This study introduces a novel intelligent enhanced cyber resilient control strategy for frequency stabilization in a cyber physical power system (CPPS). The proposed cascaded Gated Recurrent Unit-Proportional Integral Derivative (GRU-PID) controller is tested for various power transaction schemes such as Bilateral, Poolco and Contract Violation in the presence of Denial of Service (DoS) attack. The GRU-PID controller captures nonlinear temporal dynamics to improve disturbance mitigation capability and frequency stability. The controller parameters are optimally adjusted through Modified Human Conception Optimizer (MHCO). The tolerance of the controller is also validated by the inclusion of non-linearities. Furthermore, sensitivity analysis under +/- 25% parameter uncertainty is performed to demonstrate the adaptability of the controller. A comparative analysis of the MHCO trained GRU-PID controller with a conventional PID controller tuned by Particle Swarm Optimization (PSO), MHCO and Genetic algorithm (GA), Fractional Order PID (FOPID) tuned by MHCO and FuzzyPID tuned by MHCO is performed. Lyapunov stability analysis and Bode plot analysis is also performed to check the system stability. The results illustrate the superiority of the controller in ensuring stable and secure operation of the power system. The recommended controller efficiently preserves system integrity and functionality during and after cyber intrusions.
The computed tomography enables accurate lung tumor identification and early treatment planning, but existing studied face challenges inclusing image quality variations, tumor complexity and performance degradation. Therefore, an Adaptive Crossover Boosted Reptile Search optimized Adaptive Modified Generator Cross Transformer model is proposed in this paper for segmenting lung tumors by overcoming these limitations. The images are collected from diverse data sources that are pre-processed to enhance generalization ability as well as image quality while minimizing overfitting issues. The Modified Atrous Convolutional Cross Transformer model is applied as a feature extractor that has the ability to efficiently extract local as well as global features. For the segmentation task, the Improved Generator-Discriminator model is introduced that uses UNet as a base model and skip connections to preserve detailed spatial information. An Adaptive Crossover Boosted Reptile Search Optimization algorithm is implemented for optimizing hyperparameters by integrating crossover and adaptive weighting strategies for avoiding local optimal issues. Different kinds of experimental validations are performed to analyze the superior performance of the proposed model. Among all analyses, the proposed model attained the highest prediction accuracy of 98.71 +/- 0.18% and a Dice score of 0.9866 +/- 0.0018.
Face recognition systems often encounter significant challenges when dealing with variations in pose, facial expression, and occlusion, which can greatly reduce recognition accuracy. Existing deep learning approaches tend to struggle with these issues due to limited adaptability and insufficient handling of image noise and feature variability. To address these limitations, this research proposes a novel deep learning framework for robust face recognition that leverages three types of facial images: pose, expression, and occlusion. In the preprocessing stage, a Median Absolute Deviation-based Wiener Filtering (MAD-based WF) technique is applied to effectively reduce noise and enhance image clarity. Subsequently, data augmentation is performed to enrich the training dataset and improve model generalization. For feature extraction, a combination of methods, including Active Shape Model (ASM), Histogram of Oriented Structures (HoS), Improved Local Gabor Binary Pattern Histogram Sequence (ILGBPHS), and Scale-Invariant Feature Transform (SIFT) is used to capture robust and discriminative facial features. Finally, an Improved Attention Mechanism-based Parallel Convolutional Neural Network (IAM-PCNN) is employed to perform accurate face recognition. At training data 90%, the proposed IAM-PCNN model achieves maximum accuracy of 0.973 for the pose case, 0.964 for the expression case and 0.962 for the occlusion case, outperforms existing models in dataset 1.
Partial differential equations (PDEs) are fundamental mathematical tools for modeling physical phenomena across scientific and engineering domains, yet their numerical solution remains computationally challenging, especially on complex geometries. To address the limitations of existing neural PDE solvers in balancing local physical constraints with global dependency modeling, we present MH-Transolver, an enhanced architecture extending Transolver, which integrates hybrid multi-head attention mechanisms to improve PDE solving on general geometries. Building upon the physics-attention foundation of Transolver, our model introduces a hybrid multi-head attention mechanism that synergistically combines localized physical state learning with global contextual modeling. The key innovation lies in the interleaved integration of physics-aware slicing and multi-head attention blocks, enabling the model to capture both intricate local physical correlations and long-range dependencies without compromising computational efficiency. Evaluated across six standard benchmarks, MH-Transolver achieves improved accuracy with an average relative error reduction of 19.8% compared to the original Transolver, including a 56.9% error reduction on Navier-Stokes equations and a 21.1% improvement on Darcy flow. Extensive experiments confirm its superior performance in industrial-scale simulations, including complex aerodynamic designs with irregular boundaries, while maintaining linear complexity. This work establishes a new paradigm for high-fidelity physics-informed learning through attention mechanism fusion.
Precise threat detection in cloud-based environment is crucial for detecting unauthorized access, security attacks, and vulnerabilities. Several Artificial Intelligence based threat detection techniques have been implemented and they provide better threat detection result but they are struggled with lower detection results, poor scalability, and model complexity. This study introduces an automated Modified vision transformer-based Federated learning network for effective threat detection in cloud environment. The proposed framework incorporates several layer like perception, transport, cloud, and Intelligence layers. The perception layer is responsible for data collection whereas transport layer enables secure data transmission and performs mutual authentication. In the cloud layer, the threat detection is performed where Modified Vision Transformer for deriving spatio-temporal features, which is vital for attack detection. The Federated Learning is applied for contextual analysis, that utilizes Gated Recurrent Units for local model training and Artificial Hummingbird Algorithm for model aggregation. For accurate threat detection, the Hybrid Quantum Generative Adversarial Network is designed. Granular computing is applied in the Intelligence layers to offer context-aware threat actions. Simulation validation on attack datasets indicates that the proposed mechanism attains 98.89% detection accuracy and 98.48% security rate, which is superior than existing works.
This study proposes a combined prediction model to improve the accuracy of radio interference prediction in transmission lines affected by multiple environmental factors at high altitudes. The model integrates Complementary Ensemble Empirical Mode Decomposition (CEEMD), Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and Multi-Head Attention (MHA). First, a long-term observation station was set up at an altitude of 2,420 meters to collect radio interference data 20 meters from the projected edge of the phase conductor. A normality test was performed on the collected data. Spearman correlation analysis was conducted to filter radio interference data and related environmental factors. Second, to address the nonlinear characteristics of radio interference data, CEEMD was used to decompose the data into relatively stationary subsequences, reducing the impact of random fluctuations on the prediction model. Third, these subsequences were fed into the combined model, integrating CNN and BiLSTM, with the MHA mechanism assigning weights to the BiLSTM hidden layers. To improve predictive performance, an Improved Whale Optimization Algorithm (IWOA) was applied to optimize two BiLSTM hyperparameters, yielding the optimal parameter combination. Finally, the overall radio interference prediction was reconstructed by combining the predictions from each subsequence.
In this paper, a new hybrid AI-based cybersecurity system is presented that employs TabNet model to combine with the Pelican Optimization Algorithm (POA) to optimize features adaptively in clouds. The originality is in the fact that it combines time-frequency-based feature extraction with the help of STFT and metaheuristic feature selection with the help of POA which is specifically designed to work with non-stationary cybersecurity event data. This integration results in better interpretability, scalability, and detection accuracy of real-time monitoring of cloud threat. The growing popularity of cloud computing has increased the need to have advanced cybersecurity measures that are capable of combating emerging and less obvious threats of cyberattacks. This paper describes a cybersecurity system based on AI to select the most successful features in a cloud environment, where TabNet and Pelican Optimization are used. With the help of STFT, SMOTE and AES-256 encryption, it identifies the different types of cyberattacks with 98.24% accuracy. It is a resilient real-time threat identification and mitigation system that is scaled and interpretable. Such an approach can result in an increase in cybersecurity protection and can therefore be a useful asset toward the security of cloud-based systems against current and upcoming cyber-attacks.
Traditional neural machine translation systems still exhibit shortcomings in learning efficiency and output quality, particularly in handling long sentences, processing specialized terminology, and modeling syntactic structures. To address this, this paper proposes a terminology-aware cross-attention recurrent neural translation model (CA-RNN) for professional English translation in specialized domains. This model enhances performance across four dimensions within the Transformer encoder-decoder framework. Bilingual term pairs are vectorized into continuous key value pairs and injected into both encoder and decoder attention layers, enabling explicit term-level alignment. Building upon absolute position encoding, it incorporates dependency tree depth information to construct syntactic structure position encoding (STRU), enhancing modeling of long-range structural dependencies. LSTM and Masked-LSTM are employed to model term sequences and provide recursive initialization for the decoder, capturing temporal relationships between terms and between terms and context. A term preference distribution is constructed at the output layer, dynamically adjusting term generation strength through a learnable gating function combined with Softmax probability fusion. Experimental results on datasets such as WMT, Multi30k, and IWSLT demonstrate that CA-RNN significantly outperforms multiple mainstream baseline models on both BLEU and ChrF metrics, achieving 38.55 BLEU and 50.11 ChrF on the WMT dataset.
Trajectory tracking in robot manipulators is challenged by unmodeled dynamics, unknown parameters, and external disturbances, which typically degrade the performance of traditional fixed-gain controllers. This article aims to address the challenge of dynamic gain adaptation and disturbances by developing a novel robust adaptive control scheme. The proposed controller integrates dual neural networks within a PID framework. The first network, based on radial basis functions, estimates uncertainties and dynamically adjusts the PID gains. The second network, utilizing a hyperbolic tangent activation function, explicitly compensates for the manipulator's unmodeled dynamics. Theoretical analysis is provided to demonstrate that the closed-loop system achieves uniformly ultimately bounded stability. The proposed method is validated through real-time experiments on a two-degree-of-freedom robot manipulator. In addition, comparative experiments, including abrupt external torque disturbances, demonstrate that the proposed dual-network scheme outperforms state-of-the-art adaptive controllers by achieving the lowest root mean square tracking error across all joints. The main contribution of this work lies in providing one of the few existing implementations of a neural-network-based PID controller that compensates for the unknown dynamics of the system, supported by both rigorous theoretical analysis and experimental validation.
This study introduces a novel approach for ranking fuzzy numbers by integrating Cost-Benefit Analysis with vector normalization. In this method, a generalized fuzzy number is preferred when it yields higher benefits and lower costs. The left area of the fuzzy number, representing benefit, is more desirable when larger, whereas the right area, considered a cost, is more favorable when smaller. To ensure a balanced evaluation, a Profit Index is computed, incorporating both horizontal and vertical dimensions for a comprehensive ordering of fuzzy numbers. This ensures that all relevant information of the generalized fuzzy numbers is taken into account, leading to consistent and reliable rankings. The proposed technique is well-suited for application in various fuzzy multi-criteria decision-making (MCDM) frameworks, offering enhanced support for decision-makers in practical, real-world contexts.
Knee pain is a prevalent form of joint and bone pain that is frequently experienced. There are multiple factors that can lead to knee pain. A meniscus tear is among the injuries that can cause knee pain. Therefore, early detection of this issue can help to prevent more pain and healing the patients. Medical imaging is a good method of detecting this issue. In the meantime, MRI is the best imaging method to use in the diagnosis of a meniscus tear. A novel method has been proposed in the context of this research paper that enhances the identification of knee meniscus tear, depending on MRI and deep learning. In the paper, the authors suggest a hybrid method to use Xception neural network with a recently designed metaheuristic algorithm, denoted as Promoted Namib Beetle Optimizer (PNBO), which is inspired by the namib beetle. PNBO is used to modify the Xception neural network in order to improve its performance.
Human activity recognition is a quickly evolving approach that identifies the physical activities of human beings from visual or sensor data. Still, human activity recognition using visible images impacts the image quality due to varying lighting conditions. Hence, the thermal video is used for solving such issues. However, the recognition becomes complex due to noisy data and depends on handcrafted features. To tackle such issues, this article develops the novel model named Searched Binary Convolutional Neural Network (SeBCoN) for human activity recognition. The frames are extracted from the input video, and the brightness of the frames is enhanced by the Local Gamma Transform. The human segmentation is performed by the Lovasz-Softmax loss function-based PyramNet, and the pose representation is done using a multi-scale convolution fusion deep residual network (Mscf-ResNet). Finally, the developed SeBCoN recognizes the human activities into direction, eating, discussion, greeting, posing, phone talk, purchasing, smoking, sitting, taking a photo, walking, waiting, running, duckwalk, and crawl. Moreover, the developed SeBCoN-based human activity recognition attained the optimal accuracy, sensitivity, specificity, F1 score, false omission rate (FOR), and Matthews correlation coefficient (MCC) of 96.37%, 95.36%, 97.14%, 96.36%, 4.513%, and 0.935.
Accurate solar irradiance forecasting is essential for intelligent energy management in cyber-physical transportation systems such as solar-powered electric vehicles (SEVs). However, the highly non-stationary and multiscale nature of meteorological signals makes reliable prediction challenging for conventional deep learning models, which often struggle to capture nonlinear relationships between environmental variables and energy dynamics. To address this limitation, this study proposes a cybernetics-inspired hybrid forecasting framework named FCM-WaveTransNet. The proposed architecture integrates Fuzzy C-Means clustering (FCM) for regime-aware data organization, Convolutional Neural Networks (CNN) for local feature extraction, Wavelet Neural Networks (WNN) and Wavelet-enhanced Bidirectional Long Short-Term Memory networks (WBiLSTM) for multiresolution temporal learning, and a Transformer attention mechanism for modeling long-range temporal dependencies. The model was evaluated using fifteen years of solar irradiance data obtained from the Photovoltaic Geographical Information System (PVGIS) across multiple Saudi Arabian cities. Experimental results demonstrate that the proposed framework achieves near-perfect predictive accuracy ( R 2 approximate to 0.998 ) while significantly reducing RMSE and MAPE compared with state-of-the-art baseline models. These results highlight the robustness and strong generalization capability of FCM-WaveTransNet for reliable solar irradiance forecasting in intelligent renewable energy systems.
High-voltage disconnect switches are essential devices in power transmission substations. They are responsible for isolating circuits and ensuring safety during maintenance and emergencies. Failures can occur during the operation of these switches. In this context, an automated approach can assist operators in monitoring the state (open or fully-closed) of each switch to prevent accidents, equipment damage, and service interruptions. With this in mind, we present the Switch Detection Web App, an easy-to-use software that offers real-time detection options or video file analysis for inspecting the state of high-voltage disconnect switches. It abstracts the complexity of processing, prediction, video labeling, report generation, and storage of detections by integrating front- and back-end technologies. The solution supports three deep learning models trained with custom datasets for detecting two types of Lorenzetti switches (230 and 72.5 kV) and a generic version that covers both types. A pilot study showed highly satisfactory performance of the application for both file-based and live detections, with mean Average Precision (mAP@0.5:0.95) metrics indicating that the models tend to predict with a high degree of reliability (mAP@0.5 : 0.95, 230 kV = 0.930, 72.5 kV = 0.993, Generic = 0.980).
A genetic dual mutation-based Seagull optimized random majority voting machine learning model is developed for energy-efficient and secure data transmission in wireless sensor networks-based internet of things environments. The threshold-based cluster head selection is optimized to improve network longevity and reduce energy consumption. Cloud storage uses data preprocessing approaches including data cleaning, normalization, and synthetic minority oversampling to eliminate inaccurate records, transform the dataset's distribution to be more balanced, and address class imbalance. Moreover, the random majority voting-based logistic regression model is employed for classification and detection, leveraging ensemble learning to enhance classification accuracy and reduce false alarms. To refine the detection system, the genetic dual mutation-based Seagull optimization algorithm is introduced for hyperparameter tuning. This optimization framework integrates the Seagull optimization algorithm with dual mutation operations and genetic algorithm to address local optima issues, improving classification accuracy and optimizing model parameters. Experimental results demonstrate that the proposed technique achieves a classification accuracy of 98.6%, and F1 score of 98.4%, while consuming only 85J of energy, which significantly outperforms previous techniques across all measures. The findings validate the effectiveness of the developed model for securing data transmission in Wireless Sensor Network.
As Low Power and Lossy Networks (LLNs) become more prevalent, new forms of interaction are required. It was suggested that IPv6 Routing Protocols for Low-Power and Lossy Networks (RPL) be used to simplify connectivity among such low-powered gadgets. The main aspect of the proposed work is to achieve better routing performance and enhance the routing quality with the utilization of various constraints in the objective function. While building the RPL network, some of the parameters are optimized by Artificial Gorilla Troops with the African Vultures Optimization Algorithm (HAGT-AVOA) to enhance the routing performance. Thus, the novel network is built with objective measures such as energy, Expected Transmission Count (ETX), throughput, Link Quality Indicator (LQI), Received Signal Strength (RSSI), and hop count. Thus, the designed RPL network aims to satisfy the following factors, such as enhancing the more packet delivery, link quality, energy consumption, and managing the power. Hence, the performance is validated, and its results are compared with existing techniques. Based on the validation phase, the Packet Delivery Ratio (PDR) of the designed model is 96.87, which is better than traditional models. Thus, the findings demonstrate that it provides better routing performance in the RPL network.
