-anonymity and differential privacy require centralized aggregation, introducing single points of failure and failing to balance privacy with utility for effective road anomaly detection. We propose CKDP-PINNFed, dual-layer privacy-preserving federated framework that addresses these limitations. First, we apply local contextual k-anonymity combined with differential privacy to anonymize raw sensor data before storage, achieving a 78.4% reduction in composite privacy loss score and a 99.1% reduction in uniqueness risk compared to differential privacy alone. Second, we employ FedProx with a hybrid Physics-Informed Neural Network and Random Forest (PINN-RF) architecture for decentralized training on anonymized data. The lightweight framework (0.19M parameters, 0.73 MB) achieves real-time processing at 0.0322 ms total latency per sample, comprising 0.0199 ms for CK-DP preprocessing and 0.0123 ms for PINN-RF inference, enabling practical on-device deployment. Experimental results on the RoadSens-4M dataset demonstrate an F1-score of 89.94% ± 0.35% and accuracy of 89.93% ± 0.19%, outperforming FedAvg by 1.64% and all other federated baselines under extreme non-IID conditions. Ablation studies confirm component synergy, SHAP analysis verifies physically meaningful feature interpretability, and external validation achieves 84.23% ± 1.18% accuracy on an independent dataset. These results establish CKDP-PINNFed as a practical and secure solution for trustworthy crowdsourced road monitoring that protects individual privacy.

Reliable unmanned aerial vehicle (UAV) detection is critical for autonomous airspace monitoring but remains challenging when integrating sensor streams that differ substantially in resolution, perspective, and field of view. Conventional fusion methods-such as wavelet-, Laplacian-, and decision-level approaches-often fail to preserve spatial correspondence across modalities and suffer from annotation of inconsistencies, limiting their robustness in real-world settings. This study introduces two fusion strategies, Registration-aware Guided Image Fusion (RGIF) and Reliability-Gated Modality-Attention Fusion (RGMAF), designed to overcome these limitations. RGIF employs Enhanced Correlation Coefficient (ECC)-based affine registration combined with guided filtering to maintain thermal saliency while enhancing structural detail. RGMAF integrates affine and optical-flow registration with a reliability-weighted attention mechanism that adaptively balances thermal contrast and visual sharpness. Experiments were conducted on the Multi-Sensor and Multi-View Fixed-Wing (MMFW)-UAV dataset comprising 147,417 annotated air-to-air frames collected from infrared, wide-angle, and zoom sensors. Among single-modality detectors, YOLOv10x demonstrated the most stable cross-domain performance and was selected as the detection backbone for evaluating fused imagery. RGIF improved the visual baseline by 2.13

Clickbait content on video-sharing platforms poses a significant challenge to information reliability, yet progress in automated detection has been constrained by the lack of large-scale, high-quality multimodal datasets. We present YTClickbait21K, a human-annotated YouTube clickbait dataset comprising 21,238 videos collected from 40 channels across 29 countries, covering diverse content categories such as news, entertainment, education, and gaming. Each sample includes structured metadata (title, description, engagement statistics) along with associated thumbnail images, enabling comprehensive multimodal analysis. To ensure annotation quality, every video was independently labeled by three annotators using a standardized decision framework that incorporates textual, visual, and cross-modal consistency cues, with final labels determined through majority voting. The dataset exhibits substantial inter-annotator agreement (k=0.65), confirming reliable labeling despite the inherent subjectivity of clickbait detection. By combining scale, annotation rigor, and multimodal richness, this dataset provides a robust benchmark for developing and evaluating machine learning models, facilitating research in cross-modal semantic understanding, and advancing automated content moderation systems.

This study proposes an incremental machine-learning framework for photovoltaic power prediction in desert conditions, explicitly considering performance differences between hydrophobiccoated and uncoated modules. A low-cost monitoring platform recorded environmental and electrical variables for six months in Qatar. Coated modules produced higher daily energy than uncoated modules under dusty conditions, indicating improved yield stability. For forecasting, multiple baseline regressors were evaluated and compared with a stacking ensemble. The proposed incremental stacking approach updates model parameters as new data arrive, improving robustness against non-stationary behavior caused by dust accumulation and thermal drift. The final model achieved strong predictive performance for maximum power estimation (MSE = 0.0873 and R2 = 0.9794), outperforming standard single-model baselines. The results demonstrate that combining coating-aware monitoring with adaptive learning improves reliability for photovoltaic yield prediction in harsh climates.