Faridpur Engineering College

Insulator faults in power lines threaten the stability and reliability of power distribution systems. In this paper, we present an automatically fault detection scheme for power line insulators. The framework is based on the attention-guided hybrid deep learning interpretable model. It consists of the pretrained VGG16 network model and the Dilation DenseNet Block that is responsible through dilations to capture the multi-scales of features and by VGG16 for the high-level feature extraction. The model was fed every image on a normal set along with defective images and then tested on a set of evaluation metrics including training accuracy, testing accuracy, F1-score, and AUC-ROC. Experimental results demonstrate the effectiveness of the model, with a training accuracy of 98.52% and a testing accuracy of 97.06%. The framework will thus prove to be a reliable, efficient, and interpretable solution for on-line detection of faults in power line insulators, thereby improving time and resources devoted to maintenance and safety of the grid.

Skin diseases are a major public health concern worldwide, and their detection is often challenging without access to dermatological expertise. In countries like Bangladesh, which is highly populated, the number of qualified skin specialists and diagnostic instruments is insufficient to meet the demand. Due to the lack of proper detection and treatment of skin diseases, that may lead to severe health consequences including death. Common properties of skin diseases are, changing the color, texture, and pattern of skin and in this era of artificial intelligence and machine learning, we are able to detect skin diseases by using image processing and computer vision techniques. In response to this challenge, we develop a publicly available dataset focused on common skin disease detection using machine learning techniques. We focus on five prevalent skin diseases in Bangladesh: Contact Dermatitis, Vitiligo, Eczema, Scabies, and Tinea Ringworm. The dataset consists of 1612 images (of which, 250 are distinct while others are augmented), collected directly from patients at the outpatient department of Faridpur Medical College, Faridpur, Bangladesh. The data comprises of 302, 381, 301, 316, and 312 images of Dermatitis, Eczema, Scabies, Tinea Ringworm, and Vitiligo, respectively. Although the data are collected regionally, the selected diseases are common across many countries especially in South Asia, making the dataset potentially valuable for global applications in machine learning-based dermatology. We also apply several machine learning and deep learning models on the dataset and report classification performance. We expect that this research would garner attention from machine learning and deep learning researchers and practitioners working in the field of automated disease diagnosis.

Urbanization in Dhaka has led to differences in fecal sludge management (FSM) methods between urban (Mirpur) and semi-urban (Savar) areas. This research investigates how urbanization and user perceptions influence FSM practices using structured household survey data (n = 394) on toilets, containment systems, desludging frequency, service accessibility, and satisfaction. Statistical methods (correlation and two-way ANOVA) uncovered relationships among infrastructure, behavior, and perception, whereas laboratory-measured biochemical data helped in investigating the influence of retention time on sludge quality. The findings showed that Mirpur residents had larger containment volumes, more frequent desludging, and much higher satisfaction levels (mean = 4.2) than those in Savar (mean = 3.1). Tank size (r = 0.70), willingness to pay (r = 0.60), and accessibility (r = 0.53) were strongly correlated with satisfaction. According to biochemical studies, Savar’s sludge contained much higher TS, TSS, NH4–N, and pH but lower VS, COD, and TN, suggesting longer anaerobic retention leading to more stabilization. Urbanization has a positive impact on FSM service performance, service accessibility, and user perceptions, while semi-urban areas remain constrained by infrastructure and awareness gaps. Implementing decentralized treatment, planned desludging, and community-inclusive FSM planning is necessary for equitable sanitation and progress towards SDG 6.2.

Transformer-based models have evolved Natural Language Processing (NLP) although most of them have critical constraints with the learning capacity of morphologically rich and low-resource languages such as Bengali. The issues of the existing versions of Bengali BERT versions include the inefficiency of tokenization, the insufficientness of pretraining data, and the inability to work with out-of-vocabulary words, which result in the decline in task performance of the classification. These issues need to be handled to ensure enhanced training of learning Bengali. The paper includes a benchmark dataset of normalized Bengali text classification and a system that combines contextual embeddings of Bengali discriminated by BERT with subword distributed representations of FastText. The merged representations are fed to a GRU based classifier that enables it to capture sequential dependencies effectively without trying to optimize a full transformer. Experimentally, it has been found that fusion is superior to individual embeddings. The best combination, Kawsar BERT 4 layers and FastText- learned accuracy and F1-scores of 86.79 and 86.77 percent and also achieved better generalization scores with fewer epochs. These facts indicate that the fusion of the fusion is effective in solving the aspect of morphological complexity of dealing with Bengali. The direction of future work will be in form of multi-class/ multi-lingual classification, domain- adaptive pretraining and hybrid models that will involve recurrent encoder and attention-based encoder.

Accurate channel characterization across diverse propagation environments is foundational to 5G network planning, yet existing machine learning approaches rarely integrate standardized 3GPP frameworks with vendor-specific equipment parameters. This study presents a regression-based framework combining 3GPP TR 38.901 channel models with five supervised learning algorithms-linear regression, polynomial regression (degree 2), support vector regression (SVR), decision tree, and artificial neural network (ANN)-trained on 10,000 deterministic samples spanning Urban Macro (UMa), Urban Micro (UMi), Rural Macro (RMa), and Indoor Hotspot (InH) scenarios at five carrier frequencies (0.7-60 GHz). Vendor-calibrated parameterization using authenticated Nokia AirScale 64T64R, Huawei AAU5940, and ZTE AAU 5G specifications grounds the simulated link budgets in commercial equipment characteristics, providing deployment-aligned (though formula-derived rather than field-measured) performance estimates (see Limitations). All five regression architectures are evaluated identically across all five carrier frequencies and all scenario types, enabling direct comparison under controlled conditions. For throughput prediction, the ANN and decision tree achieve the highest accuracy (R2 = 0.998, RMSE ≤ 24 Mbps averaged across five independent random splits; 95% CI: R2∈[0.997,0.999], RMSE ∈[19.1,22.4] Mbps), while linear and polynomial regressors show substantial error (R2≤0.56), reflecting the strongly nonlinear throughput surface. For path loss estimation under Urban Micro NLOS conditions, all models attain near-perfect fit (R2≈1.0, MSE < 0.02 dB2), confirming that simple regressors suffice for log-distance targets. Vendor link budgets quantify the Nokia-Huawei throughput gap (1.88× at 100 m) and the ZTE 28 GHz peak capacity (1688.6 Mbps at 100 m), establishing a breakeven inter-site distance of approximately 150 m below which FR2 outperforms FR1. Cross-scenario generalization experiments reveal a critical failure mode: models trained on LOS-urban data yield strongly negative R2 on Rural Macro scenarios (<-3), while mixed-scenario training recovers generalization to R2 > 0.75 across all environments. Permutation-based feature importance identifies distance as the dominant predictor (importance 0.65-0.85), with frequency importance rising to ≈0.40 at millimeter-wave bands. Sensitivity analysis confirms robustness (R2 > 0.90) under realistic parameter perturbations (±10% distance, ±5% frequency, ±2 dB EIRP). These results provide evidence-based guidelines for model selection, training data composition, and deployment in 5G/6G network planning.