Conference Papers & Proceedings

In advancing power grid management, integrating artificial intelligence (AI) has emerged as a pivotal approach, particularly in smart grids. The modernization of power grids into intelligent, interconnected systems has introduced challenges in terms of stability prediction and efficient power management. Traditional methodologies often do not cope adequately with the intricacies associated with bidirectional communication and increased data volumes. This paper addresses the imperative need for enhancing smart grid stability through efficient power management by employing advanced algorithms and artificial intelligence (AI) techniques. Our study evaluates five algorithms for stability prediction. Leveraging bidirectional communication between the grid and consumers, machine learning algorithms, including deep neural networks, demonstrate superior performance in forecasting energy needs.

Large datasets often contain redundant or irrelevant attributes that negatively impact classifier performance. Effective feature extraction is crucial for improving classification accuracy by identifying relevant features and reducing redundancy. Traditional techniques, such as Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA), are commonly used but face limitations in balancing dimensionality reduction with feature relevance, often leading to suboptimal results in terms of classification accuracy. To address these challenges, this paper introduces a novel hybrid feature extraction approach that combines the strengths of correlation analysis and PCA. This hybrid method aims to create a minimal redundancy maximal relevance feature subset, overcoming the limitations of traditional methods by enhancing both efficiency and effectiveness in feature extraction.

Proteins play a crucial role in various biochemical activities, making accurate protein class and function prediction critical. Accurate protein function prediction is fundamental in computational biology for biological understanding. Relying solely on one modality for protein function prediction limits the model's ability to capture the protein's complete picture, potentially leading to inaccurate predictions. This paper addresses these challenges and presents a machine-learning model for precise protein class prediction, leveraging protein sequence, secondary structure, and interaction information. The model achieves a remarkable 100% accuracy in protein class prediction, highlighting the effectiveness of multimodal data. Additionally, GO notation prediction yields high accuracies of 96% for Biological Processes (BP), 97% for Cellular Components (CC), and 98% for Molecular Functions (MF).

Fabric defects are a major challenge in the textile industry, where quality is essential. These defects can appear at different stages of production. By automating defect detection, we can reduce the need for manual checks, speed up the process, and improve productivity. Existing fabric defect detection methods often rely on manual visual inspection, which is time-consuming, subjective, and prone to human error. Complex algorithms may not be suitable for real-time deployment in resource-constrained environments. This research addresses the limitations of existing methods by proposing a deep learning and Ensemble-based approach for fabric defect detection and classification. A custom fabric defect dataset was constructed specifically designed for this research. YOLOv8m and YOLOv8n-obb variants were fine-tuned for defect detection, achieving robust detection performance.

Tuberculosis (TB) is a chronic lung disease that constitutes one of the top 10 global causes of death. Timely and precise identification of TB is critical, as untreated cases can pose life-threatening risks. Traditional radiographic interpretation by trained professionals is subjective, time-consuming, and prone to human error. This study proposes an automated deep learning-based approach for TB detection from chest X-ray images, based on EfficientNet-B0 architecture. The Convolutional Neural Network (CNN) extracts discriminative features, allowing it to differentiate between normal and abnormal lung patterns indicative of TB. Our choice of the EfficientNet-B0 architecture played a pivotal role in the success of our TB detection model. Training and evaluation were conducted on three distinct datasets: Tuberculosis (TB) Chest X-ray Database, Shenzhen Hospital CXR Set, and Montgomery County CXR Set.

Diabetes poses a significant global healthcare challenge, affecting millions of individuals and straining healthcare systems financially. High hospital readmission rates for diabetic patients illuminate potential shortcomings in both the quality of care provided and the efficiency of healthcare delivery. Developing a reliable method to predict these readmissions is crucial to not only reduce the economic burden on patients and healthcare systems but also to proactively intervene and improve patient outcomes. This study aims to improve care for diabetic patients by reducing readmission costs, ultimately enhancing healthcare outcomes. We propose a machine-learning approach to predict hospital readmissions for diabetic patients. This approach innovatively integrates feature selection, feature engineering, and a specifically tailored SMOTE technique to address the inherent data imbalance challenge.

Focusing particularly on the lung region is crucial in various pulmonary disease classification. It helps to discard unwanted regions from the chest X-ray. This paper proposes the TIG-UNet++ model, an advanced modification of the Unet++-L4 architecture. A key feature of our model is the late introduction of a threshold image, that almost identifies the lung regions, which serves as a low-priority input to enhance feature fusion while minimizing its impact on the overall segmentation process. This approach surpassed recent studies in terms of dice, jaccard, auc, and accuracy, particularly in challenging cases with unclear or disease like tuberculosis affected images. It was tested on Montgomery, JSRT, and Shenzen datasets. Our model achieves a dice score of 98.0% on the Montgomery dataset, 97.8% on the JSRT dataset, 96.1% on the combination of Montgomery and Shenzen, and 97.3% on the combination of all the datasets.