Abstract Reliable, site-specific wind data are essential for accurate wind energy assessment at campus and community scales, where conventional meteorological stations may be limited or economically expensive. This study presents the design, deployment, and experimental validation of a low-cost Internet of Things-based mini weather station (IoT-MWS) for wind monitoring at Edo State University, Iyamho (ESUI), Nigeria. The system integrates wind speed (WS02), wind direction (KY-040), and atmospheric pressure (BMP280) sensors with an ESP8266-based data acquisition unit, signal conditioning circuitry, wireless transmission, and cloud-based logging. Structured calibration modelling ensured traceable conversion from analogue signals to digital outputs. The anemometer exhibited a verified sensitivity of 0.1667 V/(m/s), achieving a wind speed resolution of 0.029 m/s per ADC least significant bit with ± 0.0145 m/s quantization error, while signal scaling (α = 0.180) limited ADC input to 0.90 V within the 1.0 V reference (10-bit, Δ = 0.00088 V/LSB). Wind direction resolution was 4.5° (± 2.25° maximum error), and pressure resolution was 0.002 hPa with ± 0.001 hPa quantization uncertainty, well below the ± 0.12 hPa specified accuracy. Statistical analysis of measured wind data (N = 31) yielded a mean wind speed of 2.20 m/s, standard deviation of 0.53 m/s, and turbulence intensity of 24.1%, with northerly winds accounting for approximately 58% of occurrences. Wind power density analysis indicates low energy potential, unsuitable for utility-scale deployment but appropriate for small-scale and educational applications. The total system cost was $99.12, representing a 42% reduction relative to a benchmark academic prototype and 84–95% savings compared with commercial systems, demonstrating that reliable, statistically validated, and economically scalable campus-scale wind monitoring is achievable using low-cost IoT technologies.

Driver drowsiness is a major cause of road accidents worldwide, leading to thousands of fatalities and billions in economic losses each year. This study presents a real-time drowsiness detection framework based on a hybrid Vision Transformer Convolutional Neural Network (ViT-CNN) architecture enhanced with multi-task learning and temporal attention. Unlike traditional sensor-based or reactive vehicle behavior methods, the proposed vision-based approach provides a non-intrusive, scalable solution capable of detecting early fatigue indicators such as eye closure, yawning, and head pose. The model leverages self-supervised pretraining on over 1.2 M unlabeled driving videos and is optimized for embedded deployment using INT8 quantization and TensorRT, achieving 99.27% accuracy, F1 = 0.98, and AUC = 0.998 while sustaining 42 FPS at 42 ms latency on the NVIDIA Jetson AGX Xavier. Explainability tools (Grad-CAM + + and Bayesian uncertainty estimation) ensure transparency in safety-critical contexts. Evaluation across six datasets demonstrates strong generalization, and the framework is adaptable to other fatigue-sensitive domains such as aviation and industrial safety.

Predicting academic outcomes is crucial for enhancing educational strategies, improving student performance, and facilitating early interventions. Machine learning (ML) has emerged as a powerful tool in educational data analysis, with linear regression being one of the most widely used algorithms due to its simplicity, interpretability, and effectiveness in modeling relationships between variables. This study investigates the performance of linear regression in predicting academic outcomes, focusing on its accuracy, strengths, and limitations compared to other predictive models. The research begins with a comprehensive review of existing literature on academic performance prediction, highlighting traditional statistical methods and advanced ML techniques. It then details the methodology, including data collection, preprocessing, and the implementation of a linear regression model. Key evaluation metrics such as mean squared error (MSE), and root mean squared error (RMSE) are employed to assess model performance. The findings reveal that linear regression provides a robust baseline for academic outcome prediction, particularly when interpretability and simplicity are prioritized. However, its performance is influenced by data quality, feature selection, and preprocessing techniques. The study also discusses the implications of these results for personalized learning, educational policy, and future research directions. Recommendations are provided for improving prediction accuracy, including the integration of feature engineering and hybrid modelling approaches. This work contributes to the growing body of research on educational data analytics by validating the utility of linear regression while acknowledging scenarios where more complex models may be advantageous. Future work could explore ensemble methods, deep learning techniques, and real-time predictive systems to further enhance academic outcome forecasting.

Abstract Thyroid disease affects approximately 20 million Americans and poses significant challenges for early diagnosis. In this study, we conduct a comprehensive comparative analysis using a recently curated, well-preprocessed tabular dataset of 9172 observations to evaluate the performance of both machine learning and deep learning models. The models include logistic regression, decision trees, random forest, support vector machines, XGBoost, LightGBM, artificial neural networks (ANN), and dense neural networks (DNN). Our results show that ensemble methods—particularly stacking and bagging—consistently outperform individual models in terms of F1 score and overall robustness. Stacking models with XGBoost as the meta-learner achieved the highest F1 score of 0.9944. We further demonstrate that addressing class imbalance through undersampling and label restructuring substantially improves model performance across multiple settings. These findings highlight the importance of ensemble techniques and thoughtful data preprocessing in medical classification tasks and offer updated performance benchmarks for thyroid disease prediction.

Abstract Global health emergencies, such as the COVID-19 pandemic, have demonstrated the power and volatility of social media as a platform for public communication. Millions of people worldwide relied on digital platforms like Twitter and Facebook for real-time information, updates, and expressions of emotion. However, this rapid flow of information also led to the widespread diffusion of misinformation and changing patterns of discourse; a phenomenon that calls for a systematic investigation. This study explores topic evolution in social media discourse during global health emergencies using Natural Language Processing methods, including topic modelling (LDA and BERTopic), sentiment analysis (VADER and BERTweet), and geospatial trend analysis. A dataset of 70,452 COVID-19-related tweets collected from Nigeria between March 2020 and December 2022 was analyzed using Natural Language Processing (NLP) and data mining techniques. Topic modelling (Latent Dirichlet Allocation and BERTopic), sentiment analysis (VADER and BERTweet), and geospatial trend mapping were employed to examine the interplay between sentiment and thematic focus over time. The results reveal five dominant thematic clusters— public health awareness , lockdown and economy , vaccine and misinformation , recovery and hope , and political accountability . Temporal and spatial analyses further indicate that public sentiment and attention evolved in alignment with pandemic milestones such as lockdown enforcement, vaccine rollout, and policy controversies. The study contributes a hybrid analytical framework that integrates sentiment and topic evolution modelling for infodemic monitoring. By mapping discourse evolution across phases of a crisis, the research provides actionable insights for policymakers, communication strategists, and public health agencies to design adaptive and evidence-driven responses during global health emergencies.

Abstract In various machine learning tasks, including image classification, data augmentation is widely used to expand the training dataset and prevent overfitting. In traditional data augmentation techniques, images from training data are flipped, distorted, added with some noise, or cropped to create new samples from the original images. This paper presents an improved and more effective data augmentation technique for image classification tasks, called Flip-and-Hide (FnH). Our technique was validated on four well-known datasets, MNIST, FMNIST, CIFAR-10 and CIFAR-100. We found 0.03% and 0.12% improvements on FMNIST and CIFAR-10, respectively, compared to the conventional Hide-and-Seek (HaS) augmentation technique. Moreover, we demonstrated the superiority of the proposed FnH augmentation technique by performing some visualisations. Consequently, our technique is particularly useful for image classification tasks that require limited training data.

Abstract The increasing complexity of combinatorial algorithms, especially for solving NP-hard problems like the knapsack problem, has heightened the need for energy- and time-efficient computational solutions. This research optimized two heuristics, namely the greedy algorithm (GA) and dynamic programming algorithm (DPA), with the primary goal of investigating their energy consumption and time complexity. By applying power models and instruction-level parallelism (ILP) techniques to measure and optimize their performance across varying problem sizes, a comprehensive comparison was made. The results demonstrate that the optimized algorithms, particularly GA, exhibit significant improvements in both time and energy efficiency, making them more suitable for large-scale, energy-sensitive applications. In contrast, classical DPA, while providing optimal solutions, shows higher time complexity and energy consumption, especially for larger datasets. However, the optimized DPA addresses these inefficiencies, showing marked improvements in both time and energy use across all instances. This study highlights the critical role of algorithmic optimization in achieving a balance between computational performance and sustainability. The findings contribute to the development of energy-efficient algorithms and underscore the importance of considering both time and energy metrics when solving combinatorial problems.

Deep convolutional neural networks (CNNs) are a popular choice for many image classification tasks due to their good performance. In general, there is a correlation between the generalization performance of a deep learning model, the number of training samples available, and the number of free parameters in the model. If there are too many free parameters in the model relative to the amount of training data, the model overfits the training data. Reducing the number of trainable parameters in CNNs while maintaining performance levels is an active area of research. This is important, especially in resource-constrained environments where either the size of the training data is insufficient or the memory of the hardware where the model is to be deployed is limited, such as in mobile vision applications. In this paper, kernel transposition is proposed as a method for the reduction of free parameters in the model’s convolutional layers. This allows kernel reuse, in which a model learns a given kernel only once but uses it twice. The learned kernels and their transposes can either be used in sequence or in parallel to one another. These form series and parallel convolutional modules respectively. The modules are used as replacements of the traditional convolutional layers in a CNN model. The use of the modules reduces the number of free parameters and computational costs associated with convolution operations in the model without necessarily compromising its generalization performance. The proposed method is generic and can be adapted to existing state-of-the-art network architectures as demonstrated in the experiments using established architectures. The proposed method was validated on the CIFAR-10 and CIFAR-100 standard datasets using a series of experiments with five model architectures: a basic 5-convolutional layer CNN model, the ResNet-56, small MobileNetV3, EfficientNetB0 and ConvNext model architectures; all trained from scratch. Models based on standard convolutional layers (for basic CNN and ResNet-56 architectures) and depth-wise convolutional layers (for MobileNetV3, EfficientNetB0 and ConvNext architectures) were compared against those based on the proposed convolutional modules. The models based on the standard and depth-wise convolutional layers were used as baselines. The comparisons were in terms of the number of parameters, computational efficiencies (Floating Point Operations / FLOPs) and classification accuracies of the models. For a kernel size k, the experiments showed that, despite improvements on parameter and computational efficiencies, models based on the proposed modules for the basic CNN (k = 3) architecture had lower accuracy levels compared to their baseline model. However, the huge saving on the number of parameters (> 88%) and FLOPs (> 84%) is significant compared to the drop in accuracy (< 4%). For the MobileNetV3 (k = 3 or k = 5) and EfficientNetB0 (k = 3 or k = 5), only the model based on the series module had lower accuracy than the baseline models. The models based on the parallel module were generally at par (within a margin of $$\pm$$ 1.5%) in terms of accuracy with their respective baselines. In relation to the ConvNext (k = 7) and the ResNet-56 (k = 3 or k = 11, or k = 19) architectures, other than the improved parameter and computational efficiencies, the classification accuracies of all the models based on the proposed modules were either at par or better than those of their baseline models. In all the above architectures, the use of the proposed modules reduced the number of convolutional parameters in the networks by at least 78% with fewer FLOPs relative to the baseline models

Abstract This paper presents an innovative PWM technique for a three-phase seven-level PUC-NPC converter that achieves automatic capacitor voltage balancing without complex control systems or additional sensors, addressing a key challenge in multilevel inverters. Unlike conventional designs requiring costly closed-loop control, the proposed hybrid PUC7-NPC topology generates five-level output using a single DC source while leveraging inherent redundant switching states for natural voltage regulation. The method simultaneously minimizes total harmonic distortion (THD) through optimized switching patterns, ensuring compliance with power quality standards. MATLAB/Simulink simulations confirm the converter’s stable performance across various switching frequencies, demonstrating its suitability for diverse applications including motor drives and grid-tied solar systems. The proposed solution represents a cost-effective, high-performance alternative for modern power conversion systems, offering simplified implementation without compromising output waveform quality.