Journal of Intelligent Systems and Internet of Things

The rapid expansion of big data has accelerated the adoption of distributed computing frame- works such as Apache Hadoop and Apache Spark, enabling efficient large-scale data processing. While Spark’s in-memory computation model significantly enhances performance compared to Hadoop’s traditional MapReduce, the deployment architecture—whether Dockerized or non- Dockerized—plays a crucial role in affecting performance, scalability, and resource management. This study evaluates the impact of containerized and non-containerized multi-node cluster architectures on the performance of Hadoop and Spark, utilizing standardized workloads such as WordCount and TeraSort. Key performance metrics, including execution time, throughput, and resource utilization, are analyzed across various configurations with parameter tuning. Beyond pure performance benchmarking, the study also assesses the quality attributes of microservices in big data environments, focusing on scalability, maintainability, fault tolerance, and resource efficiency. The comparative analysis between monolithic and microservice-based architectures highlights the advantages of modularity and independent scaling inherent to microservices. Experimental findings indicate that Spark outperforms Hadoop on small to medium-scale workloads, while Hadoop exhibits superior robustness for processing extremely large datasets. Dockerized deployments offer better resource isolation and management flexibility, whereas non-Dockerized setups demonstrate reduced overhead under certain configurations. These insights contribute to optimizing deployment strategies and architectural decisions for microservices-based big data processing frameworks.

Internet and social media have become significant platforms for sharing real-time information, with rumors significantly affecting billions of people's perceptions. Considerably, Rumor recognition is the most challenging task on social media platforms. Numerous Deep Learning (DL) models have been developed to extract linguistic characteristics from short-text tweets for rumor prediction. However, these models struggles to capture the intricate spatiotemporal relationships presenting tweet interactions. To address this issues, Bidirectional Encoder Representation from Transformers with Attention based Balanced Spatial-Temporal Graph Convolutional Networks (BERT-ABSTGCN) was used. This model incorporates Spatial-Temporal Attention Mechanism (STAM) and a Spatial-Temporal Convolution Module (STCM) to effectively model the spatiotemporal dependencies within in tweet interactions to enhance rumor detection.  However, it constitutes to high degradation problem due to convergence issues. A popular solution to these problems is Residual Learning (RL), which introduces identity mappings to speed up training and enhance gradient propagation. However, traditional RL can only be used for layer-wise task refining, which severely restricts its capacity to grasp more generalized dependencies. However, conventional RL is restricted to layer-wise refinement within a single task limiting its ability to capture broader dependencies. To address this, the proposed work is included with a Cross-Residual Learning (CRL) in BERT-ABSTGCN named BERT with Attention-based Balanced Spatial-Temporal Residual Graph Convolutional Networks (BERT-ABSTRGCN) for efficient rumor detection and stance classification. CRL of BERT-ABSTRGCN enable intuitive learning across multiple tasks like rumor detection and stance classification using cross-connections. CRL establishes direct connections between shallow and deep feature representations, mitigating the vanishing gradient issue.   The fitted residual mappings in the CRL will facilitate the BERT- BERT-ABSTRGCN with the provided information by using the short cut connections and lowers the probability of model degradation. BERT-ABSTRGCN effectively identifies rumor with different stances about specific social media posts, thereby preventing the spread of rumors. Experimental evaluations show that BERT-ABSTRGCN achieves 95.62% accuracy on the PHEME dataset and 90.15% on Mendeley’s COVID-19 rumor dataset, significantly surpassing traditional models.

This paper offers a comprehensive framework for real or fake image classification based on three classifiers: a Standard Convolutional Neural Network (CNN), an EfficientNetV2 model based on transfer learning, and a re-trained GAN discriminator to address the challenges in deepfake detection. The CNN, with four convolutional blocks and dropout regularization, offers computational efficiency (87.2% accuracy, 15 ms/image inference), while EfficientNetV2 utilizes pre-trained ImageNet weights to achieve state-of-the-art performance (94.7% ac-curacy, AUC: 0.98) using hierarchical feature extraction. The fine-tuned and adversarial-pretrained GAN discriminator demonstrates niche strength in the detection of synthetic artifacts (91% recall for GAN-generated fakes). Training used augmented sets (rotation, shifts, and shear) to increase the generalization boost, with loss optimization and early stopping (binary cross-entropy) controlled through validation. Normalized test set validation affirmed EfficientNetV2's capability at balancing recall (94%) with precision (95%), although the GAN discriminator recorded a lead in adversarial resilience. All the models blended, an ensemble model achieved maximum accuracy (96.1%), under complementarities. Computational baselines showed trade-offs EfficientNetV2 accu-racy vs. resource bias (2.5-hour training), the CNN edge-compatibility, and the GAN discriminator arti-fact-sensitive specialization. The work encourages hybrid architectures and ensemble approaches to balance out single-model vulnerabilities, offering a flexible toolkit for deepfake warfare while emphasizing the need for hardware-aware deployment techniques and ongoing adaptation to changing synthetic approaches.

Diabetic retinopathy (DR) is one of the most common causes of blindness in the world, and early detection plays an important role in therapy. In this paper, we introduce a hybrid framework with the merger of sophisticated image processing techniques and deep learning models for automated DR detection from retinal fundus images. Information starts with an extensive preprocessing pipeline, which includes bilateral filtering for noise reduction, removal of artifacts, adaptive contrast enhancement and a precise segmentation in the U-Net architecture. To increase model robustness, random rotation augmentation was used to mimic different imaging positions. GLCM analysis is used to extract texture features capturing important lesion-related patterns, and deep features are extracted using a fine-tuned EfficientNet-B0 model. The hybrid feature set is then modelled by a Support Vector Machine (SVM) with the radial basis function kernel and optimized with cross-validation and hyperactive parameters. Experiments show our model can well solve the image heterogeneity problem and yields a high level of accuracy in diagnosis and grading corresponding severity requirements of DR stage.

To enhance the efficiency of edge-integrated Industrial IoT (IIoT) networks, this paper proposes a deep learning-based resource-scheduling framework for optimized asset booking in Wireless Sensor Networks (WSNs). The novelty of this work lies in the integration of a hybrid Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) model, which enables intelligent allocation of computational resources based on real-time asset demand characteristics. The proposed model is evaluated using the Intel Berkeley WSN dataset and demonstrates superior performance in terms of latency reduction, execution time, and resource utilization compared to conventional approaches such as Genetic Algorithm (GA), Improved Particle Swarm Optimization (IPSO), Long Short-Term Memory (LSTM), and Bidirectional Recurrent Neural Network (BRNN). With a maximum efficiency of 99.48% and the lowest observed average delay, the model proves effective for real-time industrial automation scenarios. This research contributes to the development of scalable, energy-efficient, and responsive WSN architectures by leveraging deep learning for asset booking in edge-IoT environments.

Air pollution, especially atmospheric particulate matter with aerodynamic diameters smaller than 10 micrometers (PM10), is one of the constant and serious environmental challenges in urban areas. Its consequences range from negative human health effects to broader ecological disruptions. With the increasing necessity of accurate and trustworthy forecasting devices in the sphere of air quality assessment, we propose a new hybrid-modeling platform that merges the sequential pattern recognition ability of Long Short Term Memory (LSTM) neural networks with fuzzy logic reasoning. The two approaches implemented in this model complement each other: while approaches taking into account the time dependence of the behavior of air pollutants address the complex temporal dynamics present in the problem, methods based on uncertainty propagate inherent uncertainties in the meteorological and environmental data. The model was trained using a well-structured, multi-variable dataset of hourly air quality and meteorological observations for five years (2019–2023) measured in Istanbul and further tested of January 2024 data. The hybrid approach outperformed all tested environments in prediction output, reaching an accuracy of 98% at the Aksaray traffic station, whereas standalone LSTM (97%) and fuzzy logic (94%) models performed lower. Importantly, it identified minute periodicity and pollution peaks with high fidelity and demonstrated robustness across diverse settings such as traffic-dense, industrial, rural and urban zones. These results place the hybrid LSTM–Fuzzy Logic model as a trusted and robust forecasting tool for predicting PM10 concentrations, providing valuable assistance to environmental policy-makers, urban planners, and public health authorities in efforts to reduce air pollution and protect the health of the population.

Facial Expression Recognition (FER) is a vital aspect of human-computer interaction with applications in healthcare, education security, and affective computing. Even with the success of deep learning, generalizability, interpretability, and efficiency of most systems, especially in uncontrolled settings, are still problematic. In this study, we propose an enhanced feature extraction technique based on Histograms of Oriented Gradient (HOG) where the central difference operator, not the conventional forward difference, used for gradient estimation. The modification enhances the accuracy of gradients, reduces truncation error, and leads to more stable facial feature descriptors. The enhanced HOG is tested on five popular datasets, CK+, JAFFE, MMI, ExpW, and AffectNet, using three traditional Machine Learning (ML) classifiers: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Random Forest (RF). Experimental results indicate uniform accuracy enhancements across all the classifiers and datasets, with improvements spiking to 7%–10% and recall and F1-score also witnessing marked increases. In this study, RF registered the maximum accuracy, 97.94%, on CK+ and 95.48% on AffectNet, hence solidifying its stability and dependability. This study shows how well mathematical optimization works with classical ML for FER. The approach we suggest provides an easy-to-understand, small, and quick alternative to deep models, making it perfect for real-time and resource-limited applications.

The transmission of energy is one of the main tasks of Electrical Engineering. Transmission lines are used for this purpose, which are susceptible to various problems such as short-circuit, overload, open circuit, and complex faults. From the perspective of smart grids, one of the open challenges is to have autonomous systems that allow the detection, classification, and location of faults in transmission lines. On the other hand, Artificial Neural Networks are computational tools used in classification and control tasks to be applied to different plants and systems. There are several ways to solve problems using ANNs; one is modularity. This strategy consists of dividing the problem into components that are easier to classify. In this way, a modular system is proposed that is composed of three ANNs: One for detection, one for classification, and one more for the location of faults in transmission lines. A simulation model of a three-phase electrical power system was built using Simulink MATLAB, employing a data transmission approach typical of smart grids. Supervised learning and WEKA software were used for network training. Databases were created using the potential difference and line current, as well as the ground fault impedance. The database was developed through cases and mathematical models, and the performance of the networks was evaluated in the simulated model. The results show that the proposed model allows the identification of all cases presented in the test stage (100%), which is a better performance than a single neural network (81.25%) that is responsible for detecting, classifying, and locating faults.

Detecting and identifying crimes in real time represents a very necessary aspect of public safety. Traditional systems are human based monitoring cameras, video surveillance systems are ineffective, time consuming and prone to mistakes. Automated solutions are much needed. Using convolutional neural networks (CNNs) to efficiently examine surveillance video footage is the main goal. This work presents a crime detection system based on deep learning. the study utilize UCF Crime dataset and four deep learning models: ResNet50, EfficientNetB2, Xception, and custom (CNN) were up-graded, trained, and tested. To guarantee best model performance, the suggested approaches required careful dataset preparation, pre-processing, and strategic data separation. By means of fine-tuning, each model addressed the constraints of conventional techniques and enhanced feature extraction and classification accuracy. With extraordinary performance measures of (99.53%) accuracy, (99.07%) precision, (98.43%) recall, and a (98.69%) F1 score, experimental findings show the superiority of the suggested system. These findings reveal the system’s high dependability in detecting and classifying criminal events, thereby far surpassing other CNN-based approaches. The model runs at an average inference speed of (30 ms per frame on CPU), with a lightweight model size of around (20 MB), These results demonstrate the system’s scalability, efficiency, and strong potential for intelligent surveillance applications. This study shows how scalable and effective deep learning models transform crime detection in surveillance systems to support public safety.

Nowadays, with the proliferation of mobile devices and the internet around the world that are available for everyone, and due to the low prices versus their high capabilities, images are considered one of the most common ways of transmitting information between users, advancement of image processing and editing tools, simplified the process of editing and changing photographs such as in magazines, newspapers, scientific journals, and on social media or on the Internet. As a result, the propagation of manipulated photographs that misrepresent the truth is prevalent, whether deliberate or inadvertent. We propose a method that uses deep learning based convolutional neural network in order to detect instances of the copy-move forgeries in images which can  help to ensure data authenticity in digital forensic investigations. In this case, our method is intended to improve digital evidence integrity by detecting complicated changes quickly and precisely. This work can supports cybersecurity applications like anti-fraud systems, fake news detection, and social media forensics. The findings of the experiment demonstrate that the suggested approach is capable of detecting forgery against multiple copies and post-processing activities. The dataset's images used for both training and testing are MICC-F2000, composed of 2,000 images, 700 tamper and 1,300 originals. The findings indicate a testing accuracy of 98.00% and a training accuracy of 99.17%.

Wireless Sensor Networks (WSNs) support intelligent data acquisition systems across environmental monitoring, industrial automation, and smart cities. As a fundamental enabler of the Internet of Things (IoT), WSNs rely heavily on battery-powered sensor nodes for sustained operation in dynamic and often remote environments. However, predicting battery lifetime in WSNs remains a critical challenge due to the complex interplay between environmental conditions and operational behaviors. Conventional energy models often fail to consider the simultaneous influence of temperature, humidity, and data traffic intensity on battery depletion rates. This study proposes a battery lifetime prediction model based on a Markov framework integrated with an exponential energy consumption function to address this issue. The model incorporates three primary variables—ambient temperature, relative humidity, and data movement to simulate energy usage dynamically. The framework calculates transition probabilities and energy load based on environmental states, enabling accurate forecasting. Additionally, the model evaluates the impact of different battery chemistries (Ni-MH, LiPo, Li-ion, and Alkaline) on lifespan performance across varying environmental scenarios. Simulation results reveal that temperature and humidity significantly influence energy depletion, while data transmission intensity plays a supporting role in high-traffic cases. LiPo and Li-ion batteries demonstrate superior performance and stability, especially under extreme environmental conditions. This study contributes a novel multi-variable model that bridges physical sensing environments with predictive battery analytics. The findings provide a foundation for strategic energy planning and adaptive deployment of WSNs in sustainability-critical applications.

The rapid development of intelligent computing has led to Internet of Things (IoT) applications and embedded devices suffering from severe constraints on energy, processing, and memory. This calls for fast and lightweight algorithms that maintain performance accuracy without draining resources or affecting response time. This paper presents a new hybrid metaheuristic algorithm that combines the advantages of four optimization algorithms to achieve efficient results and reduce computational complexity without compromising output quality. Experiments demonstrate significant improvements in performance and execution time compared to traditional algorithms, in addition to the algorithm's ability to scale and handle diverse workloads. The lowest improvement of the proposed algorithm compared to other algorithms was approximately 25.7%. This algorithm opens up prospects for effective applications in smart systems in urban and industrial areas.

Reducing energy losses and increasing power grid efficiency need accurate prediction of power consumption accurate prediction of future energy consumption requires the use of time series data. To overcome the shortcomings of conventional techniques for forecasting energy consumption in India for the period from 2 January, 2019 to 23 May, 2020, we used an attention mechanism, which is still relatively new and not well known. In this paper, we propose a new approach for predicting energy consumption by combining local feature extraction with convolutional neural networks (CNNs), long short-term memory (LSTM) to capture long-term temporal dependencies, and attention mechanisms to deal with the issue of information loss brought on by extremely lengthy input time series data. After high-dimensional features are extracted from the input data using a one-dimensional CNN layer, temporal correlations within historical sequences are captured using an LSTM layer.  In order to optimize the weighting of the LSTM outputs, strengthen the impact of important information, and enhance the prediction model as a whole, an attention mechanism is finally implemented. This integration improves the model's ability to represent complex spatio-temporal patterns. The mean absolute error (MAE) and root mean square error (RMSE) are used to assess the performance of the proposed model. The results demonstrate that the CNN-LSTM-Attention model outperforms conventional hybrid CNN-LSTM and LSTM models, demonstrating superior performance across a range of prediction scenarios. By supporting more reliable grid management, proactive intervention methods, and predictive maintenance, these developments contribute to reducing load imbalances and energy waste in India. The Future developments could see the proposed model extended to other time series prediction domains.

Area coverage and exploration of unknown environments by swarm robots autonomously is one of the challenges in the robotics domain. This paper proposes a new strategy for area coverage in two parts; firstly, enhancing a Pelican Optimization Algorithm (POA) using swarm robots to explore an unknown area. Secondly, merges many algorithms with the proposed POA, such as Timed Elastic Band (TEB) as a local planner for obstacle avoidance, SLAM (Simultaneous Localization and Mapping), and a training model which is called You Only Look Once version 8 nano (YOLOv8n) for person detection. The proposed POA algorithm successfully monitored a large area and achieved a high exploration ratio with minimal time. In this work, the new strategy is applied to a robot warehouse environment, utilizing a swarm of robots to explore the area and find targets, which are employees suffocated by the effects of chemical pollution. The simulation and real-world tests for a new strategy were done in the Robot Operating System (ROS) using the TurtleBot3 robot. The total time-consuming for exploration and detection time is less by POA, while the coverage ratio is the largest when compared with the original RRT exploration algorithm for empty, small, and large environments, respectively.

Blockchain technology has recently emerged as a fundamental pillar of decentralized and secure systems. However, many Proof-of-Work (POW) algorithms suffer from some challenges, including their inefficiency in discovering the value of Nonces due to their reliance on random attempts, which consume significant resources, energy, and time, making them difficult to use in lightweight blockchain environments, especially in resource-limited environments such as mobile devices and others. The main goal of this paper is to introduce a smart system that replaces random guessing with a more intelligent and predictive approach using deep learning models like CNN2D, GRU, LSTM, and hybrid models. The intelligent optimization algorithm (GWO) is also used, which has been enhanced with random Lévy jumps, in addition to improved clustering using a genetic algorithm. The results, after applying the system to health data across three difficulty levels (4, 6, and 8), showed that the intelligent neural model was the most stable and accurate, achieving the lowest error values ​​and the highest generalization ability, with a maximum error value of (0.0136) at the highest difficulty level (8). The hybrid GA–KMeans algorithm demonstrated high efficiency in improving clustering accuracy. It achieved the highest similarity index value (0.9980) and the lowest Davis-Bolden index value (0.0000), which plays a significant role in guiding searches efficiently and effectively. The CNN2D model also achieved ideal numerical results, but it is prone to overlearning, while the GRU neural model provided an efficient balance between stability and accuracy. Other hybrid models, such as GRU+CNN, have shown excellent performance, but with varying results. The proposed system proves to be an efficient and intelligent alternative to the low-cost random approach for Nonce discovery in lightweight blockchain environments.