International Journal of Innovations in Science & Technology
Not a member yet
813 research outputs found
Sort by
Extraction of Bio-Oil from The Pyrolysis of Banana Tree Waste Using A Fixed-Bed Reactor
Full text linkThe rapid and ongoing depletion of fossil fuel reserves is driving up energy costs and harming the environment due to greenhouse gas emissions, leading to a global energy crisis. This situation highlights the urgent need to produce renewable fuel from biomass. This research focuses on extracting bio-oil from banana tree waste under different operating conditions. In this study, the pyrolysis process of banana tree waste was carried out in a fixed-bed reactor to maintain controlled conditions and prevent unwanted cracking. To optimize the process, the effects of temperature, particle size, and nitrogen flow rate on bio-oil yield were investigated. Experiments were conducted at temperatures ranging from 400 to 600 ℃, with feedstock particle sizes of 0.5 – 2.0 mm and nitrogen flow rates between 0.5 and 2 liters per minute. The optimal conditions for maximizing bio-oil yield were determined. Under these conditions, the maximum bio-oil yield of 32.13% was obtained at a temperature of 500 ℃, with a particle size of 1.2 – 2.0 mm and a nitrogen flow rate of 1 liter per minute. The results also demonstrate how temperature, particle size, and nitrogen flow affect the bio-oil yield during pyrolysis. The study concludes that banana tree waste can be efficiently converted into bio-oil through proper processing, contributing to sustainable energy production while minimizing environmental impact. The chemical composition of the bio-oil was analyzed using the GC-MS technique, which identified various compounds, including phenols, acids, and other chemical components
Efficient Region-Based Video Text Extraction Using Advanced Detection and Recognition Models
Full text linkThis paper presents an automated process for extracting text from video frames by specifically targeting text-rich regions, identified through advanced scene text detection methods. Unlike traditional techniques that apply OCR to entire frames—resulting in excessive computations and higher error rates—our approach focuses only on textual areas, improving both speed and accuracy. The system integrates effective preprocessing routines, cutting-edge text detectors (CRAFT, DBNet), and advanced recognition engines (CRNN, transformer-based) within a unified framework. Extensive testing on datasets such as ICDAR 2015, ICDAR 2017 MLT, and COCO-Text demonstrates consistent gains in F-scores and word recognition rates, significantly outperforming baseline methods. Additionally, detailed error analysis, ablation studies, and runtime evaluations offer deeper insights into the strengths and limitations of the proposed method. This pipeline is particularly useful for tasks like video indexing, semantic retrieval, and real-time multimedia analysis
Machine Learning-Based Improvement of Smart Contract Security in Fog Computing Using Word2vec And Bert
Full text linkFog computing extends cloud computing services closer to users, improving efficiency and reducing latency. Smart contracts play a key role in authentication and resource access management within this framework. As the adoption of smart contracts in fog computing grows, ensuring their security has become a major challenge. This study enhances smart contract attack detection in fog computing using machine learning techniques. A dataset of 818 smart contracts was collected from “etherscan.io.” Feature extraction was performed using Word2Vec and BERT, while feature selection was done using the information gain method. The Random Forest (RF) and Extra Trees Classifier (ETC) achieved the highest accuracy of 0.91 with Word2Vec, while the LightGBM (LGBM) classifier attained 0.90 accuracy using BERT.
These results demonstrate the effectiveness of machine learning models in improving smart contract security within fog computing environments
AI-Based Sindhi Handwritten Alphabets Classification with Web-Based Development
Full text linkHandwriting recognition has made remarkable progress for some prominent scripts, but low-resource languages such as Sindhi have received little attention so far. In this research, we propose the design and implementation of a strong AI based model to classify handwritten Sindhi alphabets. To overcome the difficulties caused by varying handwriting and a lack of publicly available datasets, the model builds on a manually curated, heterogeneous dataset, sophisticated CNN architectures, and data augmentation techniques. To support more research, the dataset will be made publicly available in two versions: raw and augmented. This study’s key contributions include achieving approximately 93% training accuracy and 96% validation accuracy with a loss below 1%, and the creation of valuable open-source datasets for Sindhi handwriting recognition. While a web-based application is planned as future work, these achievements provide a strong foundation for digitizing Sindhi texts and educational tools, and help preserving Sindhi language heritage
HLCE: Framework for Enhanced Stock Price Forecasting
Full text linkAccurate stock price forecasting is a key element of risk management and investment decision-making. A key element of this study is the introduction of a Hybrid LSTM-Conventional Ensemble (HLCE) model, which addresses the limitations of traditional models in capturing nonlinear financial patterns. Utilizing the advantages of both deep learning and conventional forecasting techniques, the HLCE framework combines Long Short-Term Memory (LSTM) networks with traditional statistical models and machine learning methods, including Random Forest, XGBoost, and Support Vector Regression (SVR). The model is assessed using important performance metrics, such as Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and R-squared (R²), in a case study using Apple Inc. (AAPL) stock data, where MinMaxScaler is utilized for data preprocessing. With an RMSE of 0.16, MAE of 0.16, MAPE of 0.12%, and R² of 0.95, the HLCE model performs better than individual models, according to experimental results, demonstrating its greater capacity to identify intricate financial patterns. By contrast, isolated models exhibit far lower predictive efficiency and much higher error rates. These results highlight the promise of ensemble and hybrid approaches in financial forecasting, offering a more reliable and accurate framework for predicting stock prices. The work adds to the expanding body of research supporting the combination of deep learning and conventional techniques to enhance risk assessment and financial market analysis
Predictive Analytics for Smart Cities: Traffic Flow Forecasting Using Ensemble Algorithms
Full text linkTraffic flow prediction is crucial for smart transportation systems, as it plays a key role in improving traffic management and planning infrastructure. While many machine learning techniques have been used for this purpose, ensemble methods have proven to be especially effective because they enhance prediction accuracy by combining the strengths of multiple models. This paper offers a detailed overview of how ensemble methods are applied to traffic flow prediction. We start by exploring the basics of traffic flow prediction, including common data sources, types, and performance metrics. Then, we categorize ensemble methods into bagging, boosting, and hybrid approaches, reviewing important studies that show how these methods work, the datasets they use, and their performance results. Real-world examples and case studies are included to highlight the practical effectiveness of these methods in various traffic situations. Finally, we discuss the current challenges and suggest future research directions, aiming to provide a valuable resource for researchers and practitioners interested in improving traffic flow prediction with ensemble techniques
Hybrid Intrusion Detection System Based on Optimal Feature Selection and Evolutionary Algorithm for Wired Networks
Full text linkThe field of cybersecurity encounters ongoing difficulties in identifying and preventing attacks in networks, and the pervasive threat of cyberattacks demands continual advancements in intrusion detection systems (IDS) to safeguard network integrity. Traditional intrusion detection systems face the challenge of class imbalance. Addressing the formidable challenges posed by class imbalance and high-dimensional data, this research proposes a novel hybrid IDS approach. Leveraging (ACO), the algorithm navigates complex datasets to identify salient features, effectively mitigating the complexities associated with high-dimensional data. Subsequently, a Weighted Stacking Classifier amalgamates the strengths of Random Forest, AdaBoost, and Gradient Boosting classifiers, fortifying the system’s ability to handle class imbalance robustly. By strategically enhancing the importance of base classifiers with favourable training outcomes and diminishing the influence of those yielding inferior results, the hybrid IDS endeavors to optimize classification efficacy. The experimentation, conducted exclusively on the dataset named NSL-KDD, demonstrates the efficacy of the proposed model, yielding remarkable results. With a 90.13% Accuracy, 88.87% precision, 91.23% Recall, and 87.33% F1-score, the hybrid IDS exhibits superior performance in detecting malicious activity. The findings underscore the viability of the proposed hybrid IDS as a potent tool in the ongoing battle against cyber threats, positioning it for real-world deployment across diverse networks
Deep Learning-based Weapon Detection using Yolov8
Full text linkDeep learning (DL), a subset of machine learning (ML), has demonstrated remarkable success in image recognition and object detection tasks. This study presents a deep learning-based approach for offline weapon detection using the YOLOv8m architecture. A custom YOLO-formatted dataset was developed, comprising over 10,000 annotated images spanning two weapon categories: guns (all types of firearms) and knives (all types). The model achieved a Mean Average Precision ([email protected]) of 0.852. and [email protected]:0.95 of 0.622, with precision and recall scores of 0.89 and 0.80, respectively. The class-wise evaluation revealed strong detection across both weapons, with [email protected] of 0.871 for knives and 0.831 for guns. Despite occasional false positives and class confusion, the system shows promise for offline weapon detection tasks
Prediction of Molecular and Physical Properties of Non-small Cell Lung Cancer (NSCLC) Drugs using Mathematical Modelling and M-Polynomial Indices
Full text linkThe computation of M-Polynomial indices for Erlotinib, a tyrosine kinase receptor inhibitor and most widely recognized anti-cancer drug for the treatment of patients with NSCLC and advance pancreatic cancer is the main focus of this study. In order to efficiently calculate these M-polynomial indices, we used a graph-based method which renders use of the edge partitioning technique based on adjacent matrices and vertex degrees. Using Python software, we applied numerous regression models, such as numerous Linear Regression (LR), Elastic Net Regression (ENR), Lasso Regression (LR), Ridge Regression (RR) and Support Vector Regression (SVR), to develop Quantitative Structure-Property Relationships (QSPR). Based on the M polynomial indices, these models were utilized to forecast the physical properties such as melting point, enthalpy of vaporization, molar refractivity, molar volume, and polarizability, molecular weight, molecular mass, surface area, chemical hardness of NSCLC medications. According to our research, the M-polynomial indices predict these physical attributes with remarkable accuracy, providing crucial information on structural traits that maximize anticancer effectiveness. Additionally, we suggested predictive models for every physical attribute examined, proving the value of the M-polynomial index in comprehending molecular behaviour and directing the creation of innovative therapeutic medicines. This study not only facilitates the accurate prediction of physical properties for known NSCLC drugs but also holds the potential to fasten the novel drug discovery and development, uncharacterized anti-cancer compounds, thus contributing to the advancement of cancer therapeutics
Maximum Value Attribute based Decision Tree and Random Forest for COVID-19 Prediction
Full text linkThe COVID-19 pandemic emerged as one of the most disruptive global health crises of the century, affecting social and economic systems worldwide. The rapid rise in infections placed immense pressure on healthcare infrastructures, demanding fast and reliable diagnostic tools. In recent years, Machine Learning (ML) has gained considerable importance in the medical field, supporting the diagnosis of conditions such as heart failure, pneumonia, dengue, breast cancer, and diabetes. In a similar way, clinical symptoms related to COVID-19 can be utilized to support early prediction, helping limit transmission. Although ensemble learning techniques such as Decision Trees and Random Forests have shown strong predictive performance for COVID-19, they often require more time and a larger number of iterations, which can be challenging when rapid detection is needed.
This study focuses on improving the efficiency of COVID-19 prediction by integrating Rough Set Theory (RST) through the Maximum Value Attribute (MVA) approach with classical Decision Tree (DT) and Random Forest (RF) models. The objective is to reduce computation time and iterations while maintaining reliable diagnostic accuracy. The proposed method classifies patients as COVID-19 positive or negative based on eight key clinical symptoms. A dataset containing clinical records of 136,294 patients, collected from an open-source GitHub repository, was used for evaluation. Four models—DT, RF, MVA-DT, and MVA-RF—were implemented in Python using Jupyter Notebook. Standard evaluation metrics were applied to assess performance. Overall, the MVA-DT model achieved the most efficient execution, while the MVA-RF model demonstrated strong predictive capability with an accuracy of 95.82%, precision of 81.90%, recall of 59.28%, and an F1 score of 68.77%.