		FeedBack \| Contact Us \| Links \| Site Map

Welcome to Journal of Theoretical and Applied Information Technology

Submit Paper / Call for Papers

Journal receives papers in continuous flow and we will consider articles from a wide range of Information Technology disciplines encompassing the most basic research to the most innovative technologies. Please submit your papers electronically to our submission system at http://jatit.org/submit_paper.php in an MSWord, Pdf or compatible format so that they may be evaluated for publication in the upcoming issue. This journal uses a blinded review process; please remember to include all your personal identifiable information in the manuscript before submitting it for review, we will edit the necessary information at our side. Submissions to JATIT should be full research / review papers (properly indicated below main title).

Journal of Theoretical and Applied Information Technology
July 2026 | Vol. 104 No.14

Title:	ENHANCED MULTILINGUAL TEXT CLASSIFICATION WITH PRE-TRAINED LANGUAGE MODELS
Author:	KANDULA NARASIMHARAO, ANGARA S. V. JAYASRI
Abstract:	Introduction: Multilingual text classification is an important task in natural language processing (NLP), especially in the context of sentiment analysis in languages with different grammatical, semantic, and lexical properties. Labelled data sets are not available for many languages, and hence it is difficult to predict the sentiment. The use of pre-defined language models and translation-based methods is a promising solution to cross-lingual generalization. Objectives: This research has the following research objectives: (1) multilingual sentiment classification, (2) neural machine translation to translate foreign language data, (3) neural machine translation using transformer models and large language models (LLMs) combined with ensemble learning, and (4) multilingual model performance evaluation. Methods: Four languages- Arabic, Chinese, French, and Italian are translated into English using neural machine translation tools: DeepL Translator and Google Translate. The translated text was then fed into an ensemble of 3 pre-trained models: Twitter-RoBERTa-Base-Sentiment-Latest, BERT-Base-Multilingual-Uncased-Sentiment and GPT-4 from OpenAI. The sentiment predictions were aggregated output in a weighted voting mechanism, which was based on performance metrics, such as accuracy, precision, recall, and F1-score. Results: The ensemble model suggested here obtained an overall accuracy of greater than 96%, which was improved when compared to individual models. The ensemble learning with translation enhanced the understanding of context, robust across languages and showed a consistent level of performance in multilingual sentiment classification. Conclusion: The obtained results support the fact that translation-based ensemble learning using pretrained transformers and LLMs is a scalable and accurate multilingual sentiment analysis approach. This paper shows how neural translation and the best pre-trained models can overcome language barriers in NLP.
Keywords:	Multilingual Sentiment Analysis; Pre-Trained Language Models; Ensemble Learning; Machine Translation; Large Language Models (LLMs).
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	A QUANTUM ENHANCED DEEP REINFORCEMENT LEARNING FRAMEWORK FOR INTELLIGENT SMART GRID ENERGY OPTIMIZATION UNDER DYNAMIC LOAD CONDITIONS
Author:	K. SAI MADHURI, M. V. RAMESH, RAVI SANKAR SEEMAKURTI, VIJAYAKUMAR SANGAMESVARAPPA, P. KIRAN KUMAR, CHIRANJEEVI PHANEENDRA, BANDARU SATISH BABU, RADHIKA PEERIGA
Abstract:	The interconnection of renewable energy sources and variable electricity consumption has introduced new challenges to smart grid energy management, including energy imbalances, high operational expenses, and grid instability. Although deep reinforcement learning and quantum-inspired optimization techniques have demonstrated applicability in smart grid applications, very few studies have explored their integration within a unified framework for real-time energy scheduling and resource allocation under dynamic operating conditions. To address this challenge, this work introduces a novel Quantum-Enhanced Deep Reinforcement Learning (QEDRL) framework that integrates Long Short-Term Memory (LSTM)-based load forecasting, quantum-inspired probabilistic state encoding, and Deep Q-Network (DQN) reinforcement learning to optimize energy consumption in smart grids. The proposed framework was evaluated using a hybrid smart grid dataset comprising real-time load demand, renewable energy generation, battery storage status, and electricity pricing information. Performance assessment was conducted using metrics including optimization efficiency, forecasting accuracy, renewable energy utilization, convergence speed, and grid stability. Experimental results showed that the proposed QEDRL framework achieved 96.8% optimization efficiency, 97.1% load forecasting accuracy, 31.4% energy cost reduction, 27.2% transmission power-loss reduction, and 93.7% renewable energy utilization, outperforming conventional Energy Management Systems (EMS), Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and standard DQN approaches. The quantum-inspired state representation significantly improved convergence speed and exploration efficiency in a high-dimensional smart grid environment. The study makes original contributions to the knowledge base by demonstrating how quantum-inspired optimization can be integrated with deep reinforcement learning to enhance energy efficiency, renewable energy utilization, grid stability, and adaptive decision-making. The findings provide valuable insights for the design of future intelligent, sustainable, reliable, and adaptive smart grid energy management systems.
Keywords:	Smart Grid Optimization; Deep Reinforcement Learning; Quantum-Inspired Computing; LSTM Load Forecasting; Renewable Energy Management; Energy Efficiency
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	AN ADVANCED EXPLAINABLE MULTI-TASK DEBERTA-V3 FRAMEWORK FOR AUTOMATED CONTEXT AWARE SCORING OF ENGLISH WRITING PROFICIENCY
Author:	MYAGMARSUREN OROSOO, DOLJINSUREN BATBAYAR, DR. AMIT KHAPEKAR, V V M J SATISH CHEMBULY, DR. B. SYAM SUNDAR BHAGAVAN, DR. VUDA SREENIVASA RAO, DR.M.SHYAMALA BHARATHY
Abstract:	The increasing interest in automated writing evaluation that requires reliability has led to the development of intelligent systems that can be used to assess various linguistic proficiencies. Current methods of transformers can usually be unsatisfactory in terms of grammar structure, richness of vocabulary, pronunciation clues, and general linguistic integrity in a single framework. To surmount these shortcomings, an Advanced Explainable Multi-Task DeBERTa-v3 Framework is proposed, which is formulated using a multi-stage architecture, which is based on contextual embeddings, feature fusion in cross-tasks, and attention-based interpretability. The procedure of the methodology consists of preprocess text systematically, extract DeBERTa-v3 embedding, linguistic quality control, multi-task score prediction, and executed in Python and PyTorch to train and evaluate models. The results of the experiment indicate the presence of high-performance improvements, with the Grammar Error Reduction rate of 87.4%, Vocabulary Improvement Score of 84.9%, Pronunciation Accuracy Score of 88.2%, and Overall Writing Quality Score of 86.7%. These scores are higher than the performance that is usually documented in previous studies where grammar gains are usually less than 80% and vocabulary less than 75% with higher levels of generalization across linguistic levels. The multi-stage fusion strategy is effective in capturing semantic, syntactic and phonological properties that allow more analysis of writings by the learner and gives better interpretability due to attention mechanisms. Comprehensively, this framework provides a scalable, precise, and elucidable solution to advanced language proficiency assessment, which offers a more flexible and all-inclusive alternative to traditional ESL assessment practices.
Keywords:	Automated Writing Assessment, Multi-Task Learning, Transformer Architecture, Explainable Artificial Intelligence, Linguistic Proficiency Scoring
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	ENHANCING CLOUD COMPUTING EFFICIENCY WITH HYBRID MACHINE LEARNING ALGORITHMS FOR VM FAILURE DETECTION, RESCHEDULING, AND QOS MANAGEMENT
Author:	B Subramanya Anil Kumar, Dr Basant Sah
Abstract:	This paper introduces an innovative framework aimed at optimizing cloud computing environments by incorporating Virtual Machine (VM) Failure Detection, Dynamic Task Rescheduling, and Quality of Service (QoS)-Based Performance Measurement. The proposed framework applies a combination of supervised machine learning models, including Linear Regression, KNN, SVR, boost, Light, and Random Forest, to predict key performance metrics such as CPU usage, memory usage, and network traffic. The integration of these models, particularly the combined Light and Random Forest, resulted in the most accurate predictions, achieving the lowest RMSE values across all metrics. The dataset used for the study contained essential cloud metrics like CPU and memory usage, network traffic, energy consumption, and task execution details. Through this framework, hybrid machine learning techniques are leveraged to detect VM failures in real-time while dynamically rescheduling tasks, optimizing resource distribution, and minimizing system downtime. The results demonstrate a marked improvement in cloud performance, with enhanced energy efficiency and reduced execution time. The experimental outcomes also showed significant improvements in failure detection accuracy and QoS metrics, such as reliability and security, making the framework highly effective for managing modern cloud systems. This work sets a new benchmark for cloud service performance by combining machine learning with AI-driven optimization techniques.
Keywords:	Hybrid Machine Learning, VM Failure Detection, Dynamic Task Rescheduling, Quality of Service (QoS), Cloud Computing.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	STAD-MLTRS: A SMART THREAT-AWARE DETECTION MODEL WITH MULTI-LAYER TRUST AND ROUTING SECURITY FOR IOT-ENABLED SMART CITIES
Author:	GUNDALA VENKATA RAMA LAKSHMI, Dr. R. DEEPTHA
Abstract:	The Rapid expansion of IoT infrastructure as the foundation of smart cities has intensified cybersecurity risks, with modern attack vectors including phishing-triggered payloads, injection exploits, and multi-stage DDoS campaigns. To counter these threats, this study proposes STAD-MLTRS (Smart Threat-Aware Detection with Multi-Layer Trust and Routing Security), an integrated framework that enhances security through layered anomaly detection, dynamic trust profiling, and adaptive routing optimization. The framework introduces three key components: a Contextual Email Threat Filter (CETF) for detecting malicious email payloads, Temporal Trust Profiling (TTP) for evaluating node reliability over time, and a Dual-Layer Secure Route Optimizer (DLSRO) that combines trust scoring with LSTM-based anomaly detection to secure communication paths. Experiments conducted on hybrid datasets (IoT-23, CIC-Email 2019, and real-time urban telemetry) demonstrate the framework’s effectiveness, achieving 98.2% detection accuracy, reducing false positives to 1.7%, lowering SLA violations, and improving routing trust and energy efficiency compared to baseline IDS and ML-based approaches. These results confirm that STAD-MLTRS provides a scalable, resilient, and energy-conscious security solution for safeguarding IoT-enabled smart cities against evolving and complex cyberattacks.
Keywords:	IOT Security, Software-Defined Networking, Anomaly Detection, Routing Trust, Smart Cities
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	EMPIRICAL EVIDENCE OF BLOCKING PROBABILITY REDUCTION UNDER FIXED ALTERNATE ROUTING WAVELENGTH ASSIGNMENT IN WDM NETWORKS
Author:	G PRAVEEN BABU, K V RAMANA
Abstract:	Blocking of connections remains one of the key factors that limit the performance in a wavelength-routed wavelength division multiplexing (WDM) network, especially when operators use fixed alternate routes to keep the control-plane simple, provide provisioning stability, and maintain predictable implementation behavior. While various important studies on existing routing and wavelength assignment algorithms highlight adaptive routing, capacity expansion at the physical layer, learning-assisted spectrum control and survivability, there is a lack of isolation of whether blocking probability can be lowered by deterministic WAs alone by fixing the route sets, the topology, the wavelength capacity and the traffic assumptions. This paper tackles this research issue by proposing a deterministic congestion-aware wavelength assignment strategy, namely Fixed Alternate Routing-Based Blocking-Reduction Wavelength Assignment (FAR-WA), for static WDM networks that are not equipped with wavelength conversion, adaptive rerouting and learning-based control. By considering the link-wavelength utilization concentration on precomputed alternate routes, the method selects wavelengths as much as feasible without sacrificing the wavelength continuity so that premature local saturation is avoided. An approach is tested using controlled discrete-event simulation to see how it behaves in the presence of various traffic loads, for different alternate routes, and under different base wavelength-assignment strategies (First-Fit, Random-Fit, Least-Used, and Most-Used) in various configurations of WDM meshes. The results indicate that under all the different load levels considered, the proposed method consistently reduces the blocking probability compared to First-Fit e.g., 0.298 to 0.172 at 420 Erlangs and 0.072 to 0.041 at 120 Erlangs. The approach also boosts connection acceptance, decreases congestion concentration and ensures a deterministic execution overhead, which is acceptable in statically provisioned environments. It is noticed that such findings show that in fixed WDM routing, blocking reductions can be obtained without having to resort other than to disciplined wavelength assignment, that is, without assuming network adaptivity to cope with the blocking constraints, and even without requiring additional wavelengths.
Keywords:	Fixed Alternate Routing, Wavelength Assignment, Blocking Probability Analysis, Static WDM Networks, Optical Network Simulation
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	PSYCHOLOGICAL HEALTH ASSESSMENT MODEL USING WEIGHTED COUPLED MULTILEVEL VARIABLE SET WITH T5 TRANSFORMER
Author:	M PRAVARSHA REDDY, LAKSHMI RAMANI BURRA
Abstract:	The system of psychological health assessment and problem identification primarily consists of multivariate, multilevel variables with strong coupling, and is characterized by complex nonlinear interactions among different factors. The field of psychological quality and the methodologies used to assess it has seen substantial growth with its widespread application. Unfortunately, the conclusions of the present evaluation methods, which rely heavily on basic statistical analysis, are inaccurate. College and university student affairs departments must so actively seek out and address adults' mental health issues as soon as they manifest. With the advancements in intelligent technology, it is now feasible to assess and forecast adults' mental states in real time using multimodal data and deep learning models. The use of transformer-based language models in healthcare contexts is still in its infancy, nonetheless, despite their fast advancement. Text-to-Text Transfer Transformer (T5) model is used in this research, which makes use of an encoder-decoder architecture. The decoder produces the output text after processing the input text by the encoder. The encoder's job is to take in the syntactic and semantic information from the input text and store it in a contextualized representation. This research proposes a Psychological Health Assessment Model using Weighted Coupled Multilevel Variable Set with T5 Transformer (PHA-WCMVS-T5) for accurate health assessment. The proposed model when compared with the traditional models performs better in psychological health assessment.
Keywords:	Psychological Health Assessment, Multilevel Variables, Psychological Quality, Adult’s Mental Health, Multimodal Data, Deep Learning, Text-to-Text Transfer Transformer.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	CASUAL GRAPH-STRUCTURED MULTIMODAL LEARNING FOR CARDIAC ABNORMALITY DETECTION USING ECG AND CLINICAL DATA
Author:	MANDAKINI INGLR, DR. RATNESH LITORIYA
Abstract:	Arrhythmias and coronary artery disease can be detected in their earliest stages only if the models are accurate and clinically interpretable. In the case of causal, graph-structured multimodal learning, this is achieved by combining raw ECG time series data with structured clinical features. Deep Temporal Convolutional Encoder encodes ECG signals while structured clinical features are encoded within the physiologically motivated causal graph connecting patient demographics, lab features, and cardiovascular risk factors. Causal Graph Neural Network (Causal-GNN) is capable of learning embeddings both from interaction between leads in ECG and causal pathways in the clinical domain, thanks to counterfactual training that helps to ensure an invariant predictive mechanism and reduce biases introduced by the specific dataset. This model performs both arrhythmia classification and coronary artery disease risk prediction as well as provides counterfactual explanations of how the changes in certain clinical features will affect the diagnosis. Analyses of results on the MIT-BIH Arrhythmia and Z-Alizadeh Sani databases have proven the accuracy, robustness to the distribution shift, and increased interpretability compared to state-of-the-art deep neural network and non-causal fusion baselines.
Keywords:	Causal Graph Neural Network, Multimodal Learning, ECG Signal Analysis, Clinical Data Fusion, Counterfactual Reasoning, Cardiac Abnormality Detection, Arrhythmia Classification.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	A SYSTEMATIC REVIEW OF TEMPORAL SEQUENCE MODELING USING RECURRENT NEURAL NETWORKS FOR REAL-TIME HEALTH MONITORING ARCHITECTURES AND PREDICTIVE ANALYTICS
Author:	SHIVARAM REDDY K, M V NARAYANA
Abstract:	Leveraging state-of-the-art deep learning models with real-time health monitoring systems is an essential direction to this field of digital health analytics. Here, Recurrent Neural Networks (RNNs), for example, Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRU), have significant potential in accurately learning physiological time-series data for predictive diagnostics. In this systematic review, we critically analyze the current state of modelling temporal sequences using RNNs with real-time health monitoring systems. In accordance with PRISMA guidelines, we analyzed 124 peer-reviewed articles from January 2014 to December 2024, across major databases, and curated parameters relating to modelling properties, deployment strategies (i.e., cloud, edge, mobile), data acquisition pipelines, and health signal modalities (i.e., ECG, PPG, respiration). We compared the performance of vanilla RNNs against the sophisticated temporal models for value-add and sensor signal processing (e.g. detecting arrhythmias, predicting blood glucose, estimating early warning scores). There was significant heterogeneity in metrics of evaluation, and varying methodologies for addressing data noise and latency, despite minimum real-world deployment of models also being suggested. Also emerging was a fundamental research gap in scalable RNN implementations intended for energy-efficient edge-based processing. Very few studies adopted any adaptive temporal attention, while generalization across patients posed challenges to clinical uptake. Subsequently, little test activity surfaced concerning model interpretability, computational latency, or privacy-preserving inference. Representing a clear gap in current literature as well, our review provides a taxonomy of RNN-based real-time health monitoring systems, notwithstanding some serious limitations for practical deployments, as well as benchmarking modelling trends, and promotion of a common research agenda path with a strong focus for ramification from hybrid architectures, federated learning for model integrations, and explainability. Our review endeavor provides a synthesis of technical intelligence and system-level constraints to also serve as a basis for scholars and professionals interested in aligning to advance the state-of-the-art in their intelligent health monitoring systems.
Keywords:	Recurrent Neural Networks, Temporal Sequence Modeling, Real-Time Health Monitoring, Predictive Analytics, Systematic Review
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	DYNAMIC SOCIAL SKI-DRIVER OPTIMIZATION WITH DEEP CAPSULENET ARCHITECTURE FOR THE DETECTION AND CLASSIFICATION OF EMOTION USING FACIAL FEATURES
Author:	S.Srinivas, Dr. Mercy Paul Selvan
Abstract:	Facial expression is the universal language for humans to express their feelings and detect emotions via the computer vision approach and is a challenging process. Emotion detection by the machine is arduous with the presence of accessories, pose variation, non-uniform illumination, etc. Several works have been carried out in this field, however, the mutual optimization of feature extraction and detection of emotion has not been achieved. To tackle this issue, we propose an innovative technique known as optimized deep learning. To begin the process, the data that are collected are pre-processed, and the facial features associated with the Dynamic social ski-driver optimization (DSSO) model are incorporated with Deep CapsuleNet approach. The derived features are used to detect the emotions using the proposed technique. Simulations are effectuated in the MATLAB simulator and analyzed the robustness of the proposed work. The proposed approach is compared with the previous works using the statistical parameters and achieved an increased accuracy of around 98%.
Keywords:	Facial Features, Characterization, Emotion Detection, Dynamic Social Ski-Driver Optimization, And Deep Learning.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	FEDERATED LEARNING FOR STRESS DETECTION USING DISTRIBUTED WEARABLE DATA IN PRIVACY PRESERVING HEALTHCARE SYSTEMS
Author:	CHEEMALADINNE VENGAIAH, ERUKALA MAHENDER, P. PRATIMA RANI, A. SWATHI, YAMINI DEVI YKUNTAM, RAMANJANEYULU DAYINABOYINA, RIYAZ MOHAMMAD, K. MYTHRI SRIDEVI
Abstract:	Nowadays, comprehensive, continuous, and reliable monitoring is required in healthcare systems to account for stress-related impacts not only on physical but also on mental health, such as cardiovascular diseases, anxiety, and impaired cognitive function. This research aims to propose a federated learning approach for stress detection while considering the data privacy of the sensor network. FL-SDNet is proposed as a fusion approach combining a CNN and a BiLSTM for feature extraction, integrating spatial and temporal features from multi-modal physiological signals, including heart rate (HR), electrodermal activity (EDA), temperature (TEMP), and motion (MOTION). The proposed system uses a cross-modal attention fusion mechanism that dynamically balances multiple attention signals to boost performance in scenarios with diverse data distributions and introduces an adaptive federated aggregation (AFA) strategy that considers client reliability and data quality. It is fully decentralized and security-preserving, with raw data stored on local devices and only model updates sent and received in encrypted form. The results of extensive experiments with non-IID federated data show that the proposed model achieves 90.2% accuracy, 0.88 F1 score, and 0.92 ROC-AUC in a non-IID federated setting, results comparable to those of the centralized deep learning model. Moreover, it can operate in resource-limited applications (such as wearables) by reducing communication overhead by approximately 40%. A comparative study is conducted between traditional machine learning and deep learning using standard federated learning methods, indicating that the proposed method is effective. An ablation study is conducted to verify the effectiveness of the new component added to the experiment. Overall, the results of this study show that federated learning can achieve high accuracy for stress detection while ensuring data privacy, supporting the potential to exclude local data from global collection while maintaining high accuracy in real-world healthcare applications. This work serves as a valuable asset for the development of next-generation personalized real-time stress-monitoring systems for 'privacy-preserving' healthcare, remote patient monitoring, and intelligent wearables.
Keywords:	Federated Learning, Stress Detection, Wearable Sensors, Privacy-Preserving Healthcare, Multimodal Deep Learning, Non-IID Data
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	QUANTUM INSPIRED ROUTE OPTIMIZATION FOR EMERGENCY VEHICLE MOVEMENT DURING URBAN FLOODING
Author:	RAVIKUMAR GOTTHI, K. RAMESH BABU, B. RANGA SWAMY, SWATHI YALAVARTHY, SATTI HARICHANDRA PRASAD, O. RAMA DEVI, VIJAYA GOPAL NALLAGORLA, K. NARAYANA RAO
Abstract:	Urban flooding has a significant impact on the flow of vehicles and the movement of emergency vehicles, leading to more time for rescues, risk of vehicles, and human lives lost. In this research, a new Quantum-Inspired Route Optimization (QIRO) intelligent emergency vehicle navigation framework was introduced for urban flooding scenarios. Current routing methods are mainly static and are not able to adjust to the fast-changing flood conditions, traffic congestion, and route uncertainty. These constraints may lead to longer response times and less transportation security during a disaster. The main research goal was to design an adaptive and uncertainty-aware routing model that always minimizes the delay of the emergency response, the risk of exposure to flood, the impact of congestion, and the consumption of fuel. The suggested framework combined real-time flood severity estimation, traffic congestion analysis, prediction of road accessibility in the probabilistic way, and quantum-inspired multi-objective optimization and merged them into a single intelligent transportation system framework. The simulated smart city transportation environment was used for experiments, which included 500 intersections, 1200 road segments, dynamic flood propagation and emergency vehicle trajectories. Proposed QIRO framework was compared with Dijkstra, A, Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO) methods. The experimental results showed that the proposed model achieved an emergency response time of 8.4 minutes, route reliability of 94.3%, flood exposure risk reduction of 17.2%, and road accessibility prediction accuracy of 95.2%, outperforming conventional methods. Under uncertain flooding conditions, the probabilistic quantum-inspired optimisation mechanism greatly enhanced adaptive routing capability and convergence efficiency. The results show that combining QIO with timely flood information can greatly enhance the efficiency, reliability and resilience of emergency transportation during disasters in urban areas where flooding poses risks. This study provides a feasible direction for future intelligent disaster response and smart city transportation systems.
Keywords:	Quantum-inspired optimization, Emergency vehicle routing, Urban flooding, Intelligent transportation systems, Flood-aware navigation, Smart city disaster management
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	DUAL-WAY F2M3SLNN: MRI AND PET MULTI-MODALITY FUSION BASED BRAIN TUMOR DIAGNOSIS USING PALE STRUCTURED TRANSFORMER AND SKIP CONNECTED LSTM
Author:	SANTHOSH KUMAR. S, DR. S.P.SASIREKHA2, DR.R.SANTHOSH
Abstract:	A brain tumor signifies an irregular proliferation of cells that replicate uncontrollably. Within a medical diagnostic system, the precise identification of both the location and size assumes a critical role in brain tumor diagnosis. The primary diagnostic modalities for this purpose encompass Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET), both being extensively employed techniques. Identifying brain tumors in PET and MRI images presents a formidable challenge, primarily stemming from the low sensitivity of boundary pixels. To achieve precise and reliable brain tumor detection, the recent studies introduces an efficient fusion-based approach in its methodology however it declines to provide accurate results. To alleviate these issues, we have proposed multi-modality based brain tumor classification named as Dual-way F2M3 SLNN framework. Initially, for utilizing both MRI and PET information we have performed Multi-modality image fusion by proposing TangleNet via Average Norm Technique (ANT). After that, we have enhanced fused image quality by executing pre-processing in terms of noise reduction and bias field correction using Quaternion Quasi-Chebyshev with Non-Local Means (QQC-NLM) and Expectation Maximization respectively. Besides, the pre-processed image quality is ensured by performing image quality evaluation using Kernel Support Vector Machine (K-SVM). Following that, the appropriate features are extracted by Pale structure transformer and feature dimension are reduced through employing Analogous Attention Mechanism. Finally, the accurate brain tumor is classified using Skip connected Long Short-Term Memory (SC-LSTM). The proposed Dual-way F2M3 SLNN framework is conducted on MATLAB for improving model performance, additionally the performance of proposed model is evaluated by determining several performance metrics in terms of accuracy, sensitivity, specificity, F1-Score and AUC where our proposed Dual-way F2M3 SLNN framework achieves superior achievement than other existing works.
Keywords:	Multi-modality, Image Fusion, Brain Tumor Classification, MRI & PET, Pale Structured Transformer, Analogous Attention Mechanism.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	QUANTUM DEEP LEARNING FRAMEWORK FOR EARLY SKIN CANCER CLASSIFICATION USING DERMOSCOPIC IMAGES
Author:	KONDA MANOJ KUMAR, B. RANGA SWAMY, CH. NIRANJAN KUMAR, KONALA PADMAVATHI, S. MANOJ, C. SAI KALYANA DEEPTHI, JOHN T MESIA DHAS, SRILAKSHMI RAMYA SAKAMUDI
Abstract:	It is crucial that skin cancer be identified at an early stage and accurately diagnosed to decrease the mortality and enhance the clinical outcomes of treatment for skin cancer. However, standard deep learning models have limitations in handling dermoscopic image analysis due to feature redundancy, high computational complexity, and inadequate representation of the nonlinear lesions. In order to overcome these drawbacks, the present research introduced a new Quantum Attention-driven Deep Learning Framework (QADLF) system for dermoscopic images-based early skin cancer classification. The proposed framework combined advanced preprocessing, attention-guided segmentation of the lesion, quantum feature encoding, Hybrid Quantum Convolutional Neural Network (Hybrid Q-CNN) and adaptive feature fusion for comprehensive lesion analysis and classification. Two dermoscopic databases, namely ISIC 2019 and HAM10000, were used for experiments. The attention-based Quantum U-Net segmentation successfully localized regions of lesions, and variational quantum circuits improved the representation and optimization of the nonlinear features. The results of the experiments showed that the proposed QADLF framework yielded better classification accuracy with an accuracy score of 98.7%, precision of 98.2%, recall of 98.5%, F1-score of 98.3% and AUC of 99.1% than conventional CNN, ResNet50, DenseNet121, EfficientNet and Vision Transformer models. The convergence stability, classification robustness and lesion localization were considerably improved by the integration of quantum-enhanced learning and feature extraction guided by attention, under different dermoscopic imaging conditions. The proposed framework can be part of the development of intelligent, reliable, and computer-aided dermatological diagnostic systems to assist early melanoma detection and for better clinical decision-making processes.
Keywords:	Skin Cancer Classification, Dermoscopic Images, Quantum Deep Learning, Hybrid Q-CNN, Attention-Guided Segmentation, Melanoma Detection
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	ATTENTION-BASED LSTM MODEL FOR PREDICTING RECURRENCE OF FEBRILE SEIZURES IN CHILDREN
Author:	SUNEET GUPTA, R. SUKRUTA, P. ROHINI, GUNDEPOGU VENUGOPALARAO, CH. V. KIRANMAYI, T. V. SAI KRISHNA, N. V. S. SOWJANYA, LOYA CHANDRAJIT YADAV
Abstract:	Febrile seizures are one of the main neurological problems of the first years of life, and the risk of recurrence constitutes a difficult problem for timely clinical intervention. The goal of this study is to develop a predictive model that reliably classifies children at high risk of febrile seizure recurrence using their clinical information over time. To this end, a novel Dual-Attention Temporal Long Short-Term Memory (DAT-LSTM) network is presented that extends temporal sequence modelling to capture temporal sequence patterns and learn feature importance simultaneously by incorporating a dual-level attention mechanism. The model was tested on a multivariate dataset of children and compared with conventional statistical methods and machine-learning methods. Experimental results show that the proposed model performs better than the baseline models with an accuracy of ‘90%' and an AUC of 0.93. The attention mechanism also improves interpretability by selecting key clinical data, such as the early age of onset, with the pattern of fever progression. Most results highlight the effectiveness of deep learning and attention in improving predictive performance and clinical understanding. The proposed framework could improve early risk stratification, prevent problems associated with event recurrence, and support the development of intelligent clinical decision support systems in pediatric care.
Keywords:	Febrile Seizures, Recurrence Prediction, LSTM, Attention Mechanism, Pediatric Healthcare, Deep Learning
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	EXPLAINABLE QUANTUM AI FOR CRITICAL INFRASTRUCTURE FAILURE PREDICTION
Author:	M. SAILAJA, CH. NIRANJAN KUMAR, P.S.G. ARUNA SRI, LAKSHMI PANUGANTI, C. RAGHAVENDRA, GEETA KAKARLA, RAVI SANKARA SEEMAKURTI, JOHN T MESIA DHAS
Abstract:	In recent years, critical infrastructure systems—from smart grids to industrial automation platforms to cyber-physical environments—have become increasingly susceptible to failures, cyber attacks, and equipment deterioration. Many Machine Learning (ML) and Deep Learning (DL) models achieve high predictive accuracy but are difficult to interpret and require significant computational resources to deploy in safety-critical applications. The study introduced a new Explainable Quantum Artificial Intelligence (XQAI) framework for predicting failures in critical infrastructures, combining Quantum Adaptive Feature Embedding (QAFE), Variational Quantum Circuits (VQC), Bidirectional LSTM temporal learning, Quantum Attention mechanisms, and SHAP-based explainability analysis. The proposed framework leveraged heterogeneous data collected from Industrial Internet of Things (IIoT) sensors and SCADA systems to detect complex nonlinear patterns in infrastructure failures. The proposed XQAI framework was evaluated experimentally and demonstrated a very high accuracy of 98.9%, precision of 98.5%, recall of 98.1%, and F1-score of 98.3%, which are significantly higher than those of conventional machine learning, deep learning, and quantum learning models. The explainability module accurately identified the parameters that influenced infrastructure failure, particularly temperature, vibration, voltage instability, and communication latency. The results showed that integrating Explainable Artificial Intelligence (XAI) with quantum-augmented temporal learning had a remarkable impact on prediction reliability, interpretability, and computational efficiency. The proposed framework offers an intelligent infrastructure-monitoring solution that is scalable and transparent for the next-generation safety-critical industrial environment.
Keywords:	Explainable Quantum AI, Critical Infrastructure, Failure Prediction, Quantum Machine Learning, Predictive Maintenance, Industrial IoT
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	DEEP LEARNING-BASED REAL-TIME ANOMALY DETECTION IN INDUSTRIAL IOT NETWORKS: MITIGATING BOTNET ATTACKS WITH EXPLAINABLE AI
Author:	A.N. GNANAJEEVAN, G. INDUMATHI, SWATHI AGARWAL, GARA JAYA RAJU, N. SRIKANTH REDDY, J DAPHNEY JOANN, HARIKA B, ZHU HAIJING
Abstract:	The advent of the Industrial Internet of Things, more advanced botnet attacks of smart factories and other systems have emerged. There is a need to have an offline and an unexplainable solution for the real-time, explainable and secure anomaly detection in resource constrained IIoT systems, which can be accomplished using existing IDS solutions. This model applies to a novel deep learning framework for real-time botnet mitigation in IIoT networks based on Attention-based Hybrid Models and XAI, with limited scope. A novel deep learning method that includes an attention-based CNN/Transformer hybrid network for spatial-temporal feature extraction, multi-head self-attention bidirectional gated recurrent units, a conditional generative adversarial network for balancing the classes, LSTM denoising autoencoders for feature dimensionality reduction and adversarial training using Auto-PGD and Square Attack. SHAP-based explainable AI delivers global and local explainable results. This approach on edge devices like Raspberry Pi 4 and Jetson Nano with a Flask dashboard achieves 99.96% accuracy and 99.94% recall on IoT-23 and Edge-IIoTset datasets more than 99% accuracy for adversarial attacks and inference time of 32.1 ms. The real-time XAI integrated pipeline is a major achievement, while its implications span from better industrial safety and regulatory adherence to greater trust in data-driven decision-making processes to tackling the challenge of explaining complex systems. The result of this research indicates that a lightweight, explainable and robust model was developed, which is an advancement of the state of the art in real-time botnet detection for IIoT networks that can benefit safe industrial practices and regulatory compliance.
Keywords:	SHapley Additive exPlanations, Process Innovation, Industrial Internet of Things, CNN-Transformer Hybrid, Edge Computing, Adversarial Robustness.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	AN INTELLIGENT ADAPTIVE SYSTEM FOR ANEMIA CLASSIFICATION FROM HEMATOLOGICAL PARAMETERS
Author:	S.B. SWATHI, DR. N. RAMANA
Abstract:	Anemia is among the most common blood disorders worldwide, caused by a decrease in red blood cells or hemoglobin concentration, which reduces oxygen transport in the body. Traditional methods of diagnosing anemia usually involve manual interpretation of diagnostic hematological parameters, and these time-consuming processes can still be subject to human error. To mitigate some of these challenges, this study presents a Sharpbelly Fish-based Deep Neural Network Prediction Framework (SbDNNPF) that can classify anemic cases using publicly available hematological data. The SbDNNPF process begins with data pre-processing to eliminate inconsistencies and missing values, followed by feature selection to identify relevant hematological parameters. An optimized feature subset is then used to train a DNN, and the Sharpbelly Fish Optimization algorithm is used to tune network parameters to improve convergence and minimize prediction error. Experimental analysis of the SbDNNPF demonstrates that the proposed DNN-based approach outperforms traditional machine learning (ML) classifiers in terms of performance, accuracy, precision, recall, and F score, thereby effectively enhancing the reliability of anemia diagnosis.
Keywords:	Anemia classification, Hematological Data, Pre-processing, Diagnosis, Feature selection. Model.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	AN INTELLIGENT INTRUSION DETECTION SYSTEM BASED ON GIFO-PLSTM MECHANISMS FOR ENHANCING MANET SECURITY
Author:	MAREESWARI G, VENKATESH K, SURESH THANGAKRISHNAN M, GOMATHY NAYAGAM M
Abstract:	Ensuring robust security in Mobile Ad Hoc Networks (MANETs) remains a significant challenge due to their decentralized architecture, dynamic topology, and node mobility. Intrusion Detection Systems (IDS) are vital in safeguarding MANETs by identifying unauthorized or malicious activities that compromise network integrity. Existing approaches leveraging machine learning and deep learning have shown promise but often suffer from high algorithmic complexity, suboptimal network performance, and notable misclassification rates. This study introduces a novel IDS framework aimed at strengthening MANET security through a deep learning-driven model. The framework follows a structured pipeline involving data preprocessing, feature extraction, feature optimization, and attack classification. Two benchmark datasets NSL-KDD and CICIDS-2017 are used to evaluate the proposed method. Missing values in the datasets are addressed during normalization to minimize classification errors. Principal Component Analysis (PCA) is applied to derive informative feature vectors. A hybrid Greedy Improved Firefly Optimization (GIFO) algorithm is designed to select the most relevant features, combining the strengths of greedy search and firefly optimization for enhanced accuracy. The selected features are then classified using a Probabilistic Long Short-Term Memory (PLSTM) model, which effectively detects attack patterns. The system’s performance is thoroughly assessed using standard evaluation metrics, demonstrating superior detection capability and reduced false classification rates compared to traditional methods.
Keywords:	Mobile Ad Hoc Networks (MANET), Network Security, Intrusion Detection, Attack Detection, Principal Component Analysis (PCA), Hybrid Firefly-Greedy Feature Selection (GIFO), Probabilistic LSTM (PLSTM) Classifier.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	HYBRID DEEP LEARNING MODEL FOR TUNNEL CRACK SEGMENTATION AND RISK CLASSIFICATION
Author:	DIVYA LINGINENI, T. V. SAI KRISHNA, B. RANGA SWAMY, BOLLAM SRIVANI, LAKSHMI PANUGANTI, VITHYA GANESAN, JOHN T MESIA DHAS, M. KAVITHA
Abstract:	The ability to monitor the structural condition of tunnels is critical to keeping roads and bridges safe and secure, and preventing failures due to cracking and decay. The traditional manual inspection is time-consuming, labor-intensive, and subjective, which is difficult to operate under complex underground environment. A Hybrid Deep Learning Model for Tunnel Crack Segmentation and Risk Classification was proposed, consisting of an Attention-Guided Multi-Scale U-Net (AGMS-U-Net) with multi-scale convolution, spatial-channel attention, and residual decoder blocks for crack segmentation, together with a Hybrid CNN-LSTM Risk Analyzer for structural severity prediction. The proposed framework adopted multi-scale Convolution, Spatial-Channel attention mechanism and residual Decoder block to enhance the thin crack detection and crack segmentation robustness in noisy tunnel environment. The CNN-LSTM classifier was used to extract and analyze crack features of the structural cracks, such as crack length, crack width, crack density, crack orientation, and crack branching factors, to assess the risk level of the tunnel. The combined dataset of 24,700 tunnel crack images was taken in actual underground environments for experimental evaluation. The proposed AGMS-U-Net has obtained segmentation accuracy of 98.1%, Dice coefficient of 97.4%, and IoU of 95.9% while the Hybrid CNN-LSTM classifier has obtained risk classification accuracy of 96.5%, and has shown remarkable improvement in terms of precision and recall over other deep learning models such as U-Net, DeepLabV3+, Mask R-CNN, and CNN-GRU. The experimental outcomes showed the effectiveness of the proposed framework to enhance the localization of cracks, preservation of boundaries and predict the severity of the tunnel under harsh environmental conditions. The proposed intelligent tunnel inspection framework can be helpful for the development of automated structural health monitoring systems, predictive maintenance planning, and smart transportation infrastructure management systems.
Keywords:	Tunnel Crack Segmentation, Structural Health Monitoring, Attention U-Net, CNN-LSTM, Risk Classification, Deep Learning Framework
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	A MODULAR EMOTION-AWARE MULTILINGUAL AI DUBBING FRAMEWORK WITH FER-GUIDED EXPRESSIVE SPEECH SYNTHESIS AND LIP SYNCHRONISATION
Author:	ANANDU M, SHERLY KK, JITHIN MATHEWS, WILLSON JOSEPH C
Abstract:	The development of emotion-aware multilingual dubbing solutions is a new promising research direction in media localization, accessibility, digital entertainment, and cross-lingual communication. Despite the advances in existing dubbing frameworks that concentrate mostly on enhancing transcriptions, translations, speech synthesis, and lip sync, there exists little work on capturing the emotional intent of the speaker during the entire process of dubbing. This may lead to the generation of linguistically correct but unnatural speech in terms of affective aspects from the original video. In order to address this challenge, it is necessary to introduce visual emotions into the speech generation process.In this study, we introduce a modular emotion-aware multilingual AI dubbing system based on Facial Emotion Recognition (FER), Whisper ASR, Neural Machine Translation (NMT), expressive speech synthesis with YourTTS, and lip-sync generation with Wav2Lip. In contrast to other existing methods, which make use of mostly text information, the introduced system directly uses facial emotion predictions to condition expressive speech synthesis, thus allowing us to keep the emotional context intact while keeping linguistic fidelity at a high level in multilingual settings. Another advantage of the proposed approach is its modular nature, which allows for individual optimization of every module independently of the others.The designed FER system uses the EfficientNet-B0 model and is trained with a mixed dataset that is created by combining the FER2013 five-class benchmark with carefully selected target-domain face images for the purpose of robustness against the video environment. Our experimental results on the FER2013 five-class test set give an emotion classification accuracy rate of 75% along with macro-F1 value of 0.68 and weighted-F1 of 0.75 in 20 stable epochs of training, thus providing reliable recognition despite problems such as class imbalance and inter-class similarity. Detected emotions are then used for the generation of emotional multilingual speech.The efficiency of the suggested approach is analyzed through objective and subjective measurements such as performance of FER classification, ASR precision, translation quality measured in BLEU, Mean Opinion Score (MOS), Perceptual Evaluation of Speech Quality (PESQ), and speaker similarity. The results obtained confirm that the explicit emotion-driven approach to speech synthesis could effectively improve expressive qualities of multilingual dubbing without losing flexibility and maintainability that are characteristic of modular system architecture. Thus, the suggested approach could serve as a basis for development of future emotion-driven dubbing systems supporting different languages and emotions.
Keywords:	Emotion-aware dubbing, Facial emotion recognition, EfficientNet-B0, Whisper ASR, Multilingual translation, YourTTS
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	A ROBUST AND EXPLAINABLE HYBRID STACKING–VOTING ENSEMBLE FRAMEWORK FOR MALARIA DIAGNOSIS USING OPTIMIZED FEATURE ENGINEERING
Author:	PRABHAT KUMAR SAHU, SANGAM MALLA, MITRABINDA KHUNTIA, SMITA RATH, SHRABANEE SWAGATIKA, SIPRA SAHOO, DEEPAK KUMAR PATEL, SIDDHANT JAIN
Abstract:	Malaria, caused by parasites of the Plasmodium genus, remains a life-threatening infectious disease requiring timely and accurate diagnosis. Conventional diagnostic approaches, such as peripheral blood smear microscopy and rapid diagnostic tests (RDTs), although widely used, are susceptible to human error and false-negative results in certain endemic regions. To address these limitations, this study proposes a robust hybrid ensemble machine learning framework for malaria risk prediction using structured clinical and hematological data collected from endemic areas. The proposed workflow integrates data preprocessing, feature engineering through standardization, Principal Component Analysis (PCA), and Recursive Feature Elimination (RFE), followed by hyperparameter optimization using Optuna. Multiple classifiers—including Random Forest, Support Vector Machine (SVM), Logistic Regression, K-Nearest Neighbors (KNN), and Gradient Boosting—are combined using stacking and voting strategies to construct a final hybrid ensemble model. Experimental results demonstrate that the proposed hybrid ensemble achieves an overall accuracy of 93%, outperforming individual classifiers and standalone ensemble models. Statistical significance testing confirms that the observed performance improvement is not due to random variation (p < 0.05). Cross-validation and balanced accuracy analysis further validate the robustness and generalization capability of the model under potential class imbalance. An ablation study highlights the contribution of each architectural component, while LIME-based interpretability enhances transparency in clinical decision-making. The proposed framework offers a reliable and explainable machine-learning-driven solution to support early malaria diagnosis and improve healthcare outcomes in resource-limited settings.
Keywords:	Malaria Detection, Ensemble Methods, Classification Models, Logistic Regression, K-Nearest Neighbors, Naive Bayes, Decision Trees
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	GENERATIVE AI FOR EARLY DETECTION OF RETINAL DISEASES USING SYNTHETIC FUNDUS IMAGES
Author:	PONNAM LALITHA, SIVA YENIKEPALLI, K. VISALA, DATLA GANESH, R.Z. INAMUL HUSSAIN, B. RANGA SWAMY, YAMINI DEVI YKUNTAM, VITHYA GANESAN
Abstract:	Early detection of retinal diseases, particularly diabetic retinopathy (DR), is essential for preventing irreversible vision loss; however, the lack of annotated datasets, class imbalance, and privacy concerns remain major challenges for automated diagnosis. This paper proposes a hybrid retinal disease detection approach using synthetic fundus images and three techniques: StyleGAN, a diffusion model, and a multi-scale attention-based convolutional neural network (CNN). StyleGAN generates anatomical structures similar to the human eye, and the diffusion model refines pathological details to produce high-quality images. Augmented images are added to address data scarcity and improve class balance. The improved dataset is then fed into a multi-scale attention-based CNN for accurate ocular disease diagnosis. This proposed framework was tested on the publicly available EyePACS (≈35,000 images) and APTOS 2019 (≈3,662 images) image databases. Approximately 20,000 synthetic fundus images were generated and combined with the original datasets, resulting in an augmented dataset of nearly 55,000 images. The experimental results show that the proposed method achieved an accuracy of 97.8%, a precision of 97.2%, a recall of 96.5%, an F1 score of 96.8%, and an ROC-AUC of 0.991, which are better than those of some other state-of-the-art deep learning models. The ablation study also validates the effectiveness of combining synthetic image generation, diffusion-based refinement, and attention-guided feature learning. The proposed system provides a strong privacy-preserving proposal for automated retinal disease diagnosis and has great potential for clinical use in real-world medicine.
Keywords:	Generative AI, Retinal Disease Detection, Synthetic Fundus Images, Attention-Based CNN, Diffusion Models, Medical Image Analysis
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	DEEP LEARNING ENABLED VISION TRANSFORMER FRAMEWORK FOR WORKER SAFETY MONITORING IN CONSTRUCTION SITES
Author:	Dr. P. PAVAN KUMAR, RAKSHITHA KIRAN P, DANAM NOELLE, V. V. RAMA KRISHNA, C. RAGHAVENDRA, GANGULA SURENDAR, KOMALI GOVINDU, NARESH ALAPATI
Abstract:	Construction sites are dangerous places where employees are often at risk in various ways, including falling objects, unsafe equipment interactions and failure to wear personal protective equipment (PPE) as specified in the regulations. Manual safety monitoring is typically ineffective, unreliable and not able to conduct hazard analysis in real-time. This paper aimed to overcome these drawbacks by introducing a Deep Learning Enabled Vision Transformer Framework to monitor worker safety in construction sites. The suggested solution combined CNN with VC with Adaptive Contextual Risk Assessment (ACRA) to detect PPE compliance, unsafe activities, restricted area intrusion, and hazardous equipment proximity in real time. A hybrid construction safety database with 48,732 annotated images was used to evaluate the proposed approach under various scenarios, such as occlusion of workers, illumination changes, and complex scenes. The results of the experiments have shown the proposed framework to be superior to the CNN, ResNet50, Faster R-CNN, YOLOv5, EfficientNet and Swin Transformer models in terms of accuracy (98.1%), precision (97.3%), recall (96.8%), F1 score (97.0%), and mean Average Precision (mAP) (97.6%). Under complex construction environments, the Vision Transformer (ViT) encoder significantly enhanced its understanding of context hazards and reduced false alarm proportions. The proposed framework was also capable of achieving real-time processing capability with a low detection latency that is suitable for industrial deployment. Incorporating Vision Transformers and adaptive contextual risk assessment can significantly enhance intelligent construction worker safety monitoring and facilitate automated industrial surveillance for accident prevention and workplace safety improvement, the study showed.
Keywords:	Construction Safety Monitoring, Vision Transformer, Deep Learning, PPE Detection, Hazard Detection, Worker Safety Surveillance
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	CAN INTELLIGENT DATA PREPROCESSING METHODS SIGNIFICANTLY ENHANCE INTRUSION DETECTION PERFORMANCE?
Author:	JOSEPHINE. R, ANANDARAJ S. P
Abstract:	In todays world, as cyberattacks increase a lot in both difficulty and frequency, a very strong NIDS has become very important. The CICIDS 2017 dataset is widely used for testing but has major reliability problems. These problems include missing values, duplicates, and outliers. In this context, an important number of studies rely just on simple imputation and simple outlier removal. So, such methods fail to understand the very complex way the dataset works. This is exactly why smart data cleaning in the area of NIDS is still a big gap in the study. To solve this problem, this study suggests a very useful data cleaning method called the Enhanced Preprocessing Approach to Network Intrusion Detection (EPA-NID). This method combines regression imputation, hybrid outlier detection, and normalization. The findings show that useful data cleaning improves both data goodness and model results on the CICIDS 2017 dataset. The suggested EPA-NID method uses different data cleaning methods. First, to fill missing values, it includes regression models (REPTree, Isotonic Regression, and SMOreg) together with the Performance-Weighted Imputer (PWI). It well removes duplicates using a method based on cosine similarity. To find outliers, it uses a mixed method joining the Enhanced Z-score and IQR. Also, it well uses the Tanh Estimator scaling method to adjust the data in the limit of 0 and 1. Using the CICIDS 2017 dataset, EPA-NID a lot improved the model's results. Clearly, models trained on the processed data got an accuracy of 86.67%, a precision of 84.22%, a recall of 83.33%, and an F1-score of 90.90%. In this situation, these outcomes are a lot better to those got using old data cleaning ways. Overall, this study reveals that a better data pre-processing leads to a better intrusion detection performance. These findings can assist researchers in creating more dependable and precise models for cybersecurity.
Keywords:	CICIDS 2017 Dataset, Missing Value Imputation, Network Intrusion Detection Systems (NIDS), Normalization, Outlier Detection
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	HYBRID DEEP LEARNING FRAMEWORK FOR EARLY DETECTION OF BRIDGE CRACK PROPAGATION USING UAV THERMAL IMAGING AND IOT SENSORS
Author:	SHAIK KHALEEL, SRI LAKSHMI CHANDANA, LATHA VATTIPALLY, T. MURALIDHARA RAO, D. VENKATA RAVI KUMAR, VITHYA GANESAN, GEETA KAKARLA, N. SRIJA
Abstract:	It is very crucial to detect the propagation of the bridge cracks in the early stage to prevent the bridge structure failure and to ensure the transportation safety. The traditional approach to bridge examination is time-consuming and labour intensive while also being less effective at finding micro-level defects at an early stage. In this paper, a hybrid deep learning system (HDLS) was proposed to detect early bridge crack propagation by UAV thermal imaging with IoT sensor network. The proposed method included a CNN-based spatial thermal feature extractor, bi-LSTM-based temporal structural analyzer, and attention-based multimodal fusion mechanism to comprehensively analyze thermal crack patterns and real-time structural sensor data. The multimodal dataset extended to 11,300 UAV thermal images and synchronized sequences of IoT sensors was used for experiments. The proposed model achieved the accuracy of 98.4%, precision 97.1%, recall 96.8%, F1 score 98.0% and AUC 98.7% which outperforms the conventional CNN, ResNet50, MobileNet, EfficientNet-B0 and CNN-LSTM model. Experimental analysis showed that combining thermal imaging with IoT sensing was highly beneficial in early-stage crack detection capacity and the number of false positive results was significantly reduced under different environmental conditions. The attention-based fusion mechanism showed excellent improvement in multimodal feature representation and structural anomaly discrimination. The proposed framework offers intelligent, automated, and reliable real-time structural health monitoring and predictive maintenance, thus helping to make bridge structures safer and more sustainable in smart transportation systems.
Keywords:	Bridge Crack Detection; UAV Thermal Imaging; Structural Health Monitoring; IoT Sensors; Hybrid Deep Learning; Attention-Based Multimodal Fusion
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	EFFICIENT EEG-BASED ALZHEIMER’S PROGRESSION PREDICTION USING A CONVOLUTIONAL MULTI-HEAD SELF-ATTENTION RESIDUAL TRANSFORMER
Author:	ANUSHA RUDRARAJU, Dr.S VENKATA LAKSHMI
Abstract:	Alzheimer disease (AD) is a neurodegenerative dis- order that is progressive and needs proper and computationally efficient models in order to detect the disease early and predict its progression accurately. EEG signals are inexpensive and non- invasive measures of functional brain activity in relation to cognitive impairment. Nevertheless, most of the current deep learning methods to detect dementia based on EEG data are characterized by a high level of computational complexity and low scalability, limiting their use in clinical settings in real- time. To tackle these issues, a Lightweight Convolutional Multi- Headed Self-Attention Residual Transformer Network (CMhS- RTN) is presented in this paper aimed at effective prediction of Alzheimer disease progression using EEG signals. The suggested architecture combines 1D convolutional layers to extract local spectral features with multi-head self-attention mechanisms to capture global feature interactions, with minimal computational demands. The model was tested on the Synthetic EEG Dementia Dataset which included samples of Normal Control (NC), Mild Cognitive Impairment (MCI) and Alzheimer’s Disease (AD) groups. Experimental findings prove that CMhS-RTN works better and performs with an accuracy of the highest level of 95.6, which is better than the current state-of-the-art models. Notably, architecture would experience less parameter complexity and would train faster, thus suitable to the real-time clinical applications and scalable implementation in healthcare software systems.
Keywords:	Alzheimer Disease; EEG Signals; Self-Attention; Residual Transformer; Prediction Of Disease Progression;
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	A HYBRID SPIKING CΟNVΟLUTIΟNAL NЕURAL NЕTWΟRK AND SNN-U-NЕT NЕURΟMΟRPHIC FRAMЕWΟRK FΟR RЕAL-TIMЕ HЕALTHCARЕ PRΟCЕSSING AND PRIVACY-PRЕSЕRVING MЕDICAL DATA ANALYSIS
Author:	P. VIMALADEVI, P. SUGANTHI
Abstract:	Nеurοmοrphic cοmputing, inspirеd by thе brain’s еvеnt-drivеn and еnеrgy-еfficiеnt infοrmatiοn prοcеssing mеchanisms, οffеrs a transfοrmativе pathway fοr rеal-timе and privacy-prеsеrving hеalthcarе applicatiοns. This study prοpοsеs a Hybrid Spiking Cοnvοlutiοnal Nеural Nеtwοrk (SCNN) and SNN-U-Nеt architеcturе dеsignеd tο еnhancе mеdical imagе sеgmеntatiοn, physiοlοgical signal analysis, and sеcurе οn-dеvicе infеrеncе. Thе hybrid mοdеl lеvеragеs SCNN fοr prеcisе tеmpοral fеaturе еxtractiοn using spikе-еncοdеd data, whilе SNN-U-Nеt pеrfοrms finе-grainеd spatial rеcοnstructiοn thrοugh a spiking еncοdеr–dеcοdеr structurе. Еxpеrimеntal еvaluatiοns dеmοnstratе that thе prοpοsеd framеwοrk significantly οutpеrfοrms traditiοnal dееp lеarning mοdеls and standalοnе spiking nеtwοrks. Οn thе BraTS MRI datasеt, thе hybrid mοdеl achiеvеd a Dicе scοrе οf 0.92, IοU οf 0.85, Prеcisiοn οf 0.91, and Rеcall οf 0.93, surpassing cοnvеntiοnal U-Nеt (0.88 Dicе, 0.81 IοU). Οn thе DRIVЕ rеtinal datasеt, thе hybrid mοdеl rеachеd 0.94 Dicе and 0.89 IοU, οutpеrfοrming bοth SCNN and SNN-U-Nеt individually. Systеm еfficiеncy tеsts cοnfirm substantial pеrfοrmancе gains, with infеrеncе latеncy rеducеd tο 10 ms and еnеrgy cοnsumptiοn lοwеrеd tο 0.55 J/framе, rеprеsеnting a 55% rеductiοn in latеncy and 70% imprοvеmеnt in еnеrgy еfficiеncy cοmparеd tο U-Nеt. Biοsignal analysis furthеr validatеs thе mοdеl’s gеnеralizatiοn ability, achiеving 98% accuracy, 97% sеnsitivity, and 96% spеcificity οn ЕCG arrhythmia dеtеctiοn. Privacy еvaluatiοn shοws a rеductiοn in mеmbеrship infеrеncе attack succеss ratе tο 6%, cοmparеd tο 26% in traditiοnal mοdеls, duе tο spikе-cοdеd data οbfuscatiοn and οn-dеvicе prοcеssing capability. Οvеrall, thе Hybrid SCNN+SNN-U-Nеt framеwοrk еstablishеs a high-pеrfοrmancе nеurοmοrphic cοmputing sοlutiοn that οffеrs rеal-timе prοcеssing, lοw еnеrgy cοnsumptiοn, high sеgmеntatiοn accuracy, and strοng privacy guarantееs, pοsitiοning it as a prοmising architеcturе fοr nеxt-gеnеratiοn intеlligеnt hеalthcarе systеms.
Keywords:	Nеurοmοrphic Cοmputing, Spiking Cοnvοlutiοnal Nеural Nеtwοrk, SNN-U-Nеt, Hybrid Architеcturе, Mеdical Imagе Sеgmеntatiοn, Rеal-Timе Prοcеssing, Еnеrgy-Еfficiеnt Hеalthcarе AI, Privacy-Prеsеrving Mеdical Data, Spiking Nеural Nеtwοrks, Biοsignal Analysis
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	ADAPTIVE HYBRID CRYPTOGRAPHIC FRAMEWORK FOR SECURE KEY MANAGEMENT IN HART NETWORKS: A LATTICE-BASED AND ELLIPTIC CURVE APPROACH
Author:	CHAGANTI SURESH NAIDU, Dr M.N.V KIRANBABU
Abstract:	Highway Addressable Remote Transducer (HART) networks remain integral to industrial IoT environments, yet conventional cryptographic methods face challenges in achieving scalability, efficiency, and resilience against quantum-enabled attacks. This paper presents an adaptive hybrid cryptographic framework that integrates lattice-based primitives with optimized elliptic curve mechanisms to deliver lightweight, post-quantum secure key management. The design employs Ring-Learning with Errors (RLWE) for robust master key generation, paired with Curve25519-based ephemeral sessions to provide rapid rekeying and forward secrecy. The major contribution of this research is the development of an adaptive hybrid key management framework that integrates RLWE-based post-quantum cryptography with Curve25519-based elliptic curve cryptography for HART networks. This study creates new knowledge by demonstrating that quantum-resistant security, lightweight key management, and real-time communication requirements can be achieved simultaneously through an adaptive hybrid cryptographic architecture suitable for resource-constrained industrial IoT environments. To validate the framework, a dual-layer strategy was applied, including formal verification in ProVerif, quantum adversary modeling under the Quantum Random Oracle Model (QROM), and large-scale benchmarking across 12,000 simulated HART communication sessions. The results show a 92.3% reduction in key compromise probability relative to polynomial pre-distribution, while sustaining exchange latencies below 15 ms, ensuring compatibility with resource-limited field devices. Furthermore, the framework reduced the effectiveness of Shor- and Grover-based quantum attacks by over 95% compared with elliptic curve–only models. These findings demonstrate that the proposed hybrid architecture provides a scalable, efficient, and quantum-resilient pathway for securing HART infrastructures against both present and emerging threats.
Keywords:	HART networks, Industrial IoT security, Adaptive cryptography, Lattice-based cryptography, Elliptic curve cryptography, Hybrid key management, Secure key distribution.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	HYBRID DEEP LEARNING MODEL FOR PREDICTIVE MAINTENANCE OF POWER CONVERTERS IN ELECTRIC VEHICLE CHARGING STATIONS
Author:	BIRUDULA VENKATESH REDDY, PRAMEELA RANI MANNAVA, PALLE DEEPAK REDDY, NELAKUDITI KRISHNAVENI, MYLAVARAPU KALYAN RAM, NIRMAL MAHESH, G. MURALI, RADHIKA PEERIGA
Abstract:	Reliable and intelligent charging station maintenance is a growing need with the increasing deployment of electric vehicle (EV) charging stations, particularly in dynamic charging conditions, due to the power converter. The research presented a novel Hybrid CNN-LSTM-Attention deep learning framework for predictive maintenance of EV charging station power converter based on the multi-sensor data of operation. The main aim of the study was to achieve greater accuracy of fault prediction in the early stage and to estimate the Remaining Useful Life (RUL) in different working conditions of the electrical and thermal variables of the system, and to have more efficient maintenance scheduling. Charging station simulations were implemented in the MATLAB/Simulink simulation environment to create a realistic multi-sensor dataset of the voltage and current measurements, temperature, harmonic distortion, switching frequency, ripple voltage, and converter efficiency parameters. In the proposed framework, Convolutional Neural Networks (CNN) were used for the spatial feature extraction, Long Short-Term Memory (LSTM) networks were employed for the temporal degradation learning, and an Attention Mechanism was introduced to prioritize the features based on fault sensitivity. The experimental results showed that the proposed model was able to attain an accuracy rate of 98.4%, precision of 97.8%, recall of 98.1%, and F1-score of 97.9%, which is higher than conventional machine learning and standalone deep learning models. The framework also had low prediction error for Remaining Useful Life and low false alarm frequency in a noisy charging environment. The results showed that the proposed hybrid architecture was a successful approach to capture the spatial degradation signatures and also the long-term temporal operation dependency. The proposed predictive maintenance system has the potential to enhance the reliability of charging stations, lower maintenance downtime, mitigate unexpected converter failures, and contribute to the creation of intelligent and sustainable charging infrastructure for EVs.
Keywords:	Predictive Maintenance, Electric Vehicle Charging Stations, Power Converters, Hybrid Deep Learning, CNN-LSTM-Attention, Remaining Useful Life Estimation
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	VYOMA: A REAL TIME DIGITAL TWIN FRAMEWORK FOR ASTRONAUT HEALTH AND PERFORMANCE OPTIMIZATION
Author:	DR. T. ADILAKSHMI, DR. E. SHAILAJA, DR. T. JALAJA
Abstract:	Long duration human spaceflight missions introduce complex physiological and cognitive risks that are further exacerbated by communication delays limiting real time ground intervention. This paper presents VYOMA, a cyber physical Digital Twin (DT) framework developed for the ISRO Human Space Flight Centre, which enables autonomous astronaut health monitoring through a closed loop Physical–Digital–Physical (PDP) paradigm. The system fuses multi modal telemetry, including cardiovascular, neuromuscular, biochemical and environmental signals, into a unified digital replica that is continuously synchronised for real time state estimation. A hybrid artificial intelligence pipeline combining deep learning, probabilistic inference and reinforcement learning enables predictive analytics under uncertainty, while Verification, Validation and Uncertainty Quantification (VVUQ) mechanisms together with explainable artificial intelligence provide confidence aware and interpretable decision support. To address deep space latency, VYOMA performs low latency edge inference, allowing early detection of critical conditions such as bone demineralisation and radiation exposure up to 48 hours in advance. Validation on statistically calibrated datasets derived from NASA, ESA and PhysioNet repositories demonstrates a predictive accuracy of 0.96, a root mean squared error of 0.28 and an end to end latency of 47 ms, which outperforms convolutional, recurrent, gradient boosting and transformer baselines evaluated under identical conditions. The results highlight the potential of scalable Digital Twin systems for autonomous healthcare in future interplanetary missions, and the architecture generalises to terrestrial mission critical healthcare domains such as telemedicine, military medicine and disaster response medicine.
Keywords:	Digital Twin, Astronaut Health Monitoring, Cyber Physical Systems, Predictive Analytics, Edge Artificial Intelligence
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	DEEP CONVOLUTIONAL NEURAL NETWORK FRAMEWORK FOR DEPRESSION DETECTION THROUGH FACIAL EXPRESSION AND VISUAL BEHAVIOR ANALYSIS
Author:	MS. AFREEN SUBUHI, DR MUNIRAJU NAIDU VADLAMUDI
Abstract:	Visual cues such as facial expressions, eye movements, and micro-gestures often reveal subtle psychological states associated with depression. This study introduces a Convolutional Neural Network (CNN)-based framework for automated depression detection using facial expression and visual behavior analysis. The model employs transfer learning with pre-trained architectures including VGG19 and ResNet50, fine-tuned on a multimodal dataset combining the DAIC-WOZ and AffectNet corpora. The preprocessing pipeline includes facial landmark detection, histogram equalization, and temporal frame selection to ensure consistent feature representation. The CNN framework extracts spatial and affective features that correlate with depressive symptoms, such as reduced smile intensity, lowered gaze, and diminished facial dynamics. Experimental validation used 10-fold cross-validation and evaluation metrics including precision, recall, and F1-score. The model achieved 91.2% classification accuracy, surpassing previous approaches using handcrafted features and shallow learning. Results confirm that CNN-based visual emotion recognition can serve as a reliable non-invasive indicator for early depression screening, particularly when integrated into multimodal mental health monitoring systems. Future extensions include incorporating temporal dynamics through hybrid CNN-LSTM pipelines for real-time emotion tracking in telehealth environments.
Keywords:	Depression detection, Convolutional Neural Networks (CNN), Facial expression analysis, Visual emotion recognition, Affective computing, Deep learning for mental health, Non-verbal behavior modeling, Early intervention systems
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	AI BASED HYBRID TEMPORAL ENSEMBLE NETWORK FOR CONSTRUCTION EQUIPMENT FAILURE PREDICTION USING SENSOR AND MAINTENANCE DATA
Author:	K. ARUNA KUMARI, KATAM NAGA LAKSHMAN, KRISHNA SAI UJWAL KAMBHUMPATI, JUVVALA SAILAJA, CHINNARAO KURANGI, SUCHETA PANDA, NAGULMEERA SAYYED, FRANCIS MULAGANI
Abstract:	Project delays, high maintenance expenses and lower efficiency are all caused by unexpected failures of construction equipment. In this study, a new AI-based Hybrid Temporal Explainable Ensemble Network (HTEEN) for intelligent equipment failure prediction is proposed, leveraging sensor data and maintenance records. It incorporates Bidirectional LSTMs, attention-based learning, XGBoost ensemble classification, adaptive fusion, and SHAP explainability analysis within a framework for predictive maintenance in the dynamic construction environment. The proposed model was evaluated on 52,000 operational instances and achieved 97.2% accuracy, 96.4% precision, 95.8% recall, 96.1% F1-score, and 97.5% ROC-AUC, which are better than those of standard machine learning and deep learning models. The key failure indicators identified via SHAP analysis were: vibration intensity, hydraulic pressure and engine temperature. The proposed framework increased prediction reliability, reduced false alarms, and improved maintenance decision-making. The study confirms the benefits of AI-based predictive maintenance in intelligent construction systems, significantly reducing equipment downtime, optimising maintenance plans, and enhancing operational safety.
Keywords:	Construction Equipment Failure Prediction, Predictive Maintenance, Bidirectional LSTM, Xgboost, Explainable Artificial Intelligence, Iot Sensor Analytics
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text

Title:	HOW CAN A CLOUD ARCHITECTURE UNIFY DATA PROCESSING AND AI-ASSISTED DECISION SUPPORT FOR MARKETING AUTOMATION?
Author:	HASSAN ALI AL–ABABNEH, IBRAHIM ALKHALDY, HEBA MOHAMMAD ALTARAWNEH, OLHA POPOVA, AFSHAN AZAM, FARAH HANNA ZAWAIDEH
Abstract:	Services are fragmented, data management, campaign execution and analytics are all separate and this results in disjointed data sources and poor traceability of automated marketing decisions. This study tackles the question of whether a common data processing architecture can be used to merge data processing and AI-driven decision support in Marketing Automation. An exploratory cross-company analysis was used along with a design-science protocol. Requirements were identified through information systems and marketing-automation research and a multi-layered artifact spanning from data ingestion, identity management, analytical processing, decision orchestration, campaign execution, monitoring and governance was created. The artifact was assessed by using the criteria based architectural comparison and exploratory analysis of ten company cases. For the provided data set, the correlation between the automation level and the marketing ROI was found to be 0.948 with a very high significance level (p < 0.001). Simultaneous entry of automation and AI investment, however, showed that automation was significant (p = 0.006) while AI investment was not (p = 0.336). So, the results do not indicate an independent or causal impact of AI but rather an AI effect in the context of integrated automation. A primary contribution is a traceable IT architecture that connects the data flows of marketing to the analytical decisions and feedbacks. The small cross-sectional sample and the heterogeneous industries, and the secondary nature of the data, restrict the generalizability of the findings.
Keywords:	Cloud Computing; Marketing Automation; Data Processing; Decision Support; Artificial Intelligence
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st July 2026 -- Vol. 104. No. 14-- 2026
Full Text