ABSTRACT. In developing countries, the economy largely depends on agriculture. Crops such as tomato, potato, and chili play a crucial role in ensuring food security and economic stability, making their proper maintenance essential. However, leaf diseases caused by bacteria, fungi, or viruses pose significant threats to crop yield, particularly for small-scale farmers. To address this challenge, we propose a lightweight model, Augmented Enhanced Convolutional Neural Network with Attention (AECNN-A), designed for mobile-friendly and effective plant leaf disease classification. The architecture consists of two convolutional blocks integrated with MaxPooling, Dropout, and an additive attention mechanism, which emphasizes diseased regions to extract more discriminative features. The model can be deployed in real-time and offline on low-resource devices. Performance evaluation was conducted using the PlantVillage dataset (Kaggle), comprising 20,638 images of tomato, potato, and chili leaves across 15 classes, including both healthy and diseased samples. During preprocessing, images were encoded, normalized, and augmented on-the-fly with random flipping, rotation, zoom, and contrast adjustments. AECNN-A was evaluated using accuracy, macro F1-score, and confusion matrix metrics, achieving a test accuracy of 97.50\% and a micro F1-score of 0.98, outperforming models such as CNN, VGG16, VGG19, ResNet50, and FVBC in computational efficiency and classification accuracy. These results demonstrate that AECNN-A enables edge-based devices, real-time, and accurate plant disease detection, offering practical solutions for farmers. It's efficiency and rapid performance make it highly suitable for future research and applications in precision agriculture.

ABSTRACT. Accurate and early detection of plant diseases is crucial for sustainable agriculture and crop management. Automated computer-based systems are preferred over traditional human inspection due to the tedious and error-prone nature of manual approaches. Deep Learning (DL) models, particularly Convolutional Neural Network (CNN) models, are effective for this task due to their image categorization capabilities. In this work, we fine-tuned multiple pre-trained state-of-the-art CNN models, namely ConvNeXt, EfficientNetB0, and DenseNet169, on the PlantDoc dataset. Unlike many plant disease datasets, PlantDoc contains images captured in uncontrolled environments, simulating real-world conditions. Features extracted from the CNN models were subsequently classified using various DL and Machine Learning (ML) approaches, including fully connected (FC) layers, Extra Trees (ET), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Extreme Gradient Boosting (XGB), Categorical Gradient Boosting (CatGB), Light Gradient Boosting (LightGB), Logistic Regression (LR), and Random Forest (RF). The combination of ConvNeXt feature extraction with RF classification achieved the highest accuracy of 74.15%, outperforming existing works in this domain. Additionally, the approach obtained 77.27% precision, 74.15% recall, and a 74.51% F1-score. We have also incorporated 5-fold cross-validation to validate and support the evaluation results obtained on the test set. To interpret the model’s decisions, Local Interpretable Model-agnostic Explanations (LIME) were employed. This work presents an explainable hybrid DL and ML strategy for effective plant disease classification in realistic scenarios.

ABSTRACT. Tea is one of the classic beverages in Bangladesh. Since its demand and economic contribution are significant, our agronomists need to quickly find the diseases of tea and cure them. But the traditional process of finding tea diseases, such as recognizing the diseases by experts, is a very long process and difficult to find the type of disease. Furthermore, traditional methods for diagnosing these diseases depend on pathologists to identify symptoms simply by looking at them, which often results in a high rate of misdiagnosis and takes a long process to detect. For this purpose, a simple algorithm of the Convolutional Neural Network (CNN) has been proposed to eliminate the problem. Here, our proposed CNN model successfully classifies the six types of tea leaf diseases providing the accuracy of 99.23\%. To evaluate the results, we also compared our model with two transfer learning models, MobileNetV2 and Xception, which achieved accuracies of 91.23\% and 91.66\%, respectively, on the same dataset. Moreover, the proposed method can effectively classify the classes with a significantly reduced number of parameters, resulting in a lightweight model. In addition, the model performance was measured using some excellent output curves and evaluation metrics, precision, recall, F1 score, confusion matrix, ROC, PR curve and MCC value. In addition, the Grad-CAM visualization method was used to explain the outcome of the proposed CNN model.

ABSTRACT. Rice farming is a cornerstone of Bangladesh’s economic and food security, but it is constantly threatened by several leaf diseases, especially brown spot, blast, and bacterial leaf blight, which can drastically lower productivity and quality.Traditional manual disease detection methods are tedious, costly,and highly susceptible to error by people, highlighting the critical need for automated, accurate, and dependable alternatives. The present research utilizes sophisticated DL techniques by merging CNN with channel and spatial attention processes to create animproved disease detection system specialized for rice leaves.Utilizing a comprehensive dataset of 3,829 images encompassing six distinct conditions, the research employs meticulous data preprocessing and augmentation to enhance model training.Comparative analyses were conducted against established trans-fer learning models, including VGG16, ResNet50, EfficientNetB0,and MobileNet. Results demonstrate that the CNN enhanced with attention mechanisms achieved superior performance, attaining a test accuracy of 99.00%, with precision ranging from 0.97to 1.00, recall from 0.97 to 1.00, and F1-scores between 0.98and 1.00 across all classes. In contrast, transfer learning modelsEfficientNetB0 and MobileNet also performed commendably with test accuracies of 94.00% and 93.00%, respectively, along with high precision and recall metrics. However, models like VGG16and ResNet50 exhibited significantly lower performance, with test accuracies of 71.00% and 36.00%, respectively. Furthermore, the incorporation of Gradient-weighted Class Activation Mapping(Grad-CAM) gave useful visual explanations, indicating the specific sections of rice leaf images that impacted the algorithm’s predictions. This dual enhancement of performance and interpretability not only addresses critical gaps in agricultural AI but also paves the way for real-time deployment in continuous monitoring systems.

ABSTRACT. Bladder tissue refers to the layers of cells that line the interior of the urinary bladder, playing a crucial role in urine storage and elimination. Abnormal changes in this tissue can indicate the presence of bladder cancer, which requires accurate and timely diagnosis. Currently, bladder cancer diagnosis depends on invasive, expensive, and expertise-intensive methods such as cystoscopy, which may fail to detect subtle lesions. Deep learning offers a promising alternative, yet many existing methods struggle with limited annotated data, class imbalance, and difficulty in capturing both global context and fine-grained details from complex imaging modalities. This study introduces ViTXception, a novel hybrid deep learning framework tailored for accurate and robust classification of bladder tissue. The architecture combines a Vision Transformer (ViT) backbone for capturing global contextual features and an Xception backbone for extracting local discriminative details. These complementary representations are fused and processed through a regularized classification head for final prediction. Evaluated on the Endoscopic Bladder Tissue Classification Dataset—comprising WLI and NBI images across four classes (HGC, LGC, NST, and NTL), ViTXception achieves state-of-the-art performance, attaining 97.33% test accuracy and a Cohen’s Kappa of 0.970. Comprehensive comparisons demonstrate its superiority over conventional CNN and Transformer based models, positioning ViTXception as a non-invasive, scalable, and clinically reliable AI assisted solution for bladder tissue analysis.

ABSTRACT. Among all forms of cancer, lung cancer continues to account for the highest mortality, stressing the importance of precise subtype identification. Traditional centralized deep learning models are often restricted by strict data privacy requirements, such as those mandated by HIPAA and GDPR. To address this, we present a federated learning (FL) framework designed for the classification of lung cancer subtypes from histopathological CT scans, ensuring patient data privacy while managing non-IID data distributions common in real clinical settings. Our hybrid model integrates EfficientNet-B3 with a multi-head self-attention mechanism, enabling robust feature extraction across four categories: adenocarcinoma, large-cell carcinoma, squamous-cell carcinoma, and normal tissue. The system is implemented using the Flower framework with both FedAvg and FedProx strategies, and training is simulated across five clients under non-IID conditions. Across eight communica- tion rounds, the model demonstrates strong performance, with accuracy rising from 32.05% to 96.91% and loss dropping from 4.129 to 0.218. To strengthen clinical interpretability, we incorporate explainable AI methods: Grad-CAM++ pro- duces class-specific visual maps highlighting key pathological structures, while SHAP offers both image-level and cohort- level feature attributions. These explanations consistently align with expert pathology knowledge, reinforcing the reliability of the predictions. Overall, our findings confirm that federated deep learning can provide accurate, privacy-preserving, and interpretable diagnostic support for lung cancer classification.

ABSTRACT. One of the deadliest types of skin cancer, melanoma, can be successfully treated with early detection and precise dermoscopic skin lesion segmentation. However, exact lesion segmentation is difficult due to the wide range of appearances of the lesion, the lack of contrast with the surrounding skin, and the existence of aberrations such as hair and shadows. In order to improve feature representation, we supplement the modified UNet++ architecture with Squeeze-and-Excitation (SE) blocks, Atrous Spatial Pyramid Pooling (ASPP), and attention techniques called SEAA-UNet++ to create a strong deep learning-based segmentation framework. We use Test Time Augmentation (TTA) which evaluates several enhanced copies of the test image and aggregates their predictions to provide a more reliable and accurate final output, to further enhance the model’s generalization during inference. The model is trained and assessed on the ISIC 2017 dataset. According to experimental data, the suggested method greatly improves segmentation performance; after applying TTA, the test set showed an IoU of 74.56%, a Dice coefficient of 82.90%, and a precision of 93.11%. This illustrates how adding TTA to a well-thought-out network architecture improves robustness and establishes a new standard for melanoma segmentation performance.

ABSTRACT. Ophthalmic disorders such as diabetic retinopathy, age-related macular degeneration, and glaucoma are primary preventable causes of blindness. Early and precise diagnosis is critical to ensuring early therapeutic intervention and mitigating long-term ophthalmic and systemic complications. In this study, we propose a novel deep learning based hybrid framework, RetinaXNet, that integrates DenseNet201 with Multi-Head Self-Attention (MHSA) to enhance diagnostic performance and model interpretability in fundus image analysis. The hybrid framework integrates tailored preprocessing techniques to optimize retinal image quality and leverages explainable AI (XAI) through Gradient-weighted Class Activation Mapping (Grad-CAM). This is particularly to provide clinically relevant visualizations of disease-localized regions. Unlike prior models that have limitations in generalizability, poor interpretability, or excessive computational overhead, the proposed approach achieves a balanced trade-off between predictive accuracy, transparency, and efficiency. Experimental evaluation on benchmark datasets demonstrates a test accuracy of 99.26%, a cross-validation accuracy of 99.05%, an F1-score of 99.26%, and a Cohen’s Kappa coefficient of 0.9902. The performance highlights not only high predictive fidelity but also strong inter-rater agreement. Grad-CAM visual outputs align consistently with ophthalmologist-marked pathological regions for reinforcing the model’s potential for clinical adoption. Moreover, the lightweight design supports real-time deployment via a web-based interface, making it suitable for use in low-resource or point-of-care settings.

ABSTRACT. Accurate detection of kidney stones in axial CT images is essential for timely diagnosis and management, yet manual interpretation remains time-consuming and prone to inter-reader variability. To address this challenge, a deep learning–based explainable framework is proposed for binary classification of stone vs. non-stone cases. The dataset includes 3,364 original CT images and 35,457 augmented images, all clinically verified, ensuring both reliability and diversity for robust training. Three convolutional neural networks—AlexNet, DenseNet121, and MobileNetV2—were evaluated under a unified transfer-learning strategy with standard augmentation and early-stopping. To enhance transparency, Explainable AI (XAI) techniques such as class-activation/Grad-CAM heatmaps were employed, providing visual localization of decision-critical regions consistent with radiological findings. Experimental results demonstrate that the proposed MobileNetV2-based model achieved the highest classification accuracy of 99.72\%, outperforming AlexNet and DenseNet121 while maintaining computational efficiency. The integration of XAI confirmed that the model’s predictions align with clinically relevant stone features, thereby improving interpretability and clinical trust. These findings highlight the potential of lightweight CNNs combined with explainability to deliver fast, accurate, and interpretable decision-support tools for kidney stone detection in routine CT imaging, with applications in workflow triage, quality assurance, and future prospective validation.

ABSTRACT. Speech emotion recognition (SER) is an essential technology that allows machines to detect emotional cues from spoken language, significantly enhancing human-computer interactions (HCI). While most SER research emphasizes airconducted speech, bone-conducted (BC) speech provides a resilient alternative, especially in noisy settings where standard audio may fail. This paper introduces an end-to-end SER system based on the Wav2Vec2.0 transformer model, fine-tuned with the EmoBone dataset—a comprehensive, multi-national BC speech dataset featuring eight emotion categories collected from 29 speakers in 10 countries. Our method bypasses manual feature extraction by learning detailed contextual features directly from raw audio waveforms using self-supervised learning. The system combines a custom classification head with the pre-trained Wav2Vec2.0 encoder, ensuring precise and efficient emotion prediction. Evaluation results show that our approach attains an overall accuracy of 93% and a weighted average F1-score of 93% on the EmoBone dataset, markedly surpassing earlier best-performing techniques. Analysis using the confusion matrix demonstrates strong effectiveness in most emotion categories, especially in separating emotions with unique acoustic features. Nevertheless, the model encounters challenges in differentiating acoustically similar emotions like neutral-sad and fear-disgust pairs. These results underscore the promising capabilities of transformer-based architectures for BC speech emotion recognition and set a new standard for future studies in this area.

ABSTRACT. In recent years, understanding the emotional dynamics of patient-clinician interactions has emerged as a critical topic in healthcare research. With the Speech Emotion Recognition (SER), it has become possible to analyze the emotional status and expressions during the clinical conversations. The SER provides critical insights to enhance patient care, diagnostic precision, and therapeutic effectiveness. In this paper, we present a multimodal framework for Speech Emotion Recognition specifically designed for healthcare scenarios, integrating advanced transformer-based models including T5, BERT, and XLNet. Our proposed framework analyzes both audio and textual data, enabling comprehensive identification and visualization of emotions expressed by patients and healthcare providers. Audio recordings from interactions between patients and clinicians—including doctors and psychiatrists—are transcribed using the Whisper model, ensuring high transcription quality. We evaluated the framework's performance on an audio dataset comprising clinical conversations capturing a variety of emotional expressions relevant to healthcare contexts. Our experimental results demonstrate that our SER framework accurately identifies six primary emotional states, including Happiness, Anger, Fear, Sadness, and Surprise, as well as distinguishing between positive and negative sentiments. Among the evaluated models, T5 exhibited the highest confidence score at 89.12%, significantly outperforming RoBERTa (78.44%) and XLNet (36.02%) in capturing emotional content from clinical dialogues. These findings highlight the potential of SER to aid healthcare professionals by providing deeper insights into patients' emotional states, supporting communication, and improving understanding of patients' sentiment.

ABSTRACT. This study focuses on Speech Emotion Recognition (SER), a sub-field of artificial intelligence that aims to allow machines to recognize human emotions using speech. The objective was to compare how two deep learning models perform in SER: a Residual 2D Convolutional Neural Network (CNN) and a Deep Neural Network (DNN). The two models were trained using popular datasets of emotion-labeled speech, including RAVDESS, CREMA-D, TESS, and SAVEE. The data collected included audio samples that represented emotions like joy, anger, and sadness. The Residual 2D CNN was an architecture that used residual connections to facilitate efficient information flow and easier training. After training, the Residual 2D CNN achieved an accuracy of 85%, showing that the model was able to successfully recognize complex patterns in speech-related emotions. In comparison, the DNN, with its simpler architecture, achieved an accuracy of 74%, indicating challenges in recognizing nuanced emotional features. In conclusion, the study found that the Residual 2D CNN outperformed the DNN, suggesting the need for more complex models for reliable emotion classification. The findings also highlight opportunities to explore transfer learning and multi-modal emotion recognition by incorporating visual or textual data, which could further improve model accuracy.

ABSTRACT. Religious hate speech is an emerging issue in Bengali online environments and particularly malicious statements tend to provoke conflicts and interfere with societal peace. Yet, whereas Bengali is among the most spoken languages in the world, it has been observed that little work has been done to curate datasets and reliable techniques of automatically comprehending the difference between religious hate speech and regular discourse. This paper proposes a new Bengali text corpus of real-life instances of hate speech that are religiously motivated alongside neutral, common-place statements, in order to facilitate the beginning of bridging this gap. The manual positioning of each entry has been done so that supervised learning can make use of a reliable ground truth. All modern and standard methods were considered, including classic methods of machine learning, deep neural networks, and state-of-the-art models based on Transformer. The Bangla-BERT model showed the best performance in all of them with 98.89 precision and F1-score. These results demonstrate that domain-specific Transformers are useful in the under-resourced languages hate speech detection. This paper lays down a feasible foundation to subsequent studies and shows how automated systems can facilitate content moderation, safe online interactions, and informed policies. The study offers a real dataset and a tangible standard that will be crucial when creating culturally considerate AI tools that will be able to fight hate speech without infringing on freedom of expression within the Bengali linguistic environment.

ABSTRACT. Human Activity Recognition (HAR) is a key task in computer vision, with applications in surveillance, healthcare, and human-computer interaction. Traditional HAR systems using sensors like accelerometers and gyroscopes often face limitations in accuracy and adaptability. Vision-based approaches, enabled by high-resolution imaging, provide richer contextual information. In this work, we explore HAR using deep learning, focusing on pre-trained DenseNet201, Xception, and InceptionResNetV2 models with tailored modifications. We investigate alternative network topologies, data preprocessing strategies, and training procedures to enhance recognition robustness and accuracy. Additionally, we propose an ensemble framework that integrates the strengths of these models, further improving performance. Experiments on benchmark datasets, including Stanford 40 Actions, BU-101, and Willow, demonstrate the effectiveness and generalizability of our approach. Individually, DenseNet201, Xception, and InceptionResNetV2 achieved accuracies of 87.36%, 85.65%, and 84.00%, respectively, while the ensemble reached 90% on Stanford 40 Actions. These results highlight the potential of ensemble-based deep learning to advance HAR research and enable broader practical applications.

ABSTRACT. This paper addresses the problem of multilingual Speech Emotion Recognition (SER) in Bangla-English code- mixed speech with reference to a hybrid CNN-LSTM. A brand new dataset of 2,500 actual audio samples was recruited, pre- processed and augmented. The model incorporates spectral and temporal characteristics, and with a test accuracy of 78.92, the model also performs better than conventional techniques, includ- ing Random Forest (78.19%), K-Nearest Neighbors (53.73%), and Logistic Regression (28.68%). It exhibited excellent gener- alization of five emotion classes whose macro F1-scores were in the 0.75-0.83 range. Its increased precision and stability in the presence of noise highlights its practical use in real-world settings, and better application in healthcare, education, and customer care. This paper brings culturally inclusive affective computing systems by offering a powerful solution to code-mixed speech emotion recognition.

ABSTRACT. This study investigates the utilization of peanut red skin (PRS) food waste as a natural colorant and bio-functional agent for the sustainable dyeing and functional finishing of viscose fabric. Peanut red skin, rich in flavonoids and tannins, was processed to extract colorants for dyeing at various temperatures, pH, concentrations, and times. The dyeing was carried out with three different dye concentrations—20%, 40%, and 60%—in three different methods: pre- and post-mordanting with alum and without mordant. The fastness and functional performance of dyed fabrics were studied. The findings revealed that post mordant dyed samples achieved the highest color strength (K/S) value, followed by without mordant and pre mordant dyed samples, respectively. Though all three methods demonstrated outstanding UV protection ratings, post mordant dyed samples were rated the highest. All dyed fabrics exhibited good colorfastness to washing, perspiration, and water; however, the wet rubbing fastness was rated as moderate. The excellent UPF ratings of the dyed samples, regardless of the dyeing methods used, indicate the potential of valorizing PRS agrowaste in ecofriendly functional textile processing.

ABSTRACT. This research proposes an analysis of spent textile dyes Maxinol Red GRL adsorption for the wastewater textile industries by using activated charcoal derived from sawdust, a readily available and lignocellulosic waste material. Sawdust underwent a two-step chemical activation process using H2SO4 and ZnCl2 to form high-surface-area adsorbents. We conducted batch adsorption studies to analyze the effects of pH, contact time, temperature, adsorbent dosage, and initial dye concentration on adsorption. FTIR Spectroscopy confirmed the activation and adsorption of Maxinol Red GRL onto the charcoal, revealing changes in the material’s structure and functional groups, which significantly increased the material’s adsorption capabilities. Compliance with the Beer-Lambert law in the 20-150ppm concentration range was confirmed by UV-Visible Spectrophotometry which yielded a molar extinction coefficient of ε was 1.29 x 10-2 L.mg-1.cm-1. Maximum removal of the dye was achieved at 98% at pH 2, 10 hours contact time, 100ppm initial concentration with a 1 g adsorbent dosage. Thermodynamic studies showed that adsorption is an endothermic process. These studies demonstrate the potential of the activated charcoal to be synthesized from sawdust to be employed as an adsorbent material for the removal of dyes from wastewater in a cost-efficient, environmentally friendly manner, making a case for activated charcoal’s use in the circular economy of resources and sustainable wastewater treatment processes.

ABSTRACT. In Apparel washing facilities, hydrogen peroxide, an eco-friendly oxidizing bleaching agent, is frequently utilized as an affordable chemical for color removal and finishing. It provides an environmentally friendly substitute for bleaches that contain chlorine because it breaks down into water and oxygen. It has become crucial for technologists to modify denim clothing in order to meet the demands of current fashion trends since youths today desire faded or vintage denim. This study evaluates how the physical and surface characteristics of denim dyed with indigo made of 100% cotton are affected by hydrogen peroxide bleaching at three liquor ratios(1:5, 1:8, and 1:10).After desizing, bleaching was done for 20 minutes at 45 °C, then neutralization and softening were done accordingly. Performance was evaluated using GSM, tensile and tear strength, pilling resistance, dimensional stability, and color fastness (wash, rubbing, perspiration, and saliva). GSM loss was greatest at 1:5, with reductions ranging from 4.56% to 5.53%. The fastness properties of the sample 1 was quite good than the other samples but in case of tensile and tear strength, the highest losses occurred at 1:5 indicating stronger oxidative deterioration. All of the samples exhibited shrinkage in both directions, and the pilling resistance increased greatest at 1:10 (rating 4.0). It has been monitored from the experimental data that the bleaching at a 1:5 liquor ratio results in the greatest fading and surface alterations but also the greatest strength loss ,1:8 ratio provides balanced fading with a minimal loss of strength and moderate finish and more durability can be observed at the 1:10 ratio since it reduces strength loss and preserves more color.

ABSTRACT. The discharge of untreated dyeing wastewater from the textile industry poses significant environmental challenges because of its high load of pollution, including toxic dyes, chemicals, heavy metals, etc. Traditional wastewater treatment method often fails to treat the high load of pollution, and the necessity of advanced treatment of degrading waste came from this reason. The study investigates the application of the photocatalysis process by using Titanium dioxide (TiO2) nanoparticles for treating polyester-cotton blend dyeing wastewater. The TiO2 nanoparticle is formed via a green synthesis process using cinnamon and lentil powder in the laboratory. The research focuses on the efficiency of the photocatalysis process for degrading dyes under sunlight/ UV light. The study highlights the efficiency of the TiO2 nanoparticle-based photocatalysis process as a sustainable and efficient approach to treating dyeing wastewater. Future work is recommended to explore the scalability of this process and the development and recovery of TiO2 nanoparticles for reuse.

ABSTRACT. The research deals with the industrially viable development and optimization of an environmental-friendly, biodegradable packaging material involving a composite blend of banana fiber-water hyacinth-textile dust-neem leaves-natural wax for proposing an alternative to conventional plastic packaging (Figure Figure 1). Raw materials from renewable and waste-based sources were converted through a series of processes including drying, alkaline treatment, bleaching, blending, sheet production and coating with natural wax to enhance the shelf-life and performance. The Ball Burst Strength, Water Repellency, Oil Repellency and Chemical Resistance of the material were assessed in extensive tests. Results suggested that neem leaf increased the burst strength by 15.38% and reduced water absorption tendency up to 16.67% showing structural reinforcement as well moisture resistance respectively. The results of the cost analysis showed that production costs are ranged in between 300–400 BDT per kg; this makes this material an ecofriendly and less expensive one. This potentially makes it the ideal replacement for both plastic and paper-based packaging, which should result in less waste, more effective use of resources, and greater support for the principles that underpin a circular economy.

ABSTRACT. The profound influence of natural and synthetic polymeric sewing threads on the stitch performance of polyester fabrics is thoroughly analysed in this research which constitutes a vital step in garment manufacturing. This study focuses on the interaction between needle and bobbin thread pairs at varying stitch densities and their effect on seam strength and efficiency. A set of twelve samples with different thread combinations and stiches per inch (SPI) levels (8, 10 and 12) was prepared using a single needle lock stitch machine following ASTM D1683-04 standards. Extensive testing revealed that the thread type used, along with SPI, significantly impacted seam performance. Remarkably, at an SPI of 12, cotton threads in both needle and bobbin achieved the highest seam strength, surpassing even combinations involving polyester threads. However, as SPI decreased, combination of cotton and polyester threads demonstrated superior strength and efficiency compared to identical thread pairings. These findings challenge conventional assumptions about thread selection and underscore the importance of optimizing thread compatibility and stitch density to enhance garment durability and quality. This study offers valuable insights for garment manufacturers aiming to improve seam performance and optimize production processes for enhanced sustainability and cost-effectiveness.

ABSTRACT. Perovskite solar cells have emerged as a leading technology in the field of photovoltaics due to their outstanding optoelectronic properties, high power conversion efficiency (PCE), and electronic properties. However, the widespread adoption of PSCs is hindered by the use of toxic and environmentally hazardous elements, particularly lead. In this study, we present an environmentally friendly dual-absorber solar cell incorporating MASnI3 and Ba(Zr0.95Ti0.05) perovskite materials. We investigate the effect of different Hole transport layer (HTL) components on PSC performance by employing the transition metal dichalcogenide, WS2, as an Electron transport material, owing to its remarkable electron transport properties, stability, and adaptability, with the aim of enhancing the device's PCE. Key device parameters such as shallow acceptor density, defect density, temperature, and absorber layer thickness are optimized to enhance the overall PCE of the proposed PSC structure. The optimized configuration Au/CuO/MASnI3/Ba(Zr0.95Ti0.05)/WS2/FTO achieved a maximum PCE of 38.86%, with a short-circuit current density (Jsc) of 34.70 mA/cm², an open-circuit voltage (Voc) of 1.270 V, and a fill factor (FF) of 87.84%. The solar cell has achieved the best efficiency with absorber layer thicknesses of 1 µm for MASnI3 and 0.01 µm for Ba(Zr0.95Ti0.05), by using the back-connect Au (gold). Furthermore, the findings indicated that both efficiency and device workability had a slight decline above a temperature of 300K. These findings highlight the potential of MASnI3/Ba(Zr0.95Ti0.05)-based PSCs as a promising lead-free alternative for high-efficiency, environmentally sustainable photovoltaic applications.

ABSTRACT. In today’s advanced world, where intra-city public transportation has become an issue to cope with massive populations in major capitals, electricity-driven metro rail appears as a blessing. But this advancement comes with a very high energy demand, and a large portion of that is supplied from conventional sources, emitting a large amount of CO2. This paper focuses on the Dhaka Metro Rail system, which has a daily electricity demand of 160 MW. This study investigate the opportunities to install a grid-connected rooftop solar PV system specifically at Dhaka metro rail depots, aimed to reduces the net grid draw and produces clean energy, which is a part of the government’s NDC to the Paris agreement. But a larger PV injection to the grid may impact the grid stability, requiring grid feasibility analysis. Through our simulation, we analyze the grid compatibility of the proposed system along with techno-economic and environmental assessment. PVsyst simulation has been conducted to demonstrate the possible generation, performance ratio, and losses. Moreover, the grid interactions were analyzed based on the Voltage fluctuation at the common coupling point (PCC) due to PV injection and load profile Vs. grid injection analysis. Finally, to quantify the relation between all the parameters like PV output, grid-draw, load, and voltage at VCC, a correlation coordination matrix has been evaluated. The system produces 14.19 GWh per year, with a specific output of 1415 kWh/kWp/year, which causes the overall grid draw to lessen by 24.33% and a performance ratio of 87.23%, indicating high efficiency and low losses; it will also reduce 160,851.66 tons of CO2 emissions in Bangladesh over 25 years. Findings emphasize how significantly the proposed system can minimize the operational energy cost and carbon emission, and the stability impact on the grid that must be considered for large-scale implementations. This research collectively affirm the viability of the grid-connected rooftop PV system for the metro rail depot, creating scope for further potential enhancements to lower grid dependency.

ABSTRACT. South Asian countries face high levels of air pollution, including outstanding particulate matter (PM2.5), which has a substantial impact on solar energy output. PM2.5 dust particles build up on solar panels, lowering efficiency by blocking sunlight absorption and causing long-term surface deterioration. The growing industry, urbanization, and seasonal fluctuations, such as dust storms, aggravate this problem. The loss in solar panel performance caused by particulate pollution presents a significant obstacle to the growth of renewable energy in nations such as India, Bangladesh, Pakistan, and Nepal. To address this issue, improved panel cleaning techniques, sophisticated coatings, and governmental interventions must be used to reduce air pollution at its source. This study found that PM 2.5 levels are higher in cities with high population density, which can harm photovoltaic cells and reduce their production compared to other locations. He also investigated the output of photovoltaic cells in five of the most populous cities in South Asia to determine the impact of PM2.5 particles on photovoltaic cells in densely populated areas and found some long-term effects.

ABSTRACT. Dust buildup is a significant issue for solar panels in Bangladesh, which significantly reduces their effectiveness. Because there isn't much rainfall throughout the lengthy dry season, which runs from November to March, these particles can accumulate on the panels and form a thick coating that blocks light. This dust might get cemented to the surface due to Dhaka's high humidity, which makes removal much more challenging and significantly reduces current and power production overall. This study looks at how PV performance is impacted by dust collection in different parts of Dhaka, Bangladesh. In areas such as Uttara, Shahbag, Jatrabari, Mirpur-1, and Mohakhali, PV panels were installed and monitored for six months. This study investigates the impact of various dust types and loads on a solar panel's electrical performance, with a particular focus on the maximum power point current (I mp), voltage (V mp), and power (P max). A solar panel exposed to 2-gram and 4-gram dust loads was used for experimental observations. With a coal dust load, I mp decreases from 1.5 A to 0.6 A, demonstrating that an increased dust load typically has an inverse connection with current production. The investigation of P max revealed a wide range of consequences, depending on the dust type, from an unanticipated power gain with specific soil types to a 1610% reduction induced by pigeon excrement; however, the influence on V mp was less consistent. By influencing the current and total power generation capacity, the results show that dust type and quantity are essential determinants in solar panel degradation. The findings highlight how crucial it is to create efficient dust mitigation plans, especially in areas with high particulate matter concentrations. In the dry season, the effect was more pronounced. For around 120–125 days, panels with capacities ranging from 10W to 250W were tested without being cleaned. The performance ratio decreased for all panels. One constant factor lowering PV efficiency was determined to be dust. The analysis also aids in determining which zones lost the most efficiency.

ABSTRACT. Access to clean drinking water remains a critical challenge in water scarce regions. We present a portable Atmospheric Water Generator (AWG) using Peltier technology, integrated with IoT (Blynk app) for real time monitoring of temperature, humidity, and water levels. The system includes solar backup to ensure operation without grid power. Tested in various environments, the AWG produced up to 2 L/day of potable water, while remaining lightweight (≤ 1kg) and easy to operate. Its portability, affordability and real time monitoring make it a practical solution for outdoor, emergency and remote applications where clean drinking water accessibility is limited.

ABSTRACT. The following study focuses on breaking the Shockley-Quisser limit of perovskite solar cell (PSC) efficiency by combining organic FACsSnI₃ and inorganic Cs₂PtI₆, two metal halide, lead-free, three-dimensional absorbers, into a planar (n-i-p) structured bilayer perovskite solar cell model. The architecture is Glass/FTO/SnS2/FACsSnI3/Cs2PtI6/HTL/Au. We conducted a study on power conversion efficiency (PCE) by varying the bandgap, defect density, and electron affinity using four different hole transport layers (HTLs): CuO, CuI, Spiro-OMETAD, and Mo. The results are based on the photovoltaic properties of PSCs, specifically the power conversion efficiency (PCE), fill factor (FF), open circuit voltage (Voc), and short circuit current (Jsc) of the architecture, which were analyzed numerically using SCAPS-1D simulation software. We performed the process with an air mass of AM 1.5 G, illumination of 1000 W/m2, and temperature of 300 K. Variations of HTL have allowed this study to investigate PV characteristics of the absorbers. We’ve also used carbon (C) as the back metal contact.The Glass Substrate/SnS2/FACsSnI3/Cs2PtI6/Cu2O/C model we looked at worked best for our needs. It had an efficiency of 47.55%, a fill factor (FF) of 83.08%, a short circuit current density (Jsc) of 48.407586 mA/cm², and an open circuit voltage (Voc) of 1.1824 V. The glass substrate/SnS₂/FACsSnI₂/Spiro-OMETAD/CuI/c worked the best, with an efficiency of 38.20%, a FF of 84.77%, a Jsc of 33.009704 mA/cm², and a Voc of 1.3650 V.

ABSTRACT. In today's digital landscape, preventing security breaches presents significant challenges due to the rapid increase in computer traffic, making intrusion detection a paramount concern for network and computer security. The growing frequency and complexity of attacks jeopardize accessibility, confidentiality, reliability, and integrity, making it vital to implement effective intrusion detection systems (IDS). Traditional rule-based systems often struggle to address diverse and emerging cyber threats, leading to the successful implementation of machine learning algorithms, particularly random forest methods, in this field. In contrast to static rule-based models, machine learning (ML) approaches can adapt over time by learning from new threat patterns. This adaptability is crucial for identifying zero-day vulnerabilities that consecutive systems may miss. By looking at various real and fake events and spotting patterns like network type, port number, source, and destination, the random forest algorithm has shown to be very good at finding both familiar and new attacks. Its impressive accuracy rate of 99.78\% underscores its significance as a powerful and versatile tool for modern intrusion detection systems, essential for safeguarding computer networks against malicious activities and unauthorized access. Furthermore, its robustness against overfitting makes it particularly suitable for dynamic and heterogeneous network environments. As cyber threats evolve, the adaptability and scalability of random forest-based IDS frameworks continue to make them a preferred choice for both enterprise and industrial applications.

ABSTRACT. False data injection attacks (FDIAs) pose a serious threat to smart grid reliability by covertly corrupting measurement streams and misleading state estimation. To counter stealthy, Jacobian-consistent attacks that evade conventional bad-data detection, this paper introduces a hybrid deep learning detector. The architecture synergistically combines: (i) a graph neural network (GNN) to explicitly model grid topology, (ii) an attention-augmented recurrent module to capture evolving temporal patterns, (iii) a compact autoencoder (AE) for unsupervised anomaly sensitivity, and (iv) a deep belief network (DBN) for robust classification. A conditional GAN is employed exclusively during training to enrich the minority (attack) class, improving calibration without complicating inference. Using the realistic SimBench benchmark, our method achieves 98.05% precision and PR–AUC of 0.776, enabling low–false-alarm operation critical for control-room trust. While recall remains moderate (48.48%) due to near-nullspace attack overlap, the framework demonstrates how joint spatial–temporal modeling enhances resilience in human-centric, Industry 5.0 energy systems.

ABSTRACT. The rapid proliferation of heterogeneous IoT devices presents serious cybersecurity challenges because such IoT devices are characterized by limited computational capability and weak in-built protection mechanisms. This paper proposes dependable and scalable intrusion detection solutions using hybrid deep learning and ensemble machine learning techniques for robust threat detection across diverse IoT environments. Specifically, this work designs two models: (i) a P-ResNet-based Deep Transfer Learning Intrusion Detection System to enhance detection with limited labeled data by leveraging pretrained representations; and (ii) a hybrid CNN→Random Forest model that integrates deep feature extraction with ensemble decision-making for superior classification accuracy. Experimental results on the TON IoT dataset show that the proposed CNN→RF hybrid achieves state-of-the-art performance with 99.1% accuracy, outperforming standalone CNN and DTL approaches, while the P-ResNet DTL model offers strong performance, with 96.18% accuracy, at reduced inference time, hence dependable in constrained environments. Overall, the proposed models offer high accuracy, adaptability, and efficiency, thus validating their suitability for dependable intrusion detection in heterogeneous IoT networks.

ABSTRACT. Intelligent surveillance systems for violence detection play an important role in maintaining public safety. However, many violence detection systems suffer from a lack of transparency in decision-making. Explainable artificial intelligence (XAI) is rapidly advancing, particularly in violence detection surveillance systems. Our study proposes a custom ResNet-50 deep learning model with an explainable artificial intelligence technique, Shapley Additive explanations (SHAP), to detect violence in surveillance videos. It includes a meticulous preprocessing pipeline, such as equal time-domain sampling and spatial-intensity normalization. The dataset consisted of 300 videos, with 150 videos in each class. The dataset was divided into training, testing, and validation datasets. ResNet-50 extracts features to detect fights and nonfight actions in videos. The custom model achieved 98% accuracy, 93.83% precision, 97.97% recall, and 95.86% F1-Score. SHAP’s GradientExplainer provides transparency and offers clear visual insights into the model’s decision-making, highlighting critical pixel regions. The results show that our study demonstrates high performance and robust generalization for detecting violence and, as a result, improved model accuracy by reducing prediction error and increasing trust through model transparency. This improves public safety and prevents the risk of incorrect accusations by the security authorities. Moreover, the proposed method achieved better performance than the existing methods.

ABSTRACT. In this paper, we propose a framework for texture classification that improves upon the traditional local directional ternary pattern (LDTP) by integrating a performance-enhanced directional mask (PEDM). This innovative approach addresses the limitations of existing texture analysis techniques, such as time-consuming, computational inefficiency, and limited robustness, by using optimized directional masks and Gaussian second-derivative filters to achieve higher noise resilience and efficiency. PEDM captures both structural and directional texture details, enabling robust encoding of local spatial patterns with low time complexity. The resulting proposed RTC-LDTP algorithm generates distinct upper and lower pattern codes, which are combined into a discriminative feature vector for classification. Employing a Support Vector Machine (SVM) classifier, the proposed method demonstrates significant performance improvements across multiple datasets, including KTH-TIPS, KTH-TIPS2-b and CUReT achieving competitive accuracy and reducing computational overhead. This study establishes PEDM as a robust solution for real-world texture classification, setting new benchmarks in accuracy and efficiency.

ABSTRACT. Aleatoric uncertainty, arising from inherent noise in data, is a critical challenge in deploying machine learning models for healthcare applications. Even in clean electronic health records (EHR), a baseline level of aleatoric uncertainty is unavoidable due to measurement errors and variability in patient populations. In this work, we demonstrate that the injection of Gaussian noise, commonly used for privacy preservation and regularization, further amplifies aleatoric uncertainty in tabular healthcare datasets. We systematically evaluate state-of-the-art uncertainty quantification (UQ) methods, including Bayesian neural networks and deep ensembles, on disease prediction tasks under controlled Gaussian perturbations. Our findings reveal that most existing models underestimate the compounded uncertainty introduced by noise injection, potentially leading to overconfident predictions in high-noise regimes. These insights highlight the need for noise-aware UQ frameworks to ensure the reliable deployment of machine learning systems in clinical settings.

ABSTRACT. Lung adenocarcinoma (LUAD) remains a leading cause of cancer mortality worldwide. Robust gene markers are crucial for early detection and enhanced therapeutic strategies. In this study, we developed a pipeline that combines the Wilcoxon rank-sum test with a log2 fold change (log2FC) filter to identify significant genes of LUAD using TCGA-LUAD RNA−seq data. Subsequently, identified significant genes were mapped to Entrez Gene IDs, and unmapped entries were considered as intergenic non-coding RNAs (lincRNAs). KEGG pathway enrichment analysis was applied to the mapped genes, which enabled the identification of the key candidate genes. We identified 15 candidate genes, including BAK1, CDK4, CDKN2A, E2F1, E2F2, E2F3, EGF, ERBB2, GADD45G, JAK3, KRAS, MET, PRKCG, RET, and TGFA. Afterwards, survival probabilities of the candidate genes were assessed using Cox proportional hazards (CoxPH) model, LASSO-penalized Cox regression, and Kaplan-Meier (KM) estimator. LASSO-Cox regression refined the candidate genes into a six-gene signature (BAK1, CDKN2A, MET, PRKCG, RET, TGFA) that improved risk stratification by reducing overfitting. KM analysis on these candidate genes successfully separated patients into high and low-risk groups. Infine, the findings of this study can have a great impact on clinical treatment for patients with LUAD, and further research on lincRNA can open new doors both for clinical treatment and in the early detection of cancer patients.

ABSTRACT. The rapid advancement of speech synthesis and voice cloning technologies has led to the emergence of highly realistic deepfake voices, posing serious threats to security, privacy, and trust in digital communication. In this study, we propose a Convolutional Neural Network (CNN)-based framework for detecting deepfake voices across four distinct categories: Male Real, Male Fake, Female Real, and Female Fake. A primary dataset was constructed using real and synthetic voice samples from both male and female speakers. Mel-Frequency Cepstral Coefficients (MFCC) were extracted as acoustic features to capture the spectral and temporal properties of speech signals. The proposed CNN model, trained with Adam optimization and categorical crossentropy loss, achieved a remarkable accuracy of 99.33\%, with precision, recall, and F1-scores consistently above 0.99 for all classes. The confusion matrix demonstrates the robustness of the classifier, with minimal misclassification across categories. These findings highlight the effectiveness of deep learning-based acoustic feature modeling in distinguishing authentic speech from synthetic manipulations. The proposed approach contributes to the growing field of audio forensics and provides a reliable method for safeguarding against voice spoofing attacks in security-critical applications.

ABSTRACT. This paper proposes a reduced switch 23-level cross-switch multilevel inverter that integrates symmetric and asymmetric submodules to deliver high-quality output with the reduced hardware complexity. The topology uses only four isolated DC voltage sources, and nine unidirectional switches arranged in a cross-switching structure, eliminating the need for bidirectional switches and H-bridge inverters. Compared to conventional topologies, the proposed structure minimizes the voltage stress and the total number of switches, enhancing the reliability, efficiency and thermal performance. MATLAB/ Simulink simulations validate the effectiveness of the proposed inverter with three load types (purely resistive, resistive and inductive, and purely inductive load) under five loading conditions with and without a small LCL filter. Results demonstrate the excellent output voltage quality with a total harmonic distortion (THD) of 3.60% for a resistive load, and 1.13% for an inductive load with a compact LCL filter. The proposed inverter topology meets IEEE std. 519, and offers a promising solution for cost-effective, high-performance medium-voltage power conversion applications.

ABSTRACT. This paper focuses on the design of an inverter utilizing a center-tapped transformer topology. In today's power industry, energy generation is one of the most vulnerable aspects due to inefficiencies in conversion processes. Conventional power sources, such as diesel motors, produce significant carbon emissions and pollute the environment extensively. In contrast, solar energy, as a renewable source, is experiencing rapid growth, particularly in DC-to-AC inverters for agricultural applications. Solar grids can effectively compete with other power generation industries. This paper proposes a 750W inverter capable of driving a 1HP single-phase induction motor, offering two distinct operational modes. The designed inverter minimizes the number of switches, reducing conduction losses and component count. Furthermore, it delivers a 220V single-phase output, matching the power quality of conventional grids, with a center-tapped transformer providing a rated output current of 3.41A and an efficiency of 98.68%, while maintaining very low total harmonic distortion. Additionally, the inverter can step up or step down the output voltage based on demand.

ABSTRACT. This paper illustrates the purpose and implementation of a Neutral Point Clamped (NPC) converter incorporating a novel Pulse Width Modulation (PWM) technique aimed at improving output quality, switching efficiency. The proposed modulation strategy introduces an optimized switching sequence that effectively reduces total harmonic distortion (THD), with a voltage THD at 22.38% and a current THD at 1.65%. Both enhancements meet the harmonic distortion requirements of IEEE-519. These enhancements are attributed to strategic modifications in the modulating signal. Furthermore, the proposed method also exhibits consistent THD performance across a wide range of switching frequencies and modulation indices. Simulation results confirm the effectiveness of the approach, demonstrating superior THD characteristics compared to conventional PWM techniques, including Sinusoidal PWM (SPWM), Third Harmonic PWM (THPWM), Conventional Space Vector PWM (CSVPWM), and Bus-Clamping PWM (BCPWM). The proposed method's ability to enhance the performance of NPC converters in load-connected systems is underscored by this study, ensuring better power quality and extensive control.

ABSTRACT. Worldwide the use of electric vehicles (EVs) as an alternative to fossil-fueled cars is growing. If EV deployment is not effectively integrated and handled, it might pose problems for the grid. And the vast amount of empty battery storage in millions of EVs is utilized for auxiliary services to the grid, these difficulties might be turned into possibilities. This paper proposes the development of a bi-directional battery charging stations for the electric vehicles (EVs) with a focus on grid-to vehicle (G2V) and vehicle-to-grid (V2G) technologies. As the conventional batteries of EVs are charged with Direct Current (DC), which is supplied by a battery charger that converts the sinusoidal current and a unitary power factor from the power grid when the G2V operation mode is in effect. The ability to send energy from the batteries back to the power grid during the V2G operation mode contributes to the stability of the electrical system. In this system took in a Phase Locked Loop (PLL), which ensures accurate power transfer by synchronizing the frequency with the grid voltage, and Proportional Integral (PI) maintains the stable charging and discharging process by regulating the current and voltage of the charger. This proposed system bi-directional charger contains one AC-DC converter and a DC-DC buck-boost converter. For grid-to-vehicle applications, the battery of EVs is charged in buck mode, and for the vehicle-to-grid applications, it is discharged by boost mode. This study is carried out in MATLAB/Simulink R2022a.

ABSTRACT. High efficiency, low total harmonic distortion (THD), and near-unity power factor are essential for low-voltage battery charging systems to satisfy modern power quality requirements. In this study, an isolated totem-pole power factor correction (PFC) integrated buck converter has been designed and analyzed for low-voltage battery charging applications. The proposed topology combines a bridgeless totem-pole PFC front end with an isolated buck DC–DC stage to provide galvanic isolation, regulates constant-current/constant-voltage output, and improves power quality. A closed-loop control scheme employing PI controllers regulated both input and output parameters under steady-state and dynamic operating conditions. The system has been designed for an input range of 180–260 V AC and deliveres 31.6 W to the load. Simulation results showed that the converter achieved a power factor above 0.99801 and THD below 4.90%, complying with IEEE Standard 519. The converter demonstrates stable operation under input voltage and load variations, making it suitable for low-to-medium power applications such as UPS, LED lighting, security camera systems, and portable electronics.

ABSTRACT. This paper studies distributed generation (DG) siting, sizing, and Volt–VAR dispatch on a modified IEEE-57 system using Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and a Hybrid GA–PSO scheme. The goal is to cut technical losses and stabilize voltages while meeting IEEE 1547 limits, thermal ratings, and the 0.95–1.05 p.u. band. The test case removes the Bus-1 load and fixed shunts at Buses 18, 25, and 53, yielding 1195.8 MW / 319.4 MVAR on a 100 MVA base (vs. 1250.8 MW / 336.4 MVAR). Five DG technologies are modeled with capability-limited Volt–VAR droop; total pen etration is constrained to 12–25% of active load (achieved: PSO 24.1%, GA 18.5%, Hybrid 19.0%). The Hybrid method delivers the lowest losses active 22.195 MW (25.03%), reactive 76.906 MVAR (35.69%), and apparent 80.045 MVA (35.02%) followed by GA (22.954 MW, 78.209 MVAR, 81.507 MVA) and PSO (26.151 MW, 85.636 MVAR, 89.540 MVA). Voltage profiles tighten from 0.9315–1.0661 p.u. (base) to in-band ranges: PSO 0.9550–1.0500 p.u., GA 0.9560–1.0490 p.u., Hybrid 0.9542 1.0480 p.u. In the context of Industry 5.0 and Sustainable Tech nology, the framework uses AI-driven optimization to augment engineering decisions, raises hosting capacity through stability aware placement (L-index screening) and realistic VAR capabil ity, and converts efficiency gains into avoided energy waste and lower indirect emissions. Results show that effective siting with Volt-VAR support not higher penetration alone delivers superior performance under identical constraints, with the Hybrid method providing the most balanced loss-voltage outcome.

ABSTRACT. Microgrids are essential for integrating renewable energy into modern power systems, yet their stability is challenged by load fluctuations and generation intermittency. To address this, two advanced control strategies, Linear Quadratic Regulator (LQR) and Model Predictive Control (MPC) are optimized using the Ant Colony Optimization (ACO) algorithm. A discrete-time state-space model of the microgrid is employed for controller design, with performance evaluated through impulse and step response tests. Results show that ACO-tuned MPC achieves 0% steady-state error, 5.38% overshoot, and 5 ms settling time, outperforming the ACO-tuned LQR, which records 3.28% steady-state error, 7.42% overshoot, and 35 ms settling time. These findings confirm the effectiveness of ACO-based MPC in improving microgrid's dynamic response and stability.

ABSTRACT. This study examines the techno-economic viability of using the Padma Bridge, the biggest infrastructure bridge in the nation, to generate hybrid solar-wind energy. In order to capture both solar radiation and vehicle-induced wind flows and natural wind flow, the proposed approach integrates vertical-axis wind turbines implanted in the bridge's the middle divider with photovoltaic panels connected on the top of the LED streetlight poles. The hybrid model's yearly generation has been estimated at 3,432,655 kWh using HOMER Pro software. It has a 99.3% renewable fraction, a levelized cost of energy (LCOE) of $0.0181/kWh, and a 7.3-year payback period. An energy output of 1,456 MWh/year, an LCOE of $0.0225/kWh, a payback period of 4.3 years, and a performance ratio of 95.0% would be forecasted by the PVsyst simulation system. The system may decrease annual emissions by about 14,469 kg of CO2, 62.7 kg of SO2, and 22.9 kg of NOx when compared with conventional grid power. Based on the results, which show a 12% internal rate of return and a 9% return on investment, this hybrid design is technically feasible, economically appealing, and environmentally advantageous for coastal bridge applications. The proposed design demonstrates how Bangladesh's massive bridges could serve as smart energy corridors, addressing regional sustainability issues and advancing the country's renewable energy goals.

ABSTRACT. Optimizing microgrids with renewable sources is a challenging due to high initial costs, operational constraints, and uncertain generation. This paper presents a resource-efficient stochastic optimization method based on interior-point method (IPM). The approach is applied for optimal sizing and dispatch of distributed energy resources and storage in microgrids. To address this, a nonlinear cost-based model with two-stage optimization is formulated under varying renewable penetration and reliability levels. Uncertainties in solar, wind, and demand are modeled using beta distribution. Results show that a renewable penetration of 55% reduces the carbon emissions by 42.2% with only an 8.53% cost increase.

ABSTRACT. The global push towards decarbonization and energy security is hampered by the intermittency of renewable sources and the protection complexities of DC microgrids, necessitating integrated solutions that ensure both reliability and resilience. This study aims to design and assess a DC hybrid microgrid that seamlessly integrates solar photovoltaic (PV) and fuel cell technologies, with a dedicated fault detection and isolation (FDI) system, to overcome these challenges. A comprehensive model was developed in MATLAB/Simulink, featuring a PV array with Maximum Power Point Tracking (MPPT), a Proton Exchange Membrane Fuel Cell (PEMFC), an automated switching controller for energy management, and a microprocessor-based FDI algorithm for rapid fault identification. Simulation results demonstrated stable operation, with the PV system supplying 118 kW and the fuel cell providing 26 kW to the DC bus. The FDI system successfully identified and isolated faults within milliseconds, preventing system-wide collapse and maintaining voltage stability. The proposed system proves to be a robust and sustainable solution, effectively mitigating renewable intermittency through hybridization and enhancing system security with a rapid protection scheme, making it a viable model for reliable power supply in remote and critical applications.

ABSTRACT. The extensive adoption of electric vehicles (EVs) is hindered by challenges associated with traditional conductive charging, including connector wear, user inconvenience, and grid dependency. This paper presents the design, implementation, and performance analysis of a novel prototype system that integrates solar energy with both static and dynamic wireless charging to address these limitations. The system utilizes inductive power transfer (IPT) technology, where energy from a solar-charged battery bank is delivered wirelessly through transmitter coils to a receiver coil on an EV. A prototype was built using microcontrollers for control and monitoring, achieving a static charging efficiency of up to 80% with a 10W solar input. Dynamic charging experiments demonstrated the feasibility of in-motion power transfer, with analysis showing a potential 30-40% reduction in required stationary charging time for EVs. This study demonstrates a proof-of-concept for a sustainable and convenient charging solution that can alleviate range anxiety and promote renewable energy use in transportation.

ABSTRACT. Maximum Power Point Tracking (MPPT) plays a crucial role in extracting optimal power from solar photovoltaic (PV) systems in extreme weather conditions. Traditional MPPT algorithms often suffer high oscillations, slow convergence time, and low performance during rapid environmental fluctuations. This research presents an Artificial Neural Network (ANN)-based MPPT technique for modeling and simulation in MATLAB/Simulink environment to observe these limitations behavior. The proposed model is trained with long term datasets (PVGIS) with the parameters of Global Horizontal Irradiance (GHI), and Ambient Temperature (Tamb). To predict the optimal duty cycle (D), voltage and current inputs data are used, which enable fast and accurate tracking of maximum power point (MPP). An extensive solar PV model is developed in MATLAB/Simulink, that incorporates long term trained ANN- based MPPT controllers with standard solar PV features. The simulation results with long term average data and rapidly changing environmental conditions demonstrate that the ANN-based MPPT model outperforms conventional models in terms of accuracy, oscillations, tracking speed, and overall efficiency. Proposed model results were compared with traditional P&O MPPT model. The study ensures the potential of the ANN-based MPPT controllers for optimal power harvesting for next-generation solar PV systems and provide directions for researchers and engineers of hardware implementation.

ABSTRACT. In the sanitary ware ceramics industry, visual quality inspection is critical to maintaining product integrity and minimizing defects. Manual inspection, though commonly employed, is labor-intensive and unsuitable for large-scale production due to its inefficiency and susceptibility to human error. The unique challenges posed by the intricate surface geometries and reflective glazes of ceramic products further complicate defect detection using traditional computer vision techniques. To address these issues, we introduce VisuCeram, a novel publicly available dataset specifically designed for defect detection in sanitary ware ceramics. The dataset encompasses 3,265 high-resolution images across seven defect categories, including surface cracks and glaze imperfections. To evaluate the dataset, we conducted experiments using various lightweight YOLO-based models. Among these, YOLOv4-light achieved the best performance with a mean Average Precision (mAP) of 87.30%, followed by YOLOv7-light (46.56%), the proposed VisuCeramNet model (46.10%), and YOLOv3-light (41.74%). This dataset shows promise as a benchmark for improving automated defect detection in the sanitary ware ceramics industry.

ABSTRACT. Underwater microplastic pollution poses a signifi- cant threat to marine ecosystems and human health, yet accurate detection remains challenging due to the scarcity of annotated data and the small size of target objects. This study presents a synthetic data augmentation approach to enhance underwater microplastic detection using the YOLOv11 object detection archi- tecture. Instance masks are generated with the Segment Anything Model (SAM), enabling the creation of synthetic training data through blending and copy-paste techniques on diverse aquatic backgrounds. We train and evaluate five YOLOv11 models on a consistent real-world test set, demonstrating that all synthetic variants outperform the baseline model, with the SAM-blended (1BG) model achieving the highest mAP@0.5:0.95. These results highlight the effectiveness of SAM-guided synthetic augmentation in addressing data limitations and improving performance for underwater object detection tasks.

ABSTRACT. The escalating marine plastic pollution crisis requires automated, scalable solutions for monitoring. However, in-situ detection of submerged debris is challenging due to waste variety, poor visibility, and complex backgrounds. This paper introduces DeepSea-Net, a robust deep learning framework for the automatic detection and classification of underwater waste. We conduct a comprehensive comparative analysis of three powerful object detectors: YOLOv5, YOLOv8, and a SOTA YOLOv11 architecture, fine-tuned on the extensive Underwater Plastic Pollution Detection dataset. Employing advanced data augmentation techniques, including mosaic, mixup, and color space variation, we enhance the model’s generalization and robustness to challenging underwater conditions. Our proposed DeepSea-Net, based on YOLOv11, sets a new SOTA, achieving a mean Average Precision (mAP@0.5) of 79.53%. The framework demonstrates a high recall and a leading F1-Score of 74.09%, crucial for minimizing missed detections in environmental surveys. Achieving an inference latency of 16.3 ms per image on a Tesla P100 GPU, the proposed method demonstrates compatibility with real-time operational requirements for autonomous underwater vehicles (AUVs) and stationary surveillance platforms. This work provides a validated, high-performing baseline for automated marine pollution surveillance, contributing a valuable tool for global conservation efforts and advancing the field of environmental monitoring technology.

ABSTRACT. Disaster response systems are based on the ability to generate access to constant and reliable information; however, logistics database malfunctions can lead to severe losses. This study introduces a self-healing system that utilizes reinforcement learning (RL) to ensure high-availability and reduce downtime in critical environments. The data used in this study were obtained from the U.S. Coast Guard MISLE reports and New York Spill Incident Record, in real-world data, to develop a pipeline based on PostgreSQL and anomaly labeling automation. RL simulations in a custom environment that simulates an actual time database behavior and an agent trained by Deep Q-Network (DQN) that learns to approach the anomalies by considering five corrective actions including imputation, eliminating duplication, rollback, query optimization, and plan reconfiguration. The system was tested and monitored in a streaming scenario using the Mean Time to Recovery (MTTR), Downtime Success Rate (DSR), and anomaly resolution accuracy. The results demonstrate an accuracy of 90.1% ± 2.3%, an uptime of 84.4% ± 3.1%, and an MTTR of 1.14 ± 0.29 s in simulated environments. This proves that the RL guide effectively and quickly resolves issues, while underscoring the necessity for real-world validation. The proposed design can open an auto-scalable journey to self-sustaining database health in mission-critical applications with reduced human monitoring and manual adjustments.

ABSTRACT. One of the biggest challenges facing healthcare systems is maintaining the privacy of sensitive patient data while facilitating effective clinical analysis and verification of the data integrity. In order to provide precise medicine recommendation system based on patient symptoms, this paper suggests a secure framework enabling encrypted computations on distributed shares, verification of data integrity, suggestions by authorized entity based on user queries etc. For this, a hybrid cryptosystem namely Multiplicative Secret Sharing and ElGamal cryptosystem are used for secure distributed storage and computation. Again, the verification protocol confirms the integrity of the data after secure reconstruction of data from encrypted shares while necessary. To provide the medicine recommendations based on user queries, it proposes a modification of skyline query technique namely, Dominance Frequency-Based technique and compares it with sophisticated approaches, such as Block Nested Loop, Divide-and-Conquer etc., which shows a remarkable similarity. Also, it enforces Multi-Factor Authentication for approving the proposed recommendations by authorized entity. The evaluations from the proposed framework and comparisons with state-of-the-art works provide a reliable and scalable solution for privacy-aware clinical decision support by skillfully striking a balance between computational usefulness and privacy preservation.

ABSTRACT. Mental illnesses are complex disorders that signif- icantly affect individuals worldwide. In the digital era, social media has become an important medium for individuals to express their emotions and seek support. Machine learning (ML) and deep learning (DL) approaches have been used to identify mental health problems from online posts, yielding promising results. However, most research focuses on binary classification and often struggles with imbalanced datasets. This study aims to develop a BERT-based multiclass classification model to identify several mental disorders, including Normal, Depression, Suicidal, Anxiety, Stress, Bipolar, and Personality Disorder. To address data imbalance, a hybrid data balancing technique combining Random Oversampling (ROS) and Random Undersampling (RUS) is applied. The model is evaluated on a Mental Health Text Dataset collected from Kaggle and compared with traditional ML algorithms, including Logistic Regression, Random Forest, and Naive Bayes, as well as an LSTM-based DL model. Experimental results show that the baseline BERT model achieved an accuracy of 83%, which improved to 90% after applying the hybrid balancing technique. The proposed approach demonstrates superior performance compared to traditional ML algorithms and LSTM, highlighting the effectiveness of integrat- ing transformer-based models with data balancing techniques for more accurate and robust multiclass mental illness detection from social media data.

ABSTRACT. With the rapid growth of short and fragmented texts (e.g., tweets, search queries, chat messages), automatic keyphrase extraction has become increasingly challenging. Keyphrases are salient single-word or multi-word expressions that represent the core topics of a document. They are essential for tasks such as indexing, search, recommendation, and summarization—especially in short or noisy texts where explicit context is limited. Conventional keyphrase extraction methods often struggle to capture the meaningful underlying themes of short texts due to their sparse context. To address this, we propose Embedding–Neural Keyphrase Extractor (E-NeuKey), a short-text keyphrase extraction model that integrates statistical techniques (TF-IDF and PMI) with deep semantic embeddings from transformer-based models. Additionally, a capsule-based encoder is employed to capture part–whole relationships between tokens and phrases, while a Variational Autoencoder (VAE) with Stochastic Variational Inference (SVI) models phrase-level relevance, semantic richness, and contextual prominence, particularly in low-resource settings. We evaluate the proposed model on standard short-text datasets, including 20Newsgroups, BBC News, and Twitter, and compare it against state-of-the-art techniques. E-NeuKey consistently outperforms strong baselines such as PKE, YAKE, TeKET, and E-TeKET, achieving precision, recall, and F1-scores of 0.86, 0.84, and 0.82, respectively—establishing a new benchmark for keyphrase extraction from short texts.

ABSTRACT. Topic modeling automatically discovers the main underlying themes in a collection of texts by grouping related words into topics. It is widely used for organizing, searching, and summarizing large text corpora. Traditional approaches such as LDA, LSI, and NMF perform well on long, context-rich documents, e.g., news articles, scientific papers where word co-occurrence is abundant; however, they underperform on short and noisy texts, i.e., tweets, queries, and chat, due to sparse context and insufficient co-occurrence evidence. To address these limitations, this paper presents EDNeuFTM (Enhanced Deep NeuroFusion Topic Modeling), a topic model designed specifically for short texts. EDNeuFTM integrates pretrained word embeddings with a variational Autoencoder (VAE) to model document-level topics and employs capsule routing to refine token-to-topic assignments. Training minimizes a multi-objective loss that increases coherence and diversity while penalizing redundancy, and leverages a CRT-based topic–word count sampler to stabilize and accelerate convergence. Across multiple short-text datasets, including 20Newsgroups, BBC News, and Twitter, EDNeuFTM consistently outperforms strong baselines and attains the highest \(C_{\mathrm{NPMI}}\) and \(C_{\mathrm{UCI}}\) coherence scores among all evaluated methods. In short, EDNeuFTM closes the performance gap where traditional topic models struggle, delivering clearer, less redundant topics for short-text contexts.

ABSTRACT. Mental disorders such as depression, anxiety, and suicidal ideation are rising rapidly among young people in Bangladesh, with social media providing a prominent platform for expressing oneself and emotions. Early identification of these through automatic systems can provide timely intervention and even save lives. To address this problem, we propose a hybrid CNN-BiGRU model that combines linguistic, psychological, and sentiment attributes to model semantic meaning along with long term contextual relations in Bangla text with a 94% accuracy. Our method, trained on a well-curated dataset, is more precise and has superior generalization compared to standard machine learning methods like Logistic Regression, Support Vector Machines, and Multinomial Naive Bayes. These results indicate the potential of deep learning towards effective mental health detection in low-resource languages

ABSTRACT. Retrieval-Augmented Generation (RAG) grounds large language models in external corpora but suffers from noisy retrieval, redundant evidence, and inefficient context use, leading to degraded recall and correctness. Existing solutions such as re-ranking, summarization, and corrective refinement either discard low-scoring documents or compress all evidence uniformly, risking the loss of valuable evidence. We propose Quality-Based Partitioning (QBP), a lightweight post-retrieval module that partitions retrieved passages by a similarity threshold: high- scoring passages are preserved, while low-scoring ones are fused into a query aligned summary. We evaluate QBP and its cross- encoder variant on SQuAD, MS MARCO, and a custom UEFA Champions League dataset. QBP along with its cross-encoder variant achieves a stronger balance between answer correctness and context recall than RAGFusion and Self-RAG. Further, our results highlight that partition and summarization, while simple, can operate as a practical direction for RAG systems.

ABSTRACT. Regular crowdfunding platforms are excited by the mystery, centralization, and abnormally high rates and, in turn, the search for less traditional ways for this purpose. In this paper, we present the work of designing and evaluating a donation system to leverage the Ethereum-supported blockchain along with emerging web tools. The objective is a transparent, auditable system. Users can launch fundraisers, donate Ethereum, and monitor their progress. Actors are struggling to keep up. The core mechanical pieces are driven by smart contracts written in Ethereum’s Solidity. These contracts define the driving terms of the agreement, including funding amounts, degree of wet service access and agreed-upon payment plan, cash call, and payback. The front end is integrated with React.js, and Ethers.js is an easy-to-use client library that uses blockchain to communicate with MetaMask. Functionality was tested with starting drives, donations, and withdrawals using development environments of Hardhat, Ganache and Remix IDE. Avg. Transaction process- ing costs were in the ballpark of 200,000 Gas units, which were acceptable for Layer-1 Ethereum networks. The system is designed with security features, implementing access control, timestamp verification, and reentrancy protection as a step toward addressing common smart contract vulnerabilities. We will soon integrate DAO governance, the ability to convert fiat to crypto, phone support and enhanced Layer-2 scaling.

ABSTRACT. This paper presents SafeEntry, an intelligent doorlock system combining biometric fingerprint authentication, pass-code entry, and real-time camera monitoring. The system employsan Arduino microcontroller, 4x4 keypad, servo motor, I2C LCDscreen, fingerprint sensor, and ESP32-CAM module to providesecure, real-time access control and monitoring. Biometric au-thentication is provided through the fingerprint sensor, withsupplementary passcode-based security through the keypad. Theservo motor serves as the locking mechanism, and the LCD isemployed for displaying system messages and status. The ESP32-CAM is employed to capture and transmit live images or videostreams for remote monitoring. The modular nature of the systemensures scalability, flexibility, and fault-tolerant operation acrossdifferent environments. Research analysis indicates a trend inIoT security systems toward multimodal authentication basedon combined biometric and contextual knowledge for enhancedaccuracy. The article covers system architecture, deployment, andverification, pointing out the capabilities of the proposed modelin outperforming traditional security solutions.

ABSTRACT. Federated Learning (FL) enables multiple entities to collaboratively train models without sharing sensitive data, but it faces critical privacy, security, and efficiency challenges in healthcare intrusion detection systems. These issues are intensified by adversarial attacks, non-IID data, and the need for real-time performance. Existing FL methods struggle with gradient inversion, model poisoning, Sybil attacks, and high computational overhead, limiting their effectiveness in secure and scalable healthcare applications. This work proposes the Privacy-Preserving Personalized Federated Learning Intrusion Detection in Healthcare applications (P3FL-HIDS), integrating Byzantine robust aggregation, gradient masking, and Zero-Knowledge Proof–based authentication. Key features include strong adversarial resilience, protection of privacy against gradient inversion, personalized model adaptation for heterogeneous data, and secure participant authentication. Additional contributions include a dual-network training approach, adaptive clustering for personalization, and optimized secure communication for real-time healthcare scenarios. Experimental results on a Brain Tumor magnetic resonance imaging (MRI) dataset show that P3FLHIDS outperforms state of the art works in terms of accuracy, resilience, and resistance.

ABSTRACT. The explosive rise of Internet of Things (IoT) devices has intensified the demand for efficient task execution and fair resource allocation at the network edge. Conventional Mobile Edge Computing (MEC) systems face critical limitations in handling diverse, delay-sensitive tasks as they often ignore fairness, multidimensional resource coupling, and adaptive priority awareness. In this paper, we have developed an optimization framework using mixed integer non-linear programming (MINLP) for PRiority- and FAirness-aware multi-dimensional Resource allocation for task offloading, aiming at minimizing both Energy consumption and Delay, namely PRiFARED, in federated MEC environments. The PRiFARED integrates an adaptive priority mechanism, considering delay sensitivity, service class and subscription differentiation, with Jain’s fairness index applied to CPU, bandwidth, and power allocations. Experimental results depict that the PRiFARED system significantly reduces average service latency and energy consumption on an average of 14% and 9%, respectively, compared to the state-of-the-art works.

ABSTRACT. The study examines the key factors affecting the sustainability of e-commerce among the SMEs in Bangladesh in the changing industry 5.0 landscape. Primary objective is to examine the influence of TR, PA, TMS, OC, GS, and CP on the implementation of sustainable e-commerce. Using a quantitative research design, a structured questionnaire was used as a data collection tool and administered to 400 SMEs, out of which 185 valid surveys were analyzed by using SmartPLS as a PLS-SEM. The results show that TMS and PB are the best positive predictors of sustainable e-commerce adoption whereas OC and CP play a significant role. Surprisingly, the GS was negatively affecting, indicating a non-congruence between policy structures and SME realities. These results present organizational-technological-environmental interaction as a complicated one in facilitating sustainable digital change. In practice, the research provides practical insights to SME administrators and policymakers in order to increase leadership attention, prove clear benefits, and target government interventions to improve sustainable adoption of e-commerce. Socially, promoting sustainable e-commerce among SMEs would help in the inclusion of the economy, environmental protection, and social accountability on the fast-digitizing economy of Bangladesh. This study is unique as it uses the TOE model to determine the sustainability of e-commerce adoption in a developing economy in Industry 5.0, which is a big gap in literature. The use of convenience sample and cross-sectional data are the limitations that could interfere with the generalizability of the findings and require longitudinal and more widespread geographic studies in the future.

ABSTRACT. Agriculture in Bangladesh is the backbone of our economy, incorporating outdated methods with uncertain weather. To address these challenges, this paper concerns itself with advanced big data-driven methods for improving crop production. The recent methods are not accurate across different levels of classification and have poor robustness and generality. We introduce MGSX, a new Ensemble model for multi-label crop production classification that combines Machine Learning and Deep Learning. By means of a quantile-based method, the model discretizes crop production into 3 labels (Low, Medium, and High) and 5 labels (Very Low, Low, Medium, High, and Very High) based on the known production range for robust yield prediction. Our approach centres on capturing agro-environmental information from BBS, NASA, and Open-Meteo in a preliminary stage, which we refer to as preprocessing and feature engineering. A combined model consisting of machine learning, deep learning algorithms, and a stacking ensemble model has been employed (MGSX), which is based on Multi-Layer Perceptron (MLP), Gradient Boosting Machine (GBM), Support Vector Machine (SVM), and XGBoost as base learners and Light Gradient Boosting Machine (LightGBM) as meta learner. The MGSX model produces the best performance compared to other baselines, with an accuracy of 93.61% and 88.44% and a macro F1-score of 93.59% and 88.43%, for both 3 and 5 classification labels, respectively. These findings illustrate that our method is effective for learning complex and latent patterns of automated agricultural time series and provides solutions with scalability to diverse situations in crop production forecasting.

ABSTRACT. Excessive use of chemical fertilizers in agriculture contributes to environmental degradation without proportional gains in crop yield. This research proposes an AI-driven framework to optimize fertilizer application, specifically nitrogen, phosphorus, and potassium, while maintaining productivity. Using globally harmonized datasets from Coello et al. and FAOSTAT, the study integrates fertilizer usage maps with crop yield statistics for soybean cultivation in Brazil. A Random Forest regression model was trained to predict yield from nutrient input, achieving an R² of 0.951 and outperforming linear and gradient boosting alternatives. Random Forest was selected due to its robustness with small tabular datasets and ability to capture nonlinear nutrient–yield interactions while providing interpretable feature importance. To explore the trade-off between fertilizer use and yield, a multi-objective genetic algorithm, NSGA-II, was applied, enabling identification of Pareto-optimal fertilization strategies across sustainable, balanced, and high-yield zones. Scenario simulations show that up to 30% reductions in nitrogen and phosphorus inputs can be achieved with minimal yield loss, providing an interpretable, data-driven pathway toward sustainable fertilizer management.

ABSTRACT. With the rapid expansion of intelligent agricultural systems, early and accurate detection of plant diseases has become critical for safeguarding crop yield and quality. Among these, apple leaf diseases pose a significant threat to global fruit production. In this study, we propose an efficient deep learning-based approach for apple leaf disease classification using MobileNet, a lightweight convolutional neural network architecture, combined with transfer learning. The PlantVillage dataset, consisting of four categories—Apple Scab, Black Rot, Cedar Apple Rust, and Healthy foliage—was employed for model training and evaluation. To address challenges such as class imbalance and overfitting in relatively small datasets, we adopted advanced data augmentation strategies, including MixUp and CutMix. Experimental results demonstrate that these techniques substantially improved model generalization and classification accuracy. The baseline MobileNet model achieved a validation ac- curacy of 94.6%, which increased to 98.1% when using CutMix, outperforming both the baseline and MixUp-augmented models. Confusion matrix analysis further confirmed that CutMix effec- tively reduced misclassification, particularly in minority disease classes. Overall, our findings highlight that incorporating CutMix with transfer learning significantly enhances the accuracy and robustness of plant disease detection systems, offering a practical solution for precision agriculture.

ABSTRACT. This study presents the development of a solar-powered edible oil extraction machine and provides a detailed comparative analysis using various raw materials. Despite Bangladesh producing nearly 2 lakh Tones of edible oil in 2025, decentralized oil extraction systems remain largely underutilized in the country. The proposed machine is designed to process multiple types of raw materials efficiently, requiring only 250 W of power and operating on a 24 V DC source, making it suitable for rural and off-grid applications. The values obtained for the solar powered oil extractor include oil extraction rates of 40.5% for mustard seed, 49.5% for nut seed and 46% for sunflower seed, which are much higher than the oil extraction rates of the traditional mechanical driver system. Moreover, the DC motor extractor also reduces the performance time and consumes less power which brings overall cost and sustainability benefits

ABSTRACT. Halide anti-perovskites with the general formula A₃MX₃ composition have recently drawn wide interest as photovoltaic materials because of their structural versatility, tunable bandgaps, high absorption capability, and long carrier diffusion lengths. This study aims to apply density functional theory (DFT) calculations with CASTEP and device simulations using SCAPS-1D to study the structural, electronic, optical, and photovoltaic behavior of Sr₃MI₃ (M = As, Sb, Bi) perovskites. Findings from DFT reveal direct bandgaps in the range of 1.25-1.33 eV with strong p–d orbital hybridization, efficient visible-light absorption, and low electronic loss factors that support effective light harvesting and charge transport. Device modeling of Al/FTO/SnS₂/Sr₃MI₃/Ni heterostructures indicates that Sr₃BiI₃ delivers the best performance, achieving a power conversion efficiency of 30.96% with high current density and stable voltage. The study also demonstrates how absorber thickness, bandgap tuning, and doping density influence device performance. Overall, optimized Sr₃MI₃ compounds indicate high efficiency, thermal stability, and strong promise for next-generation solar cells. Importantly, this work reflects Industry 5.0 principles by using computational tools to reduce experimental waste, save energy, and promote sustainable and scalable clean-energy technologies.

ABSTRACT. Perovskite based solar cells (PSCs) are gaining popularity due to their higher photovoltaic (PV) conversion efficiency com-pared to conventional PV cells. Just a few of the many virtues of PSCs: they’re inexpensive, easy to fabricate, can be tuned to adjustable things like the energy bandgap, have excellent elec-trical and optical properties, and are highly effective at form-ing thin films. In this work, a numerical simulation and opti-mization of CH3NH3SnI3-based PSC under perfect light is demonstrated by the SCAPS-1D software. CH3NH3SnI3 perov-skite absorber layer is environmentally friendly and possesses the suitable structure, electrical and optical properties for solar applications. Our proposed structure Al/ITO/TiO2/CH3NH3SnI3/Cu2O/Ni consists of CH3NH3SnI3-based PSC, Al as the front contact, Ni is the rear contact, ITO as an transparent conductive oxide, TiO2 and Cu2O as the electron transport layer (ETL) and the hole transport layer (HTL), respectively. We analyzed the influence of the absorber thickness, doping concentration of the absorber and ETL, and of the defect concentrations at critical contacts on the perfor-mance of the system. In addition, to learn thermal stability of the device we investigated the device’s performance under temperature variations. In optimum operating condition, from the optimized structure, we have found an efficiency of 31.53%, with VOC = 1.08 V, JSC = 32.91 mA/cm² and FF = 88.46%. Our findings indicate that TiO2-based lead-free CH3NH3SnI3 perovskite cell would be good candidate for sta-ble and efficient electricity production from solar energy in an environmentally friendly manner.

ABSTRACT. Rain streaks degrade image visibility and structure, hindering vision tasks like segmentation and scene analysis. To address this, we propose ART-UNet—an Attentive Rain-Transformer U-Net with Cross-Scale Feature Fusion for accurate and consistent deraining. At its core is the Unified Rain Processor (URP), a hierarchical module spanning encoder, bottleneck and decoder stages. URP combines a Rain-Attentive Multi-Scale Transformer (RAMST) for long-range context, a Dilated Multi-Receptive Field Module (DMRFM) for spatial refinement and a Self-Supervised Rain Amplification mechanism to emphasize degraded regions while preserving structure. These elements collaboratively extract rain-aware features across scales and complexities, suppressing artifacts effectively. Cross-resolution skip fusion preserves scene detail and supports consistent information flow, while residual learning and gated local mixing enhance robustness under heavy, oblique or inconsistent rain. ART-UNet achieves effective deraining with strong quantitative scores and visually pleasing results, performing competitively with state-of-the-art models under challenging conditions.

ABSTRACT. The field of medical image analysis has been radically transformed by deep learning techniques, particularly Convolutional Neural Networks (CNNs), which have enabled high-accuracy diagnoses, even in the presence of insufficient data. Recent developments have facilitated the development of automated systems that are capable of diagnosing diseases with minimal human intervention. The increasing prevalence of skin cancer poses a substantial challenge as a result of the difficulty in identifying lesions, which are frequently distinguished by irregular boundaries, noise artifacts, and low contrast, among other health implications. This necessitates the laborious and complex process of image segmentation, which involves the separation of the lesion from the adjacent tissue, in the development of precise diagnostic systems. To address these obstacles, this study implements a hair removal step in conjunction with a segmentation model that substitutes the conventional UNet encoder with a pre-trained Xception architecture. Later, the model is fine-tuned to improve the performance of skin lesion segmentation by applying the power of transfer learning. To assess its generalisability, our model is tested on the PH2 dataset and trained on the ISIC 2018 dataset. The proposed model's effectiveness in accurately segmenting skin lesions is demonstrated by the experimental results, which indicate that it successfully obtains a segmentation accuracy of 97.221%, an IoU of 91.406%, and a Dice Coefficient of 95.418% on the ISIC 2018 dataset.

ABSTRACT. With the advancement of lithography technologies, integrated circuits are becoming smaller and their manufacturing processes increasingly complex. Consequently, wafer maps now exhibit a higher frequency of mixed-type defects. Accurate defect pattern recognition is critical for identifying root causes and improving manufacturing processes. In this paper, we present a Regional-to-Local Attention-based Vision Transformer (RegionViT) for classifying mixed-type wafer map defect patterns by leveraging both regional and local features. Experiments are conducted on a dataset containing 38 types of mixed-type defect patterns. The proposed model demonstrates strong classification performance, achieving an accuracy of 98.63%.

ABSTRACT. Skin cancer is one of the most common types of cancer worldwide, with melanoma being the most dangerous. Earlyand accurate detection is crucial for improving patient outcomesand reducing mortality. This paper introduces a Computer-Aided Diagnosis (CAD) system using an enhanced Efficient-NetB7 model with advanced preprocessing. The system integratesmorphological hair removal, image inpainting, class-weightedtraining and Grad-CAM-based explainability to produce moreaccurate and interpretable lesion identification. By integratingthese components into a single, end-to-end pipeline, the proposed system achieves improved diagnostic performance while providing clinically meaningful visual explanations. Tested on public  dermoscopic dataset, the model achieved 91.57% accuracy, 91.20 F1-score, 89.42% average precision, 89.60% recall, 0.9725 PR AUC, and 0.9669 ROC AUC, outperforming several existingmethods. The high recall demonstrates effectiveness in detectingcases, reducing missed diagnoses. Incorporating explainable AI increases transparency and trust among healthcare professionals, aligning with Industry 5.0 principles of human-AI collaboration.This system shows strong potential for telemedicine and resource-limited settings, enabling early detection and better access to dermatology care.

ABSTRACT. The study focuses on the variables that affect the adoption intentions of XAI in customer service chatbots in the Bangladeshi telecom sector. The primary objective was to examine the role of PE, EE, SI, TR, and PT in the propensity of employees and managers to use chatbots that are enabled by XAI. Quantitative research design was also used whereby 171 employees and managers were selected using a structured questionnaire. To test the hypotheses, convenience sampling was applied, and the answers were evaluated by the PLS-SEM in SmartPLS. The results have shown that PE, SI, TR, and PT were highly significant in determining the intentions to adopt AI, but EE did not show significance. These points to perceived usefulness, comprehension of AI output, trust in system reliability and peer or organizational acceptance as the most important aspects of XAI adoption and ease of use might be less significant when other factors override the decision-making process. In practice, the research can be used as advisory to telecom companies looking to adopt XAI-powered chatbots, as it focuses on transparency, building trust, and proving performance in order to increase its acceptance among staff. Socially, the research helps enhance the quality of customer service and satisfaction through the establishment of effective explanatory AI technologies integration. The novelty of the research is in the fact that it expands the UTAUT2 model with the help of XAI-focused Constructs, which provided information about technology acceptance in the new markets. Cross-sectional design and convenience sampling pose some limitations because they may reduce generalizability. Future studies can use longitudinal designs and examine the attitudes of the customers in order to offer a more holistic picture of XAI adoption in service management environment.

ABSTRACT. The exponential global data traffic rise, driven by cloud computing, artificial intelligence, and digital services, has led to an unprecedented surge in data center energy consumption. This research presents a comprehensive review of the global energy dynamics of data centers, examining regional growth patterns, consumption trends, and the operational contributions of hyperscale providers. It investigates the escalating power demand, projected to reach 1,479 TWh by 2030, and identifies key challenges like grid stress, carbon emissions, e-waste, and regulatory conflicts. Simultaneously, it explores promising opportunities in energy efficiency, renewable integration, smart grid participation, and demand-side flexibility. The analysis draws on industry data, scholarly literature, and regulatory reports to provide insights into how the sector can align rapid digital expansion with sustainable energy practices. The findings underscore the urgency of adopting cross-cutting technological, regulatory, and operational strategies to ensure that data center growth supports rather than undermines global climate and energy transition goals.

ABSTRACT. Artificial intelligence (AI) has the potential to transform education by enabling personalized learning and intelligent academic assistance, yet most solutions rely on cloud infrastructure, limiting accessibility in low-resource environments. This paper presents the design and optimization of a fully offline AI-powered educational assistant based on LLaMA 3.1–8B, fine-tuned with structured OpenStax content and instruction-following Alpaca data using Low-Rank Adaptation (LoRA). The merged model is quantized to a 4-bit GGUF format via llama.cpp, enabling efficient inference on consumer-grade devices. We detail a reproducible training and deployment pipeline, including evaluation using BLEU, ROUGE-L, latency, and memory metrics, and introduce a desktop GUI for local interaction. Experimental results demonstrate substantial improvements in instructional quality and responsiveness, achieving near-interactive latency on a 16 GB RAM machine. The paper discusses design trade-offs, ablation studies, safety and ethical considerations, and reproducibility, establishing a scalable approach for offline academic support in underserved regions.

ABSTRACT. Scalable mitigation strategies are made possible by machine learning-based predictions of agricultural methane (CH$_4$) emissions.  A stacking-ensemble framework for predicting continuous CH$_4$ emission rates and categorical emission levels in Maize and Napier grass systems amended with bio-slurry is presented in this paper.  Using a linear meta-learner and polynomial, interaction, temporal-lag, and rolling-window engineered features along with leakage-aware target encoding, four tree-based base learners (Random Forest, LightGBM, XGBoost, and CatBoost) achieve regression performance of $R^2 = 0.6643 \pm 0.0479$ and classification accuracy of $68.57\% \pm 2.35\%$ on $N = 350$ samples using 5-fold GroupKFold cross-validation.  A statistically significant improvement is confirmed by a paired $t$-test against XGBoost ($t = 2.9047$, $p = 0.0439$).  CatBoost ($R^2 = 0.6584 \pm 0.0508$) was the best individual model by mean $R^2$. The most significant predictors, according to SHAP interpretability, are the percentage of bioslurry, treatment–replication–day interactions, and emission temporal lag. Although the stacking ensemble has a significant computational cost (mean fit time 4.5760s versus 0.0183s for linear regression; $\sim250.4\times$ slower), it improves predictive performance. An economic assessment using the U.S. EPA social cost of methane (\$1.50\,kg\(^{-1}\)) for maize at Day~42 shows treatment-specific annualized social emission costs:\( T1 = \$48.86\;\text{ha}^{-1}\text{yr}^{-1} \),\( T2 = \$88.09\;\text{ha}^{-1}\text{yr}^{-1} \),\( T3 = \$123.52\;\text{ha}^{-1}\text{yr}^{-1} \),\( T4 = \$59.96\;\text{ha}^{-1}\text{yr}^{-1} \),and \( T5 = \$48.77\;\text{ha}^{-1}\text{yr}^{-1} \) (close to T1). The study presents a validated, interpretable framework for predicting CH\textsubscript{4} emissions and identifies accuracy and computational cost trade-offs for real-world implementation.

ABSTRACT. Diabetic Retinopathy (DR) affects over 93 million individuals globally and represents a leading cause of preventable blindness [1], [12]. Automated DR screening systems face two key challenges: severe class imbalance in medical datasets and high computational complexity limiting deployment accessibility. We present a three-stage framework addressing these issues: (1) class-conditional Progressive Growing GAN synthesis to create balanced training data, (2) a dual-branch ensemble combining EfficientNet-V2-S and Vision Transformer (ViT-B/16) for superior feature extraction, and (3) knowledge distillation to compress the ensemble into a lightweight MobileNetV2 model. On APTOS 2019, our ensemble achieves 95.8% accuracy with 0.975 quadratic weighted kappa. Cross-dataset validation on DDR and Messidor-2 confirms robust generalization. The distilled student model maintains 92.4% accuracy while achieving a 96.8% reduction in model parameters, 97.6% reduction in computational operations (GFLOPs), and 82.2% reduction in inference latency, demonstrating practical feasibility for mobile deployment in large-scale DR screening programs.

ABSTRACT. Dust accumulation is a significant issue for solar energy systems in Bangladesh, especially in urban and industrial areas that are heavily polluted. The panels' efficiency and power production noticeably decline due to this soiling effect, making them less transparent. Dust's diverse composition, a combination of soil, soot, and industrial pollutants, worsens the problem, necessitating frequent and time-consuming cleaning to maintain system functionality and financial sustainability. Solar power is a potent and increasingly common renewable energy source. Dust and grime buildup significantly reduce the solar panels' ability to generate energy. Dust comes from a variety of sources and functions as an atmospheric aerosol, which has an impact on climate change in addition to solar performance. This study aimed to assess the effects of both manufactured and natural dust on the performance degradation of crystalline photovoltaic modules. The PROVA 1011 PV Analyzer and an artificial sun simulator were used to test a 40-watt solar panel in a laboratory environment with various materials, including salt, aluminum, iron, wood, coal, dirt, sand, cement, and lime.

ABSTRACT. Growing urban demand for electric vehicles (EVs) presents bigger challenges for an energy infrastructure that is largely fossil-fuel-dependent in countries like Bangladesh. This study attempts to consider the hybrid EV charging station concept set with solar photovoltaic (PV) generation and hydrogen production, storage, and fuel cell technologies. Using the HOMER Pro simulation, the system was modeled for an EV load of 10,000 kWh/day with a peak demand of 763.85 kW, comprising PV arrays, converters, PEM electrolyzers, hydrogen tanks, and backup generators. From the results, it was seen that PV generated 85.6% of the total electricity, supported by grid purchases (8.13%) and fuel cell output (6.27%). The electrolyzers produce 7,703 kilograms of hydrogen every year, leaving an excess of 493 kilograms after the fuel cells use their share. Economically, the levelized cost of energy (LCOE) is $0.0056/kWh, the levelized cost of hydrogen (LCOH) is $8.98/kg, net present cost (NPC) is $894,222.20, payback period is 6.66 years, and IRR 16.6%. These findings confirm that renewable–hydrogen hybrid charging stations are both technically feasible and economically attractive, offering a pathway to reduce emissions, improve grid resilience, and support Bangladesh’s EV adoption targets.

ABSTRACT. Accurate long-term forecasting of solar irradiance and wind speed is critical for renewable energy integration and grid planning. This paper proposes an optimized deep learning framework based on the Autoformer architecture for multivariate forecasting of Global Horizontal Irradiance (GHI), Diffuse Horizontal Irradiance (DHI), Direct Normal Irradiance (DNI), and wind speed. Using historical NSRDB data, the model employs series decomposition and auto-correlation to capture climatological trends and recurrent dependencies. Hyperparameter tuning with Ray Tune and ASHA scheduling enables efficient parameter search, while robustness is validated through K-Fold cross-validation. Experimental results demonstrate significant improvements: up to 74% MSE reduction for DHI compared to DWT-BiLSTM and more than 80% improvement over MSMVAN across metrics, with wind speed accuracy maintained at 79–84%. Under seasonal variability, the model achieves GHI MSE values of 611.73, 447.85, and 363.46 for summer, rainy, and winter conditions, demonstrating resilience to high-radiance, cloud induced intermittency, and winter variability. These findings confirm the Optimized Autoformer’s superior accuracy, stability, and scalability, positioning it as a strong candidate for smart grid scheduling and renewable energy resource management.

ABSTRACT. Rapid urbanization and globalization have intensified environmental challenges worldwide, particularly through rising anthropogenic greenhouse gas (GHG) emissions such as carbon dioxide and methane. Open dumping of municipal solid waste (MSW) is a major contributor, as the organic fraction undergoes anaerobic decomposition and releases significant amounts of methane, a potent GHG with considerable energy potential if effectively captured. This study investigates landfill gas (LFG) generation, with emphasis on methane recovery, from MSW in Dhaka city for the period 2016–2030 using the LandGEM 3.02 version. The results indicate cumulative LFG production, methane emissions, and energy recovery potential of 177.10 × 106 m3, 1.20 × 109 m3, and 2.95 × 109 kWh/year, respectively, with a corresponding CO2 avoidance of approximately 2.54 × 109 kg CO2 per year. Economic analysis demonstrates feasibility with a positive Net present value (NPV) of 1.95 billion BDT, an LCOE of 2.10 BDT/kWh and payback period (PBP) of approximately 16 years. These findings suggest that methane utilization from MSW not only mitigates GHG emissions but also provides a sustainable pathway for energy generation, supporting Industry 5.0 goals of cleaner, smarter and more responsible technologies.

ABSTRACT. This work investigates how effectively the optimization algorithms based on artificial intelligence (AI) can tune the gains of a proportional–integral–derivative (PID) controller in load frequency control (LFC) systems, which are a crucial part of the smart grid. Applying such LFC tuning in a real power grid would improve frequency regulation, reduce oscillations, enhance reliability, prevent large deviations, and enable smoother integration of renewable energy. Genetic Algorithm (GA), Random Search and Particle Swarm Optimization (PSO) based approaches were used to find the optimal PID parameters in order to enhance the system's dynamic response in the event of a sudden load variation. To optimize and analyze the performance, a Simulink model of a single area LFC system was developed. From the analysis of some important transient parameters : settling time (Tₛ), Deviation from Desired (D_d), Steady-state error (E_ss), and integral time absolute error (ITAE) that is evident that the PSO-tuned PID demonstrates the best results. By using this approach we can achieve plant’s stable condition within just 2.6s having a very little E_ss of 1 × 10⁻¹⁰ and no observable D_d.

ABSTRACT. Renewable energy adoption is essential to addressing both climate change and global energy security, especially in fast-developing countries. Among the various renewable energy sources, wind power holds significant potential, and its effective use relies on accurate wind resource assessment. Existing research lacks region-specific studies, integrated explainable AI (XAI) analysis, and publicly accessible predictive models for wind speed. This study aims to develop an interpretable and deployable wind speed prediction framework, using 24 years (2000–2024) of NASA POWER daily meteorological data. Five machine learning models—KNN, Polynomial Regression, SVM, Random Forest, and XGBoost—were trained and benchmarked. XAI techniques (SHAP, LIME, ICE) were employed to interpret feature influence, and the best-performing model was deployed via a user-friendly Streamlit-based graphical user interface. Results show that XGBoost achieved the highest test performance (R2 = 0.71, MAE = 0.52, RMSE = 0.72), with temperature- related parameters identified as the most influential predictors. The proposed framework for wind speed prediction enhances transparency, accessibility, and decision-making capabilities for renewable energy planning in underexplored regions, providing a scalable reference model for global wind resource assessment initiatives.