Rajesh Kumar Dhanaraj
56884774100
Publications - 12
Proximal Policy Optimization-based Task Offloading Framework for Smart Disaster Monitoring using UAV-assisted WSNs
Publication Name: Methodsx
Publication Date: 2025-12-01
Volume: 15
Issue: Unknown
Page Range: Unknown
Description:
Unmanned Aerial Vehicles (UAVs) are increasingly employed in Wireless Sensor Networks (WSNs) to enhance communication, coverage, and energy efficiency, particularly in disaster monitoring and remote surveillance scenarios. However, challenges such as limited energy resources, dynamic task allocation, and UAV trajectory optimization remain critical. This paper presents Energy-efficient Task Offloading using Reinforcement Learning for UAV-assisted WSNs (ETORL-UAV), a novel framework that integrates Proximal Policy Optimization (PPO) based reinforcement learning to intelligently manage UAV-assisted operations in edge-enabled WSNs. The proposed approach utilizes a multi-objective reward model to adaptively balance energy consumption, task success rate, and network lifetime. Extensive simulation results demonstrate that ETORL-UAV outperforms five state-of-the-art methods Meta-RL, g-MAPPO, Backscatter Optimization, Hierarchical Optimization, and Game Theory based Pricing achieving up to 9.3 % higher task offloading success, 18.75 % improvement in network lifetime, and 27 % reduction in energy consumption. These results validate the framework's scalability, reliability, and practical applicability for real-world disaster-response WSN deployments. • Proposes ETORL-UAV: Energy-efficient Task Offloading using Reinforcement Learning for UAV-assisted WSNs • Leverages PPO-based reinforcement learning and a multi-objective reward model • Demonstrates superior performance over five benchmark approaches in disaster-response simulations
Open Access: Yes
A hybrid physics-informed neural and explainable AI approach for scalable and interpretable AQI predictions
Publication Name: Methodsx
Publication Date: 2025-12-01
Volume: 15
Issue: Unknown
Page Range: Unknown
Description:
Air Pollution is a critical environmental issue affecting public health, climate, and ecosystems. However, accurately predicting and classifying Air Quality Index (AQI) levels across different regions remains a challenging task due to the complex nature of air pollution patterns. Conventional and ensemble ML and DL models often fail to capture the physical laws goverming the air pollution, which leads to inaccurate predictions. This study addresses these issues by introducing an approach that employs Physics-Informed Neural Networks (PINN) with Explainable AI (XAI) techniques for AQI classification (AirSense-X). The proposed approach utilizes PINN for regression, along with mapping for classification and XAI for interpretation. PINN ensures that the model learns from physical laws governing air quality rather than relying solely on data. The dataset utilized in this study is a publicly available dataset containing the AQI data at daily levels from various stations across multiple cities in India. The proposed AirSense-X approach achieves an accuracy of 98 %, with 97 % precision, 95 % recall, and an F1 score of 0.96, ensuring reliability. Similarly, the confusion matrix for the proposed approach indicated that the model correctly classified 21,306 and misclassified 268 instances. The key focuses of this study include: • Introducing a novel approach, AirSense-X, which employs PINN for accurate AQI prediction and XAI for enhanced interpretability. Additionally, the study also involves comparative analysis with conventional and ensemble ML and DL models. • Employing structure mapping technique for classification based on the predicted AQI values. • Integrating physical laws governing air pollution using a PINN model enhances prediction accuracy and ensures that the model learns beyond relying on data-driven insights.
Open Access: Yes
Reliable generative interpretable framework for efficient predictive analysis of air quality index
Publication Name: Egyptian Informatics Journal
Publication Date: 2025-09-01
Volume: 31
Issue: Unknown
Page Range: Unknown
Description:
Air quality management is one of the most important sustainability goals in the era of Industry 5.0. The magnitude of air pollution and impact of drastic pollutants increase day by day despite the significant efforts of the environmental enthusiasts and researchers. The role of Artificial Intelligence (AI) in determining the Air Quality Index (AQI) is significant with reasonable accuracy of classification achieved. The proposed model is a multi-class problem, that classifies the AQI into six different classes. Various ML models such as Decision Tree (DT), Random Forest (RF), Adaptive Boosting (AdaBoost), Gradient Boosting(GB), Logistic Regression (LR). The RF provided reliable performance metrics for AQI category prediction, achieving an accuracy and Precision of 0.99. This model is selected for the implementation of Explainable AI (XAI) models such as Local Interpretable Model Agonistic Explainer (LIME) for explanation using the local surrogacy plots and SHapley Additive exPlanations (SHAP) explainer for the global surrogacy plots. The Generative Adversarial Network (GAN) can generate synthetic data, which addresses critical issues such as missing data, class imbalance, noise, and redundant data. The performance the GAN shows optimized performance in classification of the AQI data with accuracy closer to 100 %. This is mainly due to the synthetic data generated by the GAN which enhances the performance of the classification. The proposed work integrates the efforts of the GAN-AI-XAI that enhances the performance, reliability, trustworthiness and robustness of the AQI classification model.
Open Access: Yes
Adaptive few-shot tiny neural systems for real-time traffic intensity prediction in smart cities
Publication Name: ICT Express
Publication Date: 2025-01-01
Volume: Unknown
Issue: Unknown
Page Range: Unknown
Description:
The rapid evolution of urban mobility and smart city demands an intelligent transportation system which can make real-time decisions using lightweight and adaptive AI models. This research introduces a novel application of tiny machine learning which will combine the features of Few-shot learning algorithm and it will classify the traffic intensity levels on regional traffic data. By converting the traffic volume into three dynamic classes (Low/ Medium/ High), a compact neural network model is trained on episodic few-shot tasks that can mimic real-world low-data learning conditions. The proposed work supports open set classification which is more suitable for detecting unknown traffic behavior analysis by considering the previous day traffic level and how the future traffic intensity level can be predicted effectively. The accuracy of the proposed method is compared with the existing methods which lie with the baseline CNN (90 %) and SVM (89 %). But the average episode accuracy achieved through the proposed model is 95.2 % which makes this model promising for low-power edge deployment in intelligent transportation system.
Open Access: Yes
Heart disease prediction with a feature-sensitized interpretable framework for the Internet of Medical Things sensors
Publication Name: Frontiers in Digital Health
Publication Date: 2025-01-01
Volume: 7
Issue: Unknown
Page Range: Unknown
Description:
Introduction: Cardiovascular health is increasingly at risk due to modern lifestyle factors such as obesity, smoking, stress, hypertension, and sedentary behavior. Post-pandemic health practices and medication side effects have further contributed to rising cases of early heart failure, particularly among individuals aged 25–40 years. This highlights the need for an automated and interpretable framework to predict heart disease at an early stage. Methods: In this study, body vitals acquired from a secondary dataset. Machine learning models including Support Vector Machine, Random Forest, Decision Tree, and Logistic Regression were employed for classification. Model performance was evaluated using accuracy, F1-score, and k-fold cross-validation. Results: Among the tested models, the Random Forest classifier demonstrated superior performance with an accuracy and F1-score of 0.955. The interpretability is enhanced with model predictions were explained using Local Interpretable Model-Agnostic Explanations (LIME) for local surrogates and SHAP values for global surrogates. SHAP decision plots provided clear insights into classification behaviour and feature contributions. Discussion/Conclusion: The proposed interpretable machine learning framework successfully predicts heart disease with high accuracy while maintaining transparency in decision-making. With the integration of sensor data with cloud-based analysis and explainable AI techniques, this study contributes to reducing the incidence of early heart failures and supports more reliable decision-making in healthcare applications.
Open Access: Yes
MABAC model based on linguistic (p, q)-rung orthopair fuzzy Z-number and their application in green supply chain management
Publication Name: International Journal of Cognitive Computing in Engineering
Publication Date: 2026-12-01
Volume: 7
Issue: Unknown
Page Range: 247-267
Description:
The problem and complication arise from the growing environmental inefficiencies and concerns in traditional supply chains, for instance, poor accountability, excessive waste, and lack of transparency. The green supply chain practices aim to reduce or minimize the environmental impact of supply chain activities, but these efforts often face problems, for example, difficulty in monitoring sustainability performance, data manipulation, and limited traceability across numerous stakeholders. The main problem is that without effective techniques to verify and track eco-friendly practices, enterprises struggle to utilize and enforce green initiatives reliably. The blockchain technique is being derived as a solution because of its capability to give decentralized, transparent, and immutable records of processes and transactions. By integrating the blockchain into green supply chain practices, we aim to design the model of linguistic (p, q)-rung orthopair fuzzy Z-number sets with algebraic and Sugeno-Weber operational laws for the construction of the power weighted averaging operator and power weighted geometric operator. These operators can be used in the utilization of the multi-attributive border approximation area comparison model, which is also explained step-by-step with the help of examples to simplify the supremacy and validity of the invented model by comparing their ranking values with the ranking values of the existing approaches.
Open Access: Yes
Reliable power management and predictive analysis of domestic appliances with insights of XAI
Publication Name: Energy Reports
Publication Date: 2025-12-01
Volume: 14
Issue: Unknown
Page Range: 3704-3718
Description:
The unanimous focus of the sustainable technological development is energy conservation and environmental friendly production. Power management is an essential aspect of sustainable development. It not only support energy production and conservation, but also increases the life time of domestic appliances and thereby reducing the global electronic wastage. The existing systems involving Artificial Intelligence (AI) were mere prediction models, without the evidence on the detailing behind the prediction. Traditional AI systems have focused on predictive analysis but often lack transparency in decision-making and limiting consumer trust. This study proposes a solution combining remote power monitoring with the ZigBee module and Explainable Artificial Intelligence (XAI) to offer both predictive accuracy and interpretability. XAI models are more consumer oriented in every area of application, similar to the problem discussed, which tells about the impact of various parameters in power management in domestic appliances. Local Interpretable Model Agonistic Explainer(LIME) and SHAP explainer are used in the proposed work, providing explainability in the local and global surrogates. The proposed work applies various regression models such as Decision Tree (DT), Random Forest(RF), Support Vector Regressor (SVR), Gradient Boost Regressor (GBR) and Extreme Graident Boost Regressor (XGBR). The RF provides the best R2-Score of 94.71% , which is 1.5%–3.0% more than the rest of the models, and also with variance score of 68.82% , had been chosen for explainability. This study demonstrates how XAI can improve transparency and reliability in AI-powered domestic energy systems, offering actionable insights for more sustainable power consumption.
Open Access: Yes
A vision explainability method for image captioning using transformer decoder attention maps
Publication Name: Methodsx
Publication Date: 2025-12-01
Volume: 15
Issue: Unknown
Page Range: Unknown
Description:
Image Captioning is a crucial task that enables systems to generate descriptive sentences for visual content. Though image captioning systems bloom at the intersection of Computer Vision and Natural Language Processing, these models act mostly as black boxes offering little or no insight into how captions are derived. We present a novel explainable image captioning framework that integrates a Convolutional Neural Network encoder with a Transformer decoder. Attention-based heatmaps are used to explain the visuals offering transparency in the decision making process. The method evaluates captioning quality and interpretability on the MS COCO dataset using BLEU, METEOR, CIDER and SPICE. The method enhances the trustworthiness and transparency, making it reliable for applications like healthcare, education, security, surveillance and forecasting.A reproducible method for integrating visual explainability into image captioning exploring transformer decoder attention maps.The method contributes to the growing body of eXplainable AI (XAI) by addressing the transparency gap in vision-language modelsBalance performance with interpretability paving the way for more transparent and trustworthy AI systems.
Open Access: Yes
Spectral-aware CNN with learnable biorthogonal units and depthwise convolutions for multi-class blood cell classification
Publication Name: Methodsx
Publication Date: 2025-12-01
Volume: 15
Issue: Unknown
Page Range: Unknown
Description:
For effective and early diagnosis of diseases such as leukemia and anemia, accurate classification and interpretation of peripheral blood cells are critical. A novel hybrid deep learning model is proposed in this study for multi-class blood cell classification, called Spectral-Aware CNN with Learnable Spectral Biorthogonal Downsampling Units (LSBDUs) and Depthwise Separable Convolutions. The model replaces conventional pooling layers with wavelet-inspired LSBDUs for improved feature retention. This results in reduced computational overhead through efficient separable convolutions. The research used a balanced dataset of 17,092 images across eight blood cell classes. The techniques, such as stratified data splitting, advanced augmentation, and label smoothing, are included in the training pipeline for improving generalizability. As a result, the model achieves 99.18 % of overall classification accuracy with superior class-wise performance. • Replaces pooling layers with spectral-aware LSBDU blocks for better feature preservation. • Integrates Depthwise Separable Convolutions to reduce parameter count and training cost. • Demonstrates superior generalization across all classes without overfitting.
Open Access: Yes
Deep-learning based adaptive fusion of CC and MLO views for improved mammographic cancer diagnosis
Publication Name: Methodsx
Publication Date: 2026-06-01
Volume: 16
Issue: Unknown
Page Range: Unknown
Description:
Breast cancer remains the most prevalent malignancy among women worldwide. The timely detection of this cancer type is critical for improving survival outcomes. Despite advancements, mammogram classification using deep learning strategies still faces challenges. These include inter-view feature inconsistency, loss of diagnostic details, and limited interpretability. In order to address these issues, MammoFusion-Net, a dual-branch deep learning framework, is proposed for mammogram-based breast cancer classification. Using residual convolutional streams, the framework processes craniocaudal (CC) and mediolateral oblique (MLO) views independently. This supports preservation of view-specific anatomical information. In the proposed framework, a Gates Cross-View Fusion mechanism adaptively integrates features across views. As a result of experimental analysis, the proposed framework achieved 92.116 % (VinDr-Mammo dataset) and 95.556 % (INBreast dataset) of improved classification performance.•Employs a dual-branch architecture to independently process CC and MLO views using residual convolutional streams.•Integrates Gated Cross-View Fusion and attention mechanisms adaptively and refines multi-view features for stronger discrimination.•Demonstrates the explainability of the model through Grad-CAM visualizations that highlight lesion-relevant regions.
Open Access: Yes
AI-Driven Stacked Ensemble Intelligence for Robust Link Quality Classification and Adaptive Resource Management in Satellite-Terrestrial Integrated Networks
Publication Name: IEEE Open Journal of the Communications Society
Publication Date: 2026-01-01
Volume: 7
Issue: Unknown
Page Range: 4899-4913
Description:
The Satellite-Terrestrial Integrated Networks (STIN) were emerged as a key architectural pattern for attaining seamless, global and resilient wireless connectivity by adding extensive coverage of satellite systems with the high capacity and low latency of terrestrial networks. In spite of their advantages, STINs face significant challenges arising from heterogeneous link qualities, dynamic network topologies, long propagation delay and highly variable channel conditions which may complicate reliable and adaptable communication. The accurate and timely assessment of link quality is essential to enable effective resource management, adaptive modulation and coding and robust network control in space-ground integrated environments. In this study, AI-based link quality classification model for STINs based on a stacked ensemble learning architecture. This model combines multiple lightweight machine learning classifiers and a meta-level learner to capture complex non-linear relationships among satellite orbital parameters, spatial characteristics, and link dynamics. The framework categorizes satellite-terrestrial links into three operational states as Good, Moderate and Poor, which provides actionable intelligence for cross-layer resource allocation and adaptive communication strategies. The extensive experimental evaluation demonstrates that the proposed work attains 96.79% classification accuracy and 96.88% macro averaged F1-score where Decision tree as 85% and Machine Learning Based attain 88%. This indicates highly balanced and robust performance across all link classes. The confusion matrix analysis reveals that the misclassification occurs only between adjacent link quality states with no critical misclassifications between good and poor links. This ensures high reliability for operational decision-making. When compared to single-model baselines, the proposed approach increases prediction stability and robustness under heterogeneous and dynamic STIN conditions. The results confirmed the machine learning-assisted link quality intelligence can serve as a practical and efficient enabler for dynamic resource management in STIN. The model is computationally effective, scalable and readily deployable within next-generation STIN control planes which supports reliable communication for broadband access, emergency services and IoT applications.
Open Access: Yes
Hybrid NLP-based speech augmentation with explainable AI approach for enhancing reliability and explainability in Human-Robot Interaction
Publication Name: ICT Express
Publication Date: 2026-01-01
Volume: Unknown
Issue: Unknown
Page Range: Unknown
Description:
Ensuring task safety in Human-Robot Interaction (HRI) environments is a critical requirement for reliable and trustworthy robotic systems. AI can be used effectively to estimate robot task safety. However, existing systems suffer from limited data availability and class imbalance, resulting in inaccurate detection of unsafe events. To address these issues, a hybrid speech data augmentation approach is proposed, which combines acoustic and linguistic approaches to train the ML models effectively. The experimentation involves implementing the hybrid augmentation approach, with acoustic transformations for features such as audio level and linguistic transformations for speech data. Results indicate that various Machine Learning models show enhanced performance, achieving up to 0.97 accuracy with the hybrid approach, while the other augmentation approaches achieve lower results, with accuracy ranging from 0.66 to 0.92. In addition, Explainable AI (XAI) strategies are employed to highlight key contributions of significant characteristics such as speech data, audio level, and robot position.
Open Access: Yes
