Title of the Paper: HGTDL: A Hybrid Graph-Temporal Deep Learning Framework for Crop Yield Prediction Using Multi-Source Satellite and Agronomy Data
Authors: J Amani, K Sekar, D Shobha Rani, Sathya Narayana Pola, A Naresh Kumar 
Abstract:
Accurate crop yield prediction is essential for agricultural planning, supply chain management and risk-aware advisories, particularly in regions where yield variability is highly influenced by rainfall patterns and variety selection. Traditional machine learning methods are unable to fully capture the combined effect of phenological growth and agronomic suitability, resulting in limited accuracy in diverse field conditions. This study proposes a hybrid deep learning framework that integrates spatial similarity, temporal phenology and agronomy knowledge to improve crop yield prediction. The proposed HGTDL model combines a graph neural network to represent agro-ecological proximity across fields with a long short-term memory encoder to capture the seasonal evolution of vegetation indices derived from multi-temporal satellite imagery. Crop-specific attributes such as duration and recommended production zones from the Crop Recommendation Dataset [22] are embedded through a fusion layer. Experiments were conducted using dataset [22], with a 70-15-15 train–validation–test split. Performance was assessed using RMSE, MAE and R² scores. The model achieved an RMSE of 0.37 t/ha, MAE of 0.28 t/ha and R² of 0.86, outperforming Random Forest, XGBoost and CNN–LSTM baselines by 5–28%. Stratified evaluation showed 6% higher accuracy in recommended zones and stronger stability for medium-duration varieties. Statistical tests confirmed that improvements were significant at p < 0.05. The results show that combining spatial and temporal learning with agronomy metadata provides reliable yield forecasts across crop types and regions. The framework supports practical applications in crop advisory systems, procurement planning and insurance risk estimation.
Keywords: Crop yield prediction, graph neural networks, LSTM, remote sensing, satellite imagery, agronomy data, hybrid deep learning, spatial–temporal modelling, vegetation indices
ISSN (Online): 2456-9852 | Year: 2025 | Volume: 10 | Journal Article | Publisher: IJECRT
Title of the Paper: A Secure and Intelligent Web-Based Hostel Management System with Role-Based Access Control and Blockchain-Enabled Data Integrity
Authors: Ketha Dayakar Reddy; S Vydehi
Abstract:
Hostel Management System (HMS) is an online solution that was created to automate and computerize the administration processes in student hostels. It is a centralized system, which ensures interaction among five diverse stakeholders, including Admin, Students, Wardens, Heads of Departments (HODs), and Parents. With the help of the HMS, students are able to register, make out passes, monitor the status of requests and send a feedback in an electronic format. Wardens have the power to track student activities, check on the feedbacks as well as communicate directly with parents whereas, the HODs have the authority to approve or disapprove out pass requests. Parents can have a real-time view of the activities in their child hostel and reply to messages by the wardens. The Admin underlies complete control with user management which includes student, warden in addition to HOD records besides providing analytical reports. The system is developed on the frontend interface based on HTML5, CSS 3 and JavaScript, and Python Django on the backend layer. Priority is given to security by determining the use of the encryption by using the Hash Algorithms Sha-256 to protect the personal information of the students and maintain confidentiality and integrity. This automated platform removes serious issues with the traditional hostel management which are manual error, sluggish approvals, and communication breach as well as data security lapses. The HMS helps in terms of decreasing the administrative workload, increasing the operational efficiency, and creating a more effective communication between all the stakeholders by simplifying the workflow and enhancing the level of transparency. The system has also provided a secure effective and user-friendly platform of administering a modern hostel system, which in the long-term benefits the students, parents and the staff of the institution.
Keywords: Hostel Management, Automation, Student Registration, Room Allocation, Out pass Requests, Warden Monitoring
ISSN (Online): 2456-9852 | Year: 2025 | Volume: 10 | Journal Article | Publisher: IJECRT
Title of the Paper: A Secure Federated Deep Learning Framework for Privacy-Preserving Real-Time Patient Monitoring in IoT Healthcare Systems
Authors: TVS Raghavendra; T Durga
Abstract:
The increasing adoption of Internet of Things (IoT) technologies in healthcare has enabled continuous patient monitoring and data-driven clinical decision support; however, it has also raised serious concerns related to data privacy, security, scalability, and computational efficiency. Conventional centralized learning approaches require sensitive medical data to be transferred and stored at central servers, making them vulnerable to data breaches and regulatory violations. The objective of this study is to design and evaluate a secure, privacy-preserving, and computationally efficient IoT healthcare framework that supports real-time patient monitoring without exposing raw medical data. The proposed framework integrates lightweight deep learning models with federated learning to enable decentralized model training across distributed healthcare clients. Physiological IoT sensor data are locally preprocessed and used for model training at the edge, while only encrypted model updates are shared for global aggregation. The system is evaluated using a publicly available Healthcare IoT dataset, with performance assessed under varying data distributions, noise levels, and federated configurations. Experimental results demonstrate that the proposed approach achieves an accuracy of 96.2%, a precision of 95.1%, and an F1-score of 94.9%, outperforming traditional machine learning and centralized deep learning baselines. Inference latency is limited to 18.4 ms per sample, making the framework suitable for real-time deployment. Robustness analysis shows stable performance under noisy and partially missing sensor data, while energy consumption remains lower than centralized deep learning models. The study concludes that the proposed federated IoT healthcare framework effectively balances intelligence, privacy, and efficiency, offering a practical solution for scalable and secure real-world healthcare monitoring applications.
Keywords: Internet of Things (IoT), Smart Healthcare, Federated Learning, Deep Learning, Patient Monitoring, Privacy Preservation, Edge Computing, Healthcare Analytics
ISSN (Online): 2456-9852 | Year: 2025 | Volume: 10 | Journal Article | Publisher: IJECRT
Title of the Paper: A Volatility-Aware Hybrid Deep Learning Framework for Robust Stock Price and Trend Forecasting Using Multi-Channel Market Signals
Authors: A Naresh kumar; D Shobha Rani; Sathya Narayana Pola; K Suresh d
Abstract:
Financial markets often experience irregular swings, sudden shocks, and unstable price behaviour, making short-term forecasting difficult for traditional and even modern learning-based systems. Many existing approaches depend heavily on past prices alone, causing them to perform poorly during high-volatility periods. This creates a practical need for models that can understand both stable market movements and unpredictable fluctuations. The objective of this study is to develop a volatility-aware hybrid deep learning framework that improves the accuracy and reliability of short-term stock price and trend prediction. The proposed method integrates a multi-channel feature pipeline, where historical prices, technical indicators, and volatility-specific signals are processed through parallel feature extractors. Temporal dependencies are encoded using Bi-LSTM and GRU layers, and the extracted representations are fused through a dedicated feature fusion block. The framework is evaluated using the S&P 500 five-year dataset, preprocessed with sliding windows and chronological splits. The model achieved an MAE of 0.87, RMSE of 1.14, MAPE of 1.92%, and a directional accuracy of 72.8%, outperforming baseline models including ARIMA, LSTM, CNN-LSTM, and GRU-based architectures. Statistical testing showed that the improvements were significant at p < 0.05, especially during volatile intervals. The findings indicate that the proposed approach offers stronger stability, more consistent trend prediction, and better handling of abrupt market swings. This contributes to practical applications such as automated trading, portfolio risk assessment, and short-term financial planning, demonstrating its potential as a dependable forecasting tool.
Keywords: Stock price forecasting; volatility modelling; hybrid deep learning; Bi-LSTM; GRU encoder; multi-channel feature fusion; financial time series; trend prediction; S&P 500 dataset; sequence learning; directional accuracy; technical indicators; deep neural networks; short-term market prediction.
ISSN (Online): 2456-9852 | Year: 2025 | Volume: 10 | Journal Article | Publisher: IJECRT
Title of the Paper: An Imbalance-Aware Deep Learning Framework for Real-Time Credit Card Fraud Detection Using PCA-Enhanced Transaction Patterns
Authors: Sathya Narayana Pola; D Shobha Rani; A Naresh kumar; K Suresh
Abstract:
Credit card fraud continues to pose a serious challenge for financial institutions due to the subtle and evolving nature of fraudulent behaviour in large-scale digital transactions. The high imbalance between legitimate and fraudulent records often limits the effectiveness of traditional classification systems. This study aims to develop an imbalance-aware fraud detection framework capable of identifying minority fraud patterns with high reliability while maintaining real-time response efficiency. The proposed method integrates behaviour-driven feature modelling, PCA-transformed transaction representations, class-weighted loss optimisation, and a compact deep neural network architecture. Experiments were conducted on the European Credit Card Fraud Dataset consisting of 284,807 transactions, where fraud accounts for only 0.172% of the total. The model achieved 93.12% precision, 91.44% recall, 92.26% F1-score, and a PR-AUC of 0.9479, outperforming logistic regression, random forest, SVM, and baseline neural networks across all major evaluation metrics. The system also demonstrated stable performance under varied train–test splits and maintained an average inference time of 0.31 ms per transaction, supporting real-time deployment requirements. Overall, the study provides an efficient and adaptable solution for financial fraud detection, offering enhanced accuracy, computational stability, and practical applicability in modern banking environments.
Keywords: Credit card fraud detection, deep learning, imbalanced datasets, PCA-transformed features, real-time transaction monitoring, class-weighted loss, anomaly detection, financial security analytics, neural network classification, precision-recall optimisation
ISSN (Online): 2456-9852 | Year: 2025 | Volume: 10 | Journal Article | Publisher: IJECRT
Title of the Paper: Privacy-Aware AI Models for IoT Multi-Sensor Predictive Analytics in Intelligent Clinical Decision Support Systems
Authors: K Suresh; D Shobha Rani; Sathya Narayana Pola; A Naresh kumar
Abstract:
Continuous monitoring of patients using wearable and bedside devices often produces large volumes of multi-sensor data, making it difficult for clinical teams to detect early signs of risk. Many existing systems analyse only a single signal or struggle to handle noise and irregular sampling, which limits their usefulness in real hospital settings. This study focuses on improving the reliability and accuracy of predictive analytics in such conditions. The main aim of this work is to build a practical and privacy-preserving model that can study multi-sensor readings and identify potential risk levels at an early stage. The method uses a cleaned and windowed version of the Multi-Sensor Medical IoT Dataset, followed by a hybrid CNN–BiLSTM architecture that learns both local signal changes and longer temporal patterns. A dual-threshold alert mechanism is included to convert model scores into simple categories-normal, mild risk, and high risk-so that clinicians can interpret the results easily. The proposed model achieved an accuracy of 94.82%, a recall of 93.71%, and an AUC of 0.95, all of which were higher than those of earlier approaches tested under similar conditions. It also showed lower false-negative rates and more stable behaviour across different noise levels and validation folds. In conclusion, this work offers a dependable way to convert multi-sensor patient data into timely alerts that support hospital staff in early decision-making. The approach can be extended to wearable devices, tele-health monitoring, and continuous ward surveillance with further optimisation.
Keywords: Multi-Sensor Healthcare Analytics, Predictive Modelling, CNN–BiLSTM, IoT Medical Data, Risk Scoring, Clinical Decision Support, Privacy-Preserving Preprocessing, Threshold-Based Alerts, Physiological Signal Analysis.
ISSN (Online): 2456-9852 | Year: 2025 | Volume: 10 | Journal Article | Publisher: IJECRT