Journal of Science Engineering Technology and Management Science

Flooding disasters pose significant challenges to emergency response systems due to rapidly changing environments, infrastructure damage, and limited resource availability. Efficient and timely dispatch of rescue teams is critical to minimize loss of life and property. This paper presents Mobirescue, an intelligent and adaptive system designed to optimize the dispatching of rescue teams during flood emergencies. The proposed system integrates real-time data sources such as weather forecasts, water levels, geographic information, and victim locations to dynamically assess disaster severity and prioritize rescue operations. Mobirescue employs advanced optimization techniques and machine learning algorithms to allocate resources effectively, ensuring minimal response time and maximum coverage of affected areas. The system models the disaster environment as a dynamic network and applies route optimization strategies to identify the fastest and safest paths for rescue teams while considering constraints such as road blockages and resource limitations. Additionally, a priority-based scheduling mechanism is introduced to handle critical cases, ensuring that high-risk victims receive immediate attention.Experimental results demonstrate that the proposed approach significantly improves response efficiency compared to traditional dispatch methods, reducing rescue time and enhancing decision-making under uncertainty. The system also provides a scalable and user-friendly interface for emergency management authorities, enabling real-time monitoring and coordination. Overall, Mobirescue offers a robust and intelligent solution for disaster management, contributing to more effective emergency response and improved resilience in flood-prone regions

Keywords :

Author : R V S Ratna kumar,Penaganti Swathi,Kapusetti Prasanna Lakshmi Satya Sai Shirisha,Gutti. Venu Gopal, Kolakalur Hima Raja Sai Krishna

Title : Mobirescue Optimal Dispatching of Rescue Teams Under Flooding Disasters

Volume/Issue : 2026;03(04)

Page No : 215-222

Another reason of death in the world is still heart and blood vessel diseases and their early identification is highly significant in their treatment. The retinal fundus imaging is a painless method of examining the alterations in blood vessels of the heart that are as a result of heart conditions. Regular research methods will not, however, necessarily identify small patterns of blood vessels that indicate disease. Clinical features were obtained using retinal pictures by using Kaggle EyePACS dataset to enable multimodal prediction. In the preprocessing stage, the images were downsized, random horizontal flips and rotations were added, and transformed into tensors to be classified. Simultaneously, the files with detections were merged and configured to the object localization based on YOLO. The baseline classification models were ResNet50, VGG16, and a regular CNN, as well as the mixed EfficientNetV2 feature extractor, which is a combination of 1D CNN layers. Object recognition was done through YOLOv5, YOLOv8, YOLOv9, and YOLOv11 to identify and label issues. Accuracy, precision, recall, F1-score, specificity, sensitivity, and error were used to measure classification performance. Recognition was measured by precision, recall and mAP. The suggested hybrid classification model had the best accuracy (99.9%), as well as the best precision, F1-score, and specificity. YOLOv5 was the highest in terms of detection performance with a mAP of 0.787. The most significant areas to make predictions were demonstrated by means of GradCAM, and a web interface based on Flask allowed individuals to input data and view the outcomes in real-time. This was more predictive and localised the retinal issues on how to determine the early cardiovascular risk

Keywords : Retina , Heart , Feature extraction , Accuracy , Convolutional neural networks , Biomedical imaging , Risk management , Machine learning , Deep learning , Data models

Author : A Dhanasekhar Reddy, P Sernivasulu, Dunna Nikitha Rao

Title : Advanced Retinal Image Analysis for Cardiovascular Disease Risk Prediction Using Hybrid Deep Learning

Volume/Issue : 2026;03(04)

Page No : 223-232

Today, cloud storage systems have huge issues with the process of securely erasing duplicate data and verifying the integrity of data fast and accurately. It is particularly so when active processes such as the addition or removal of blocks are taken into consideration. The traditional compression requires a lot of additional computing power since whole files must be re-encrypted and verified each time the data is modified. To bypass these issues, DIADD introduces a deduplication and data integrity check safe system, which is compatible with changing data. The scheme allows the owners of data to transmit encrypted data to other parties besides generating unique file tags through a multi-set hash facility. This will ensure that future users do not end up uploading the same information twice through proof of ownership methods. Using hash-based authenticators, integrity checks are given to a TPA. These authenticators are only used to verify the modified blocks rather than the entire file which reduces communication and computing expenses. It is also equipped with a safe and searchable trapdoor that allows you to search keywords in encrypted files stored in the cloud without revealing any information about you. The system also ensures that the storage process is efficient, the data operations are facilitated dynamically, data integrity is ensured and privacy ensured when searching. It offers a safe and scalable method of outsourcing data management in clouds.

Keywords : Cloud storage, data integrity auditing, data deduplication, and data dynamics

Author : D Asha, D Nagaraj

Title : Secure Cloud Storage with Dynamic Deduplication and Integrity Verification

Volume/Issue : 2026;03(04)

Page No : 233-242

Proper examination of the well-being of the fetus is quite significant in preventing issues throughout the course of pregnancy and ensuring that the mother and the baby achieve optimal outcomes. The data of CTG that is mostly utilized to monitor the health of the fetus provides valuable data regarding the health of the fetus. Manual analysis is however time consuming and may give various interpretations. In this study the UCI Fetal Health Dataset is utilized in testing various predictive modeling techniques. These are PCA, LDA and KPCA. This is aimed at enhancing the accuracy of the diagnosis through a decrease in the number of dimensions. The class mismatch was overcome using both the SMOTE and SMOTEENN resampling techniques. Some of the machine learning models that were experimented on their performance included LR, XGBoost, RF, KNN, Support Vector Machine SVM, and Voting Classifier which is a combination of KNN and XGBoost. With SMOTE, PCA-XGBoost and PCA-Voting Classifier were able to get 98% accuracy, LDA-Voting Classifier got 98.1%, and KPCA-XGBoost got 96.6%. With the use of SMOTEENN, PCA-Voting Classifier scored 98.9%, LDA-KNN scored 98.9% and KPCA-Voting Classifier scored 97.9%. Also, the model was improved with LIME and SHAP to enhance the interpretability of the model by showing characteristics of CTGs that are important contributors. This ensured that there was clarity in clinical decision-making. A Flask-based framework was also developed in such a way that the end predictive system would be able to monitor the health of a fetus in real-time. These results indicate that the number of dimensions reduced and sophisticated resampling, explainable AI, and explainable machine learning significantly enhance the capacity to predict the health of a fetus, which is strong and accurate and offers automatic diagnostic assistance

Keywords : Fetal health, cardiotocography (CTG), dimensionality reduction, PCA, LDA, KPCA, SMOTE, ensemble learning

Author : G Viswanath, D Chandini

Title : Optimized Machine Learning Approaches for Fetal Health Assessment Using PCA, LDA, and KPCA with Advanced

Volume/Issue : 2026;03(04)

Page No : 243-255

The use of WSNs is gaining significance in the real-time data gathering, environmental monitoring, and security. But it is easy to be hacked into them and undermine their security and reliability. An explainable ensemblebased intrusion detection system was developed using the Kaggle Wireless Sensor Network Dataset that was capable of addressing this issue. In the case of hyperparameter optimization, the algorithm is Particle Swarm Optimization (PSO) and GridSearchCV. It also applies various models of machine learning, including DT + PSO, RF + PSO,KNN + PSO, XGB + PSO, and a hybrid ensemble based on LightGBM and ExtraTree, (RF + DT) + PSO, (RF + KNN) + PSO, (RF + KNN + XGB) + The model results were elucidated with the help of AI methods that could be described, such as LIME and SHAP, which simplified the intrusion detection process. The experiment indicated that the stacking ensemble performed better than any single model and accurately detected threats with 98.1 precision and recall and F1-score. It demonstrates that it is a good and comprehensible solution to WSN intrusion detection. The web application is a Flaskbased application that allows users to log in and add features dynamically, prepare attacks and categorize them into the list of Normals, Blackhole, Flooding, Grayhole, and TDMA.

Keywords : Wireless Sensor Networks, Intrusion Detection, Particle Swarm Optimization, Ensemble Learning, Decision Tree, Random Forest, K-Nearest Neighbors, Explainable AI

Author : K S Lokesh, Dunna Nikitha Rao, C Likhitha

Title : Smart Intrusion Detection in Wireless Sensor Networks Using Optimized Ensemble Learning

Volume/Issue : 2026;03(04)

Page No : 256-266

The correct detection of brain tumors in MRI images is required to deliver quick detection and efficient treatment, whereas various imaging characteristics complicate the functionality of automated practices. Clinical choices to be supported need to be supported by classification and localization techniques. The dataset is constituted by MRI scans which are labeled by tumor classification labels and bounding box labels. It has a great variety of tumor types and imaging characteristics, which makes it a good point of entry when it comes to the testing of automatic detection and classification systems. The proposed method uses transfer learning-based architectures such as VGG16, ResNet, Inception and DenseNet in classifying brain tumors. The models are evaluated using accuracy, precision, recall and F1-score which offer a strict method of comparative performance. To further improve things, we may resort to Xception, a collection of DenseNet and Xception, variants of YOLO such as YOLOv5, YOLOv8, YOLOv9, and YOLOv11 in order to locate and identify tumors, Grad-CAM to simplify the heatmap interpretation, and an interface based on Flask that allows interactive clinical usage. The ensemble model performs better than all the other classification models with 99.3 across all the metrics. The YOLOv8 has the highest detection mAP of 89.6 and a good Precision Recall balance. This demonstrates that the framework is robust and sound enough to be used in clinical practice on the precise classification and localization of brain tumors.

Keywords : Brain tumor detection, transformative transfer learning, MRI, deep learning, transfer learning, medical imaging

Author : K Vishnu Vardhan, G Swapna

Title : Brain Tumor Analysis in MRI Using Hybrid Transfer Learning and YOLO-Based Detection for Improved Diagnostic Accuracy

Volume/Issue : 2026;03(04)

Page No : 267-276

Reservation cancellations are a massive issue in the hospitality sector as they disrupt the demand forecasts and cost businesses away 20 percent of their revenues. You must be capable of making proper predictions regarding cancellations in order to optimize the operations, price and resource management of your hotel. The dataset that we used was Hotel Booking Demand, with 32 characteristics of the hotel of a resort and city. These characteristics consist of categorical and numerical data regarding customer booking and their type of customer and the specifics of their booking. Preprocessing involved a process of dealing with missing values, eliminating duplicate values and the process of evening out the distribution of classes using the RandomUnderSampler. Some of the classification techniques employed are LR, DT, RF, XGBoost, Gradient Boosting, LGBMClassifier, SVM, and MLP. The hyperparameters were fine-tuned with the help of gridSearchCV. The precision of the forecasts was enhanced through Stacking Classifier where the major model consisted of XGBoost, Random Forest, Gradient Boosting, and Logistic Regression. Categorical features were further encoded as labels to make them even more accurate with the help of Label Encoding and RFE was utilized to identify the most significant factors that influenced cancellations. The Voting Classifier which used a combination of Decision Tree and MLP models increased the accuracy to 98.7%. The importance of each feature was also determined using XAI tools such as LIME and SHAP. Flask-based user interface enabled real time and live prediction to aid good decisions by the hotel management systems.

Keywords : Cancellation forecasting, hotel booking, artificial intelligence, machine learning, revenue management, explainable artificial intelligence

Author : R R Shantha Spandana, M S Arumugam, Syed Jeelan

Title : Hotel Booking Cancellations Prediction Using Machine Learning Techniques for Optimized Management

Volume/Issue : 2026;03(04)

Page No : 277-286

In bilingual e-commerce systems, it's important to get a good sense of how customers feel in order to increase user satisfaction and purchase behavior. Traditional methods of mood analysis often have trouble with different languages and unclear contexts. The data came from online shopping and product review datasets on Hugging Face and GitHub, which have textual comments in two languages. In the preprocessing step, the text was normalized, HTML and URLs were removed, it was tokenized, labels were encoded, and sentiment was extracted using NLTK's SentimentIntensityAnalyzer. Next, NLTK and BERT were used to create the embedding. A lot of different models were used, such as Linear Regression (SGD), SVM (SGD), Random Forest, XGBoost, RF + XGB, BiLSTM, ELECTRA, and BERT. There were also advanced hybrid designs like BERT with LSTM and LSTM+GRU, and a Voting Classifier that mixed GB, XGB, LGBM, and CatBoost. SMOTEENN was used to balance the datasets, and Explainable AI tools (LIME and SHAP) made the results easier to understand. A Flask-based interface also let real-time mood prediction and topic modeling work with user input. Testing the Voting Classifier with Accuracy, Precision, Recall, and F1-score showed that it was the most accurate, with a score of 99.1% on Dataset A and 98.7% on Dataset B. This proved that it was stable and had the best predictive performance in classifying sentiment in two languages.

Keywords : E-commerce, customer satisfaction, sentiment analysis, natural language processing (NLP), sentiment keywords, feature-level sentiment analysis, consumer behavior, customer feedback

Author : M Tharun Kumar, E Pavithra, R R Shantha Spandana

Title : Improving Bilingual E-Commerce Customer Review Analysis Using a Combined Machine Learning and Deep Learning Approach for Sentiment Prediction

Volume/Issue : 2026;03(04)

Page No : 287-297

The Industrial Internet of Things (IIoT) makes Industry 4.0 possible by making industrial control systems more automated, efficient, and able to watch things in real time. Even with these benefits, IIoT infrastructures that are highly connected are open to major cybersecurity threats that can compromise their privacy, integrity, and availability. This paper shows a cyber risk assessment method based on machine learning that can be used to find threats before they happen in IIoT environments. A lot of different supervised models are tested, such as Decision Tree, Support Vector Machine, XGBoost, LightGBM, Random Forest, Logistic Regression, K-Nearest Neighbors, Decision Tree, FSVM, FXGBoost, and ensemble vote classifiers. The experimental results show that the group voting-based model works best, with a score of 0.993 and a 99.3% accuracy rate, doing better than individual learners. Artificial intelligence methods that can be explained, like LIME and SHAP, are used to improve clarity by figuring out how features affect predictions. For real-world use, the trained model is put together using a Flask-based web framework that lets users connect and assess risks in real time. The system sorts IIoT cyber risk levels into outputs that are Very Low, Low, Medium, High, and Very High. This helps people make quick choices about how to reduce the risk. Overall, the suggested method shows a strong, clear, and expandable answer that effectively boosts IIoT cybersecurity resistance.

Keywords : Industrial Internet of Things, cybersecurity, cyber threats, risk assessment, machine learning, federated learning, cyber threat intelligence, STRIDE threat modeling

Author : N Harini, Jithendra Kukutla, K Yatheendra

Title : A Machine Learning Driven Framework for Proactive Cyber Risk Detection and Classification in Industrial Internet of Things Systems

Volume/Issue : 2026;03(04)

Page No : 298-307

Cerebral stroke is a big global health problem that causes a lot of death and long-term disability. This shows how important it is to find people who are likely to have a stroke early on so that they can be prevented and treated quickly. The study uses Kaggle's Stroke Prediction Dataset, which has information like gender, age, high blood pressure, heart disease, glucose level, body mass index (BMI), and smoking status to figure out how likely someone is to have a stroke. After label encoding, one-hot encoding, and Random Forest Imputation are used to fix missing BMI values in the data, it is visualized using class distribution charts and correlation matrices. Normalizing the dataset with MinMax, Standard, and Robust scalers is done, and to fix class imbalance, oversampling methods like SMOTE, ADASYN, and Random oversampling are used. A number of machine learning and deep learning models are used to improve performance. These include Logistic Regression, KNearest Neighbors, Decision Tree, Random Forest, Support Vector Machine, Gaussian Naive Bayes, 1D-CNN, GRU, LSTM, ANN, and BiRNN. To make the predictions even better, ensemble learning methods are looked into. A Voting classifier that combines XGBoost, Random Forest, and AdaBoost does the best, with 99.4% accuracy, precision, recall, F1-score, and ROC-AUC. Explainable AI techniques like LIME and SHAP are also used to figure out how important each trait is, and a web-based interface built on the Flask framework and integrating SQLite is created to let users make predictions in real time

Keywords : Stroke prediction, Machine learning, Ensemble learning, Deep learning, Explainable AI, Flask framework, Data preprocessing, Health informatics

Author : R Venkatesh, K Dhanamjay, K Yatheendra

Title : Early Detection of Stroke Risk Through Intelligent Health Data Modeling and Machine Learning

Volume/Issue : 2026;03(04)

Page No : 308-318

For modern financial analytics to work, credit scoring needs to be strong across both organized and unstructured data. This is because traditional statistical and shallow learning methods don't work well with complex non-linear, temporal, and semantic dependencies. A unified system based on deep learning is created for both categorizing text and predicting numerical credit risk. A public credit risk dataset and standard text datasets (News, SciHTC, and UTCD) are used in the experiments. A common pipeline for preprocessing, visualization, and encoding is used, with changes made based on the job. We use contextual transformer embeddings to model text data and compare basic neural classifiers with a hybrid loss-driven LSTM and an ELECTRA-based design to improve how the data represents meaning. For numerical prediction, recursive feature elimination is the best way to choose features, and SMOTE helps even out class mismatch. We test basic machine learning models against a mixed LSTM and a softvoting ensemble that combines XGBoost and LightGBM. Accuracy, precision, recall, F1-score, and AUC are some performance measures. The ensemble is 95.7% accurate at predicting credit risk, and ELECTRA is 96.1%, 97.3%, and 99.3% accurate on News, SciHTC, and UTCD. LIME and SHAP make the predictions easy to understand, and a Flaskbased interface lets people guess and interact with the models in real time

Keywords : Credit scoring, deep learning, LSTM, transformer embeddings, explainable AI, financial risk prediction

Author : K Bhaskar, S Mohan Krishna, K Keerthana

Title : Integrated Deep Learning Strategy for Hybrid Credit Risk Prediction and Text Classification Tasks

Volume/Issue : 2026;03(04)

Page No : 319-331

The rapid growth of connected vehicles has positioned Vehicular Ad Hoc Networks (VANETs) as a critical component of intelligent transportation systems, where accurate vehicle identification and efficient routing prediction are essential for improving traffic safety and reducing congestion. Traditional approaches, including manual, heuristic, and rule-based methods, are limited in their ability to process high-velocity and high-dimensional data, resulting in suboptimal route selection and inaccurate vehicle spacing predictions. Existing machine learning techniques such as K-Nearest Neighbors (KNN), Gaussian Process (GP), Stochastic Gradient Descent (SGD), and Classification and Regression Tree (CART) provide partial improvements but fail to capture complex non-linear relationships among key vehicular features. To address these limitations, this study proposes a novel hybrid model named Fusion Mind CART, which integrates a Multi-Layer Perceptron (MLP) for deep feature extraction with a Random Forest (RF) ensemble for classification and regression tasks. The model effectively learns complex interactions among features such as vehicle speed, traffic density, signal strength, packet loss rate, route stability, and vehicle spacing. Experimental evaluation on realtime VANET datasets demonstrates superior performance, achieving 100% accuracy, precision, recall, and F1-score for route optimality classification, along with an R² score of 0.999 and minimal error rates for vehicle spacing prediction. The system is implemented using Python and Flask, enabling scalable and real-time deployment. The proposed framework provides a robust, data-driven solution for enhancing routing efficiency and predictive accuracy in dynamic vehicular environments

Keywords : Connected Vehicles, Machine Learning, Regression, Routing Prediction, Vehicular Ad Hoc Networks

Author : Kandikonda Sravanthi, R. Swathi, Rekha Gangula, Gadipelly Susmitha, Gudla Rahul, Mohammad Sharuk

Title : HIGH-PERFORMANCE MACHINE LEARNING FRAMEWORK FOR RAPID VEHICLE IDENTIFICATION IN CONNECTED ENVIRONMENTS

Volume/Issue : 2026;03(04)

Page No : 332-345

Coral reefs are among the most biologically diverse ecosystems, supporting nearly 25% of marine life while providing vital ecological, economic, and coastal protection benefits. However, factors such as climate change, ocean acidification, and human activities have significantly accelerated reef degradation, creating an urgent need for accurate and efficient coral species identification. Conventional coral identification approaches utilize handcrafted features such as shape, texture, and color, but these methods struggle with scalability and generalization across diverse underwater environments and visually similar species. To overcome these limitations, this study proposes an intelligent and scalable framework for underwater coral species identification using deep feature extraction and machine learning techniques. A Vision Transformer (ViT) model is employed to extract robust and high-dimensional feature representations from coral images. These features are then used to train multiple classifiers, including Natural Gradient Boosting (NGB), Histogram Gradient Boosting (HGB), Extreme Gradient Boosting (XGB), and a proposed Scope Rules Classifier (SRC). The proposed system is evaluated using performance metrics such as accuracy, precision, recall, F1- score, confusion matrix, and ROC analysis. Results indicate that the SRC model outperforms baseline methods in classification performance. Additionally, an Explainable Artificial Intelligence (XAI) module enhances interpretability, while a Telegram bot interface enables real-time coral species identification. This framework offers a practical, scalable, and interpretable solution for supporting marine ecosystem monitoring and conservation efforts.

Keywords : classification task, coral reefs, deep learning, marine ecosystems, vision transformer, XAI

Author : Meerupelli Meghana, P. Shivakumar, Korivi Sahitya, Basaraju Sriram, Nelakuditi Ranjith, Kinnera Sandeep

Title : SMART CORAL CLASSIFICATION FROM REEF IMAGERY USING DEEP CNN TECHNIQUES

Volume/Issue : 2026;03(04)

Page No : 346-356

The rapid increase in global wildfires, driven by climate change, demands a transition from reactive suppression to proactive, intelligence-based detection systems. Traditional wildfire monitoring methods, such as human surveillance and satellite remote sensing, often suffer from high latency, high operational costs, and limited temporal resolution. A major challenge in current automated systems is the "false alarm" problem, where modern Convolutional Neural Networks (CNNs), including You Only Look Once (YOLO), misclassify non-fire elements like sunsets, industrial glare, or red-colored objects as fire. These inaccuracies lead to unnecessary emergency responses and increased alert fatigue among authorities. To address these limitations, this research proposes VLM-FireNet, a hybrid cascade architecture that combines the speed of edge computing with advanced contextual reasoning. The system utilizes YOLOv8 for rapid initial detection at the edge, achieving inference times below 50 milliseconds. Detected instances are then validated using a Transformer-based Vision-Language Model (VLM), which applies a global self-attention mechanism to analyze the broader scene context and eliminate false positives. This dual-check framework significantly enhances detection reliability. The system is implemented using a multithreaded Python environment, integrating a local Tkinterbased interface with a remote Telegram Bot API for real-time alert notifications. The proposed approach improves detection accuracy while maintaining real-time performance. By reducing false positives by approximately 20%, VLM-FireNet provides a scalable and cost-effective solution for smart city and forest monitoring, contributing to more efficient and reliable disaster management systems.

Keywords : Wildfire Detection, False Alarm Reduction, Convolutional Neural Networks (CNN), YOLOv8, Vision-Language Model (VLM), Hybrid Cascade Architecture

Author : Thotakuri Rachana, G. Neeraja, Rekha Gangula, S. Mahanvitha, Kaluva Manasa, Kandula Nagaraju

Title : Deep Learning–Driven Object Detection for Wildfire Management and Early Intervention

Volume/Issue : 2026;03(04)

Page No : 357-366

Texture classification has become a significant area of research in computer vision–based industrial inspection, where accurate surface analysis is essential for maintaining product quality and detecting manufacturing defects. In earlier practices, inspection was carried out manually by human experts who visually identified defects such as cracks, holes, and foreign objects. However, these methods were time-consuming, inconsistent, and prone to human error. Later, traditional automated approaches based on classical image processing techniques were introduced, but their reliance on handcrafted features and simple classifiers limited their ability to capture complex texture patterns. With the rapid growth of industrial datasets in terms of size and diversity, there is an increasing need for intelligent systems capable of extracting meaningful features and performing accurate multi-class classification. To address these challenges, this study proposes a transfer learning–based texture classification framework that combines deep feature extraction with machine learning techniques. A pretrained Visual Geometry Group 19 (VGG19) network is employed to extract high-level visual features from texture images. These features are then used to train multiple classifiers, including Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), Support Vector Classifier (SVC), and Extra Trees (ET). Experimental results show that the proposed VGG19 combined with the Extra Trees classifier achieves the highest classification accuracy of 97.31%, effectively distinguishing between texture categories such as good, hole, and objects. The results demonstrate that integrating deep learning–based feature extraction with ensemble machine learning methods significantly improves the reliability and efficiency of industrial texture classification systems.

Keywords : Texture Classification, Transfer Learning, VGG19, Deep Feature Extraction, Extra Trees Classifier, Industrial Inspection, Ensemble Learning, Multi-class Classification.

Author : E. Saivarshith, A. Nagarani, Rekha Gangula, Chitte Ganesh, Kusuma Vishnu Vardhan, C Shashidhar

Title : PATCH-ORIENTED SURFACE TEXTURE IDENTIFICATION IN REMOTE SENSING IMAGES LEVERAGING VISION TRANSFORMERS

Volume/Issue : 2026;03(04)

Page No : 367-376

The quality assurance of biscuits in modern food manufacturing environments plays a vital role in maintaining product consistency, consumer satisfaction, and brand reliability. Conventional manual inspection methods have been widely used; however, they are often inefficient, time-consuming, and susceptible to human errors, especially when identifying minor or subtle defects. To overcome these limitations, an AI-driven biscuit defect detection system was developed to automate the inspection process and enhance quality control standards within the industry. The system utilized advanced image processing techniques combined with artificial intelligence to analyze biscuit images and classify them into categories such as color defects, shape defects, object contamination, and defectfree samples. Initially, the dataset was pre-processed through resizing, normalization, and transformation into numerical formats suitable for model training. Various machine learning models, including Extra Trees Classifier (ETC), Linear Discriminant Analysis (LDA), and Light Gradient Boosting Machine (LGBM), were implemented and evaluated. In addition, a ResNet-based Convolutional Neural Network (CNN) was proposed to automatically extract deep and complex visual features from the images. The performance of all models was assessed using evaluation metrics such as accuracy, precision, recall, and F1-score, along with visualization tools like confusion matrices and ROC-AUC curves. Experimental findings indicated that the ResNet CNN model significantly outperformed traditional approaches, achieving an accuracy of approximately 98.98%. Furthermore, the system was integrated with a user-friendly graphical interface that facilitated dataset handling, model training, real-time prediction, and result visualization. Additional features, including batch processing and remote monitoring through a Telegram bot, enhanced system usability.

Keywords : Biscuit defect detection, quality assurance, image processing, machine learning, Deep learning, defect classification, food manufacturing.

Author : Eppakayala Karunya, P. Shivakumar, Dayakar Thalla, Racharla Harish, R Karthika Hasini, Vasam Siddartha

Title : AI-Driven Biscuit Defect Detection System for Advanced Quality Assurance in Food Manufacturing

Volume/Issue : 2026;03(04)

Page No : 377-386

Early detection of foliar diseases in cauliflower crops is essential for minimizing yield loss and improving agricultural productivity. Conventional disease identification methods rely on manual inspection, which is often subjective, time-consuming, and impractical for large-scale monitoring. To address these limitations, intelligent systems based on Machine Learning (ML) and Deep Learning (DL) have emerged as effective solutions for automated and accurate disease diagnosis. This study presents a Transfer Learning (TL)-based framework for multi-disease in Cauliflower Leave Disease Classification (CLDC), integrating both DL and traditional ML approaches within a unified system. The dataset is preprocessed by resizing images to 64×64 pixels, normalizing pixel values, and organizing them into eleven disease categories. A hybrid Inception Residual Network Convolutional Neural Network (IRNCNN) model is developed using Convolutional Neural Networks (CNN), where deep features are learned through customized inception-residual blocks. The extracted features are subsequently classified using Logistic Regression (LR) to improve classification performance. In addition, baseline models including LR, Decision Tree Classifier (DTC), and Artificial Neural Network (ANN) are implemented to enable comparative evaluation. The system is deployed using a Tkinter-based Graphical User Interface (GUI), supporting dataset upload, preprocessing, model training, evaluation, and prediction. Batch prediction on image folders is also supported, with results exported in CSV format. Furthermore, an Explainable Artificial Intelligence (XAI) module is integrated to analyze input images, verify whether they correspond to cauliflower, and provide insights such as disease severity and affected plant parts. A Telegram Bot interface is incorporated to enable real-time mobile-based disease detection and user interaction. Experimental results demonstrate that the proposed IRNCNN model achieves high classification accuracy and robust performance across multiple disease categories. The integration of DL-based feature extraction, ML-based classification, XAI analysis, and multi-platform deployment establishes the system as a scalable and effective solution for real-time crop disease monitoring and precision agriculture

Keywords : Cauliflower Leaf Disease Classification (CLDC), Transfer Learning (TL), Deep Learning (DL), Machine Learning (ML), Image Classification, Plant Disease Diagnosis.

Author : D. Gouthami, Chowdari Samitha, S Vamshi Krishna, Keerthi Shiva Kumar, Maidam Vandana

Title : Exploring Hybrid Convolutional Representations for Early-Stage Foliar Disease Recognition in Cauliflower

Volume/Issue : 2026;03(04)

Page No : 386-397

Customer churn prediction has emerged as a vital analytical process for organizations seeking to improve customer retention, strengthen marketing strategies, and ensure stable revenue growth in competitive markets. With the rapid increase in digital customer interactions, there is a growing need for intelligent systems that can accurately identify customers who are likely to disengage or reduce their activity. Conventional churn analysis methods relied on manual evaluation using simple statistical techniques such as averages, rule-based thresholds, and spreadsheet analysis, which are insufficient for capturing complex behavioral patterns in large datasets. To overcome these challenges, this study introduces a machine learning-based predictive framework built on a Classification and Regression Tree (CART) approach. The system utilizes supervised learning to analyze multiple customer-related features, including demographic information, purchase history, and digital engagement behavior, to predict both churn risk and customer satisfaction levels. The framework evaluates several models, including Passive-Aggressive (PA), Support Vector Machine (SVM), Extra Trees (ET), and a novel hybrid model called Poly-BoostNet. The Poly-BoostNet model integrates Recurrent Polynomial Network (RPN) feature expansion with Categorical Boosting (CB), allowing it to effectively capture nonlinear relationships and complex feature interactions. Experimental results indicate that while PA and SVM offer faster learning and moderate accuracy, Poly-BoostNet achieves superior performance, delivering the highest R² score in regression tasks and achieving 100% accuracy in classification of customer engagement. The system is developed using Python-based machine learning techniques and deployed via a Flask web application, providing a complete pipeline for data preprocessing, model training, evaluation, and real-time prediction. This framework offers a scalable and efficient solution for data-driven customer retention strategies.

Keywords : Customer Churn Prediction, Machine Learning, Classification and Regression Tree (CART), Supervised Learning, Poly-BoostNet, Recurrent Polynomial Network (RPN), Categorical Boosting (CB)

Author : D. Ramireddy, Gurram Murali, Chevva Anjali, Umme Aiman Faiza, Avula Karthik

Title : Poly-BoostNet Enhanced Consumer Churn Analytics With CART-Driven Insights for Business Decision Systems

Volume/Issue : 2026;03(04)

Page No : 398-409

This study addresses the growing need for intelligent, data-driven systems in the education sector to analyze student performance and behavioral patterns. Educational datasets encompass critical factors such as study habits, attendance, and engagement, which can be utilized to predict academic outcomes and stress levels. However, traditional machine learning models including Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting (GB), Decision Tree (DT), Logistic Regression (LR), and Linear Regression (LinR) often struggle to capture complex nonlinear relationships and sequential dependencies within such data. Additionally, these models face challenges such as overfitting, limited generalization, and reduced interpretability, particularly when handling both classification and regression tasks within a unified framework. To address these limitations, this research proposes a hybrid model, Bidirectional LSTM with Attention and Rule-based Model (BARM). The model integrates Bidirectional Long Short-Term Memory (BiLSTM) with an attention mechanism to effectively capture temporal patterns and feature importance, alongside rule-based approaches including Optimal Decision Rule List Classifier (ODRLC) and Optimal Decision Rule List Regressor (ODRLR) for enhanced interpretability. The system performs classification tasks such as Final Grade and Stress Level prediction, as well as regression tasks like Exam Score estimation. A dynamic model selection mechanism is employed to evaluate and choose the most suitable model based on performance metrics. The proposed system is implemented as a Flask-based web application, enabling real-time predictions, batch processing, and visualization through Exploratory Data Analysis, By improving scalability, usability, and decision-making in academic environments.

Keywords : Educational Data Mining, Machine Learning, Deep Learning, Hybrid Model, Bidirectional LSTM (BiLSTM), Attention Mechanism, Rule-Based Learning

Author : N. Rajesh, Mukkala Rachana, Konduru Rakesh, Thipparapu Anjali, Gantala Chandu

Title : Deep-Feature Stacking Ensemble for Predictive Modeling of Student Academic Performance Using Socio-Behavioral and Institutional Factors

Volume/Issue : 2026;03(04)

Page No : 410-418

Urban traffic management and public safety have become increasingly dependent on real-time monitoring of traffic and environmental events. Traditional methods for detecting traffic accidents, congestion, and urban crimes often rely on manual surveillance, human observation, or conventional sensors, which are limited in scalability, accuracy, and real-time responsiveness. Such approaches are prone to human error, delayed reporting, and incomplete data capture, making it challenging to ensure timely interventions and efficient traffic management. To address these limitations, this research proposes an automated audio-based classification system for urban traffic and acoustic events. The system leverages modern machine learning and deep learning techniques to process audio signals captured from urban environments and classify them into multiple event categories, such as accidents, honking, traffic jams, and urban crime-related sounds. The core of the system uses the Transformer Encoder for Representations of Audio (TERA) model for feature extraction, which efficiently encodes temporal and spectral patterns in raw audio into high-dimensional embeddings. These embeddings serve as input for several supervised classifiers, including Categorical Boosting Classifier (CatBoost), Histogram Gradient Boosting (HGB), Extra Trees Classifier (ETC), and a proposed Tree-based Generalized Additive Model (TGAM). The proposed approach significantly improves classification performance compared to traditional methods by capturing complex audio patterns and providing robust multi-class predictions. Experiments demonstrate that the TGAM model achieves high classification accuracy, outperforming baseline ensemble models, with macro-average precision, recall, and F1-scores consistently above 90% on a balanced urban traffic dataset. The system also includes visualization tools such as confusion matrices, ROC curves, and waveform overlays, allowing users to interpret predictions and model performance intuitively. The research is implemented as a role-based Tkinter GUI application, providing separate dashboards for administrators and general users. Administrators can upload datasets, extract features, train models, and evaluate results, while users can perform real-time predictions on new audio files. User authentication is secured using TinyDB with SHA-256 password hashing.

Keywords : Urban Traffic Monitoring, Audio Event Classification, Machine Learning, Deep Learning, Transformer Encoder, Feature Extraction, TGAM Model

Author : K Srinivas, Gandla Sai Deepthi, Banoth Anjali, Shaik Fareed, Mohd Atif Ahmed

Title : A Deep Transformer-Based Framework for Intelligent Acoustic Event Detection in Urban Surveillance Systems

Volume/Issue : 2026;03(04)

Page No : 419-427

Demand forecasting is a crucial component in the retail domain, as it directly affects inventory planning, pricing decisions, and overall operational efficiency. The increasing variability in customer preferences, seasonal fluctuations, and competitive market conditions has reduced the effectiveness of conventional forecasting techniques. Traditional systems relied heavily on manual evaluation, historical data summaries, spreadsheets, and basic statistical approaches. Such methods were dependent on human judgment and past trends, which limited their scalability, adaptability, and prediction accuracy. Additionally, they lacked automation, real-time processing, and the capability to model complex relationships among multiple influencing factors, often resulting in inefficient inventory management. To address these challenges, the proposed system introduces a Machine Learning (ML)-based automated framework for demand forecasting. The system integrates multiple regression models, including K-Nearest Neighbors (KNN), Decision Tree Regressor (DTR), and Gradient Boosting Regressor (GBR), along with a custom TabularNet with Extreme Gradient Boosting (XGB) model designed to capture meaningful feature representations from structured retail datasets. In the current implementation, although TabNet-XGB is trained to learn feature patterns, the final predictions are generated using a XGB tree-based regressor, without direct integration of TabNet outputs into the prediction stage. For performance evaluation, baseline models such as KNN, DTR, and GBR are compared using standard regression metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R-squared (R²) score. Artificial Intelligence and ML (AIML) Engineers are enabled to perform model training.

Keywords : Demand Forecasting, Machine Learning Models, XGB Regressor, Retail Analytics, Inventory Optimization.

Author : G. Naveen Kumar, Bandi Greeshma, Puli Kavya, Kalthi Viswateja, Mora Sai Manoj

Title : Retail Sales Demand Prediction using TabNet-Enhanced XGB in an Interactive Python Desktop Interface

Volume/Issue : 2026;03(04)

Page No : 428-435

The evolution of Natural Language Processing (NLP) has shifted from static, rule-based systems to sophisticated Transformer architectures, particularly in the domain of technical recruitment and academic assessment where the automated evaluation of coding interview questions has become a critical benchmark. Traditional systems for interview analysis typically rely on manual tagging or basic keyword-matching algorithms, which serve as legacy frameworks that often fail to capture nuanced difficulty levels or the underlying sentiment of complex technical problems, leading to inconsistent evaluations. These existing methodologies suffer from significant limitations in semantic depth and are unable to provide real-time, context-aware explanations or adaptive scoring. Consequently, there is an urgent need for an automated system that integrates classical machine learning with advanced generative reasoning to handle growing volumes of technical data. Therefore, this research introduces an integrated framework utilizing a Distributed Multimodal Language Model (DMLM) Interface alongside a multi-model supervised learning suite comprising Logistic Regression Classifier (LRC), Stochastic Decision Trees Classifier (SDTC), Gradient Boosting Classifier (GBC), and Adaptive Boosting Classifier (ABC) models. The proposed system processes a proprietary dataset through a pipeline employing TF-IDF vectorization for structural classification and a DMLM-driven inference engine for high-level semantic reasoning and explanation generation. The significance of this research lies in its ability to bridge the gap between statistical pattern recognition and generative intelligence by evaluating model performance through metrics such as precision, recall, and F1-score, ultimately offering a scalable, hybrid solution for academic strategists and recruiters to optimize interview question banks with high-fidelity sentiment and difficulty analytics

Keywords : Natural Language Processing, Transformer Architectures, Distributed Multimodal Language Model (DMLM), TF-IDF Vectorization

Author : D. Gowthami, D Sandeep Kumar, Bojjam Indhu, Bandaru Ganapaiah, Sada Swathi

Title : A Hybrid Stochastic Decision Tree Framework for High-Fidelity Classification and Semantic Analysis of Technical Interview Questions

Volume/Issue : 2026;03(04)

Page No : 436-450

With more than 4,500 active satellites currently orbiting Earth and a global dependency valued at over $300 billion annually on satellite-based services, ensuring mission reliability has become a critical operational priority. Real-world reports indicate that nearly 40% of unexpected mission losses arise from undetected telemetry anomalies, highlighting the urgent need for predictive intelligence capable of interpreting complex spacecraft data. Satellite telemetry captures vital information regarding subsystem health, environmental conditions, and mission progress, but its high dimensionality, noisy signals, missing values, and heterogeneous behavior pose substantial challenges for conventional analytics. The research presents a robust and scalable solution for satellite telemetry analysis by developing a locally deployed ensemble learning system for mission success classification and duration prediction using OPSSATAD data. The system integrates multiple machine learning algorithms by combining classification and regression, referred to as classification and regression tree (CART), including Random Forest Classifier (RFC), Gradient Boosting Classifier (GBC), and Support Vector Classifier (SVC) for classification, along with Random Forest Regressor (RFR), Gradient Boosting Regressor (GBR), and Support Vector Regressor (SVR) for regression tasks. Additionally, the proposed framework introduces an advanced ensemble model that combines Extra Trees Classifier (ETC) and CatBoost Classifier (CBC) for classification, and Extra Trees Regressor (ETR) and CatBoost Regressor (CBR) for regression using a voting-based aggregation method, referred to as the proposed Ensemble Voting (CART) model. The ETCART models effectively capture hierarchical decision patterns, while CatBoost-CART improves performance through boosting and handling complex feature interactions.

Keywords : CatBoost (CBC/CBR), Extra Trees (ETC/ETR), Gradient Boosting, Random Forest, Support Vector Machines (SVM)

Author : B. Prashanthi, Pittala Amulya, Devunuri Sreeja, Bhilla Noordas, Bhukya Rohith

Title : Astro Ensemble: A CART-Boosted Learning Model for Satellite TelemetryBased Decision Systems

Volume/Issue : 2026;03(04)

Page No : 451-460

Cloud storage systems have become an integral part of modern data management by enabling users to store and access data remotely. However, traditional cloud storage architectures rely on centralized servers, which introduce several critical challenges such as single-point failure, redundant data storage, increased storage costs, and security vulnerabilities. In earlier systems, data was stored in centralized data centers where duplicate files were repeatedly maintained, leading to inefficient utilization of storage resources. Although basic deduplication methods were implemented, they often compromised data confidentiality and lacked transparency in metadata management. Furthermore, failure of the central server could result in permanent data loss. To overcome these issues, this research proposes the Blockchain-enabled Heuristic Optimized Deduplication Model (BHODM), which integrates blockchain technology, InterPlanetary File System (IPFS), Convergent Encryption (CE), and heuristic-based chunking techniques. In this model, files are divided into optimized chunks based on file size using a heuristic approach. Each chunk is encrypted using CE, where the encryption key is derived from the hash of the data itself, allowing secure deduplication without exposing plaintext data. Duplicate chunks are identified through hash comparison, ensuring that only unique data is stored. The encrypted chunks are stored in IPFS, a decentralized peer-to-peer storage network, while metadata such as file names, block numbers, and hash values are securely maintained in an Ethereum blockchain smart contract, ensuring immutability and transparency. The system is implemented using Django, Web3, IPFS API, and AES-CTR encryption. Experimental results based on storage utilization and computation time demonstrate improved efficiency over traditional approaches.

Keywords : Blockchain, InterPlanetary File System (IPFS), Deduplication, Convergent Encryption, Heuristic Chunking, Ethereum Smart Contract, Decentralized Storage, Data Security, Storage Optimization.

Author : Goutham Kunamalla, Buddineni Vinitha, Bairi Sai Vardhan, Gootla Ganesh, Chaparthi Sravani

Title : BHODM: A Blockchain-Based Heuristic Deduplication Model for Secure, Transparent, and Redundancy-Aware Cloud Storage

Volume/Issue : 2026;03(04)

Page No : 461-469

The rapid evolution of Twitter as a real-time information-sharing platform has led to an unprecedented surge in short-text data reflecting ongoing political and social events. Analyzing this vast and unstructured data manually is both inefficient and unreliable, particularly for time-critical applications such as crisis response and governance. Conventional machine learning techniques often fail to capture deep semantic and contextual relationships within textual data, thereby limiting their effectiveness in complex event classification tasks. To address these limitations, this study proposes an efficient political event detection framework that integrates SBERT-enhanced Lightweight Robustly Optimized Bidirectional Encoder Representation from Transformer (RoBERTa) for feature extraction with advanced machine learning classifiers. Initially, raw tweet data undergoes comprehensive Natural Language Processing (NLP) preprocessing, including tokenization, stop-word removal, normalization, and lemmatization, ensuring high-quality input. Exploratory Data Analysis (EDA) is conducted to uncover data distribution and linguistic patterns. Subsequently, semantic-rich embeddings are generated using a lightweight SBERT-based RoBERTa model, enabling context-aware feature representation. These features are then utilized by multiple classifiers, including Stochastic Gradient Descent (SGD), Histogram Gradient Boosting (HGB), Random Forest Classifier (RFC), and Greedy Tree Classifier (GTC). Furthermore, a Deep Neural Network (DNN) is incorporated for intermediate feature learning, followed by an optimized SGD classifier to enhance predictive performance. The proposed system classifies tweets into six categories: disaster, political, positive, protest, riot, and terror. Experimental results demonstrate superior accuracy, scalability, and generalization compared to traditional methods, highlighting its effectiveness for real-time political event detection and decision-support systems

Keywords : Political Event Detection, NLP, RoBERTa, SBERT, DNN, Machine Learning, Twitter Analysis, Text Classification, Real-Time Systems.

Author : Goutham Kunamalla, Mohammad Afroz, Macherla Pavithra, Mahankali Shivani, Eppalapally Sathwik

Title : An Efficient SBERT-Enhanced RoBERTa Feature Extraction Model with Machine Learning Classifiers for Real-Time Political Event Detection on Twitter

Volume/Issue : 2026;03(04)

Page No : 470-479