This paper presents a survey of retrieval algorithms used in advertising recommendation and organic content recommendation systems. Modern digital platforms rely on retrieval-based models to efficiently match users with relevant advertisements or personalized content. This survey reviews key techniques including inverted index methods, collaborative filtering, content-based filtering, hybrid recommendation models, and the two-tower neural network architecture widely used in large-scale recommendation systems. The paper compares the objectives, data utilization strategies, and evaluation metrics of ad targeting and organic retrieval systems. Practical challenges such as cold-start problems, data quality, scalability, and privacy considerations are also discussed. This survey further highlights the growing connection between industrial recommendation pipelines and emerging retrieval mechanisms used in large language model (LLM) systems. This survey provides insights into the design principles of modern retrieval systems and outlines future research directions at the intersection of recommendation systems and LLM.
The June 2026 issue of this journal presents a comprehensive body of research advancing efficient engineering intelligence from foundational theory to real-world deployment, with strong alignment to the Sustainable Development Goals (SDGs). A significant cluster addresses SDG 7 (Affordable and Clean Energy) and SDG 9 (Industry, Innovation, and Infrastructure) through innovations in microbial fuel cells, high-voltage insulation reliability, artificial intelligence (AI) based battery management systems, and energy-efficient LoRa/LoRaWAN frameworks. These works emphasize energy sustainability, system resilience, and infrastructure optimization. A second cluster focuses on advanced electronics, control, and communication systems, including memcapacitor design, hybrid model predictive control, reflectarray antennas, and embedded intelligence for autonomous systems, demonstrating efficiency-driven engineering across hardware and system levels. A dominant cluster highlights SDG 3 (Good Health and Well-being), with applications in medical imaging, sepsis detection, breast cancer classification, and mental health analysis, leveraging deep learning, transformers, and hybrid AI models. Finally, contributions aligned with SDG 4 (Quality Education) explore gamified learning systems, virtual reality adoption, and SDG-integrated educational information systems, while complementary studies in agriculture, finance, and Internet of Things (IoT) further demonstrate the societal impact of intelligent systems. Collectively, these works reinforce the role of efficient, scalable, and data-driven engineering in addressing global challenges.
In this study, we test three machine learning methodologies − binary tree, k-nearest neighbors (k-NN), and neural networks (NN) − using a range of hyperparameters. These methods are applied to a dataset consisting of extracted time series characteristics (root mean square (RMS), skewness, and kurtosis from vibration signals of various bearings subjected to different fault conditions from the intelligent maintenance systems (IMS) dataset. We evaluate how effectively these methods classify the condition of the bearings using the provided dataset. We observe the top two methods, artificial neural network (ANN) 99.29% and binary tree 98.84%. With a difference of 0.45%, the binary tree is preferred over the complex ANN due to its ease of interpretation, transparency, and minimal computation requirements. Its integration as code in embedded controllers or electronic control units (ECUs) is more efficient, which makes them faster for real-time processing and safety-critical electric vehicle (EV) systems.
This study investigates the effect of surface pollution on a single-unit 11 kV glass suspension insulator using two-dimensional (2D) axisymmetric simulations in COMSOL Multiphysics. The developed model incorporates the electrical properties of glass, cement, steel electrodes, surrounding air, and a uniform pollution layer, with an applied AC voltage of 11 kV under quasi-static conditions. Simulation results demonstrate pronounced electric field intensification in the polluted configuration, particularly at the air–glass–cap triple junction region, where localized electrical stress is significantly higher compared to the clean condition. While the clean insulator operates within IEC 60383 recommended limits, the polluted model exhibits elevated peak electric field magnitudes, indicating increased flashover vulnerability. The findings highlight the strong influence of surface contamination, material permittivity, and geometric configuration on electric field distribution along the creepage path. This study establishes a reliable and computationally efficient predictive framework for optimizing insulator design, improving maintenance strategies, and enhancing the long-term reliability of high-voltage transmission systems, especially in pollution-prone environments.
In this research, the authors present a study analysis and compares two types of embedded internal permanent magnet synchronous motors (IPMSM) with U-type and V-type magnet configurations using finite element method (FEM) modeling to apply these motors to the currently popular electric vehicle industry. Parameters such as magnetic flux density, torque, cogging torque, back electromotive force (back-EMF), torque oscillation, and harmonic components were analyzed and compared; thereby identifying the advantages and disadvantages of the two IPMSM structures. Specifically, the V-type IPMSM motor offers higher efficiency, more stable torque, and a higher quality back electromotive force waveform with lower losses, making it suitable for high-performance applications such as electric vehicles and industrial automation. Meanwhile, the U-type structure has lower cogging torque, suitable for low-speed applications or those requiring high precision. Simulation results from the ANSYS Maxwell software show that the IPMSM motor is energy-efficient, has high power density, and operates smoothly, allowing for rapid acceleration, long range, compact configuration, and low maintenance; it uses permanent magnets on the rotor to eliminate losses, making electric vehicles lighter and more efficient than traditional motors.
Stock price forecasting presents a challenging problem due to factors like nonlinearity, seasonality, and economic volatility in financial data. Deep learning approaches can handle nonlinearity and complexity of financial data, but they often face limitations in capturing both local and global dependencies. This study introduces a hybrid Transformer–bidirectional long short-term memory (BiLSTM) model to improve stock price forecasting. Our method combines the strength of BiLSTM with the global context understanding of the Transformer by embedding a 1D convolutional layer. The model can efficiently capture short-term and long-term dependencies in stock data. Experimental results on various datasets show that our hybrid model outperforms other well-known models.
An analytic solution to the Riccati algebraic equation has been investigated by employing eigenvalue–eigenvector techniques combined with the Gram–Schmidt orthogonality process. An analytic solution to the Riccati algebraic equation has been investigated by employing eigenvalue–eigenvector techniques combined with the Gram–Schmidt orthogonalization process. The applied method is used to improve robust control of second and third-order state-dependent systems by handling nonlinearities. An H∞ controller is designed in this context via backstepping technique to enhance robustness and reduce computational effort. The effectiveness of this method has been demonstrated on a two-degree-of-freedom (2-DOF) robotic manipulator arm. Simulation results validate the performance of the controller, showing improved tracking accuracy, disturbance rejection, and overall system stability, thereby confirming the efficiency and applicability of the combined analytic Riccati algebraic equation and H∞ backstepping approach for nonlinear robotic systems.
Phishing emails continue to pose a significant cybersecurity threat, particularly through the increasing use of malicious attachments to evade traditional text-based detection systems. Most existing approaches focus primarily on email content, creating a blind spot in attachment-aware phishing detection. This paper proposes a multi-modal phishing email classification model that integrates email header features, body text analysis, and attachment inspection within an ensemble learning framework. Independent machine learning classifiers are employed for each email component, and a majority voting mechanism is used to determine the final classification decision. The proposed model is evaluated using publicly available email and attachment datasets that are combined to simulate attachment-bearing phishing emails. Experimental results demonstrate strong detection performance across multiple evaluation metrics. Nevertheless, the study acknowledges the limitation of using synthetically paired email bodies and attachments, which may not fully capture real-world semantic relationships. The findings highlight the importance of incorporating attachment-aware analysis into phishing detection systems and provide a foundation for future research on semantic consistency modeling and transformer-based architectures.
This study employs the mixed logical dynamical (MLD) framework for modelling, simulating, and controlling hybrid dynamical systems. Hybrid systems, which combine continuous-time dynamics and discrete logical events, pose significant challenges for conventional control strategies, such as proportional-integral-derivative (PID) controllers, particularly under complex operational constraints. To address these challenges, the MLD formalism provides a unified representation that integrates differential equations, logical rules, and inequality constraints. Based on the MLD model, a multivariable hybrid model predictive control (HMPC) approach is designed to optimize control system performance and operational efficiency over a prediction time horizon. At each sampling time step, a mixed quadratic programming (MIQP) optimization problem is solved online to determine the control law. The proposed control approach is applied to a three-spherical tank system, where simulation and experimental results demonstrate its effectiveness in ensuring stability, minimizing tracking errors, and satisfying physical constraints. These results underscore the relevance of MLD-based predictive control approaches for the optimization and advanced control of complex multivariable hybrid dynamical systems in industrial fields.
Solar cell technology has advanced rapidly in efficiency and material innovation. As a renewable energy source, solar cells help mitigate the global energy crisis. Perovskite-based solar cells have recently achieved efficiencies above 25%, surpassing conventional silicon cells. Among emerging materials, chalcogenide perovskites show great promise due to their superior stability compared to halide perovskites. However, they remain in the exploration stage, making accurate predictions of their electrical properties, especially bandgap, essential for assessing potential in solar cell applications. This study predicts bandgap values using computational methods, emphasizing efficiency and cost reduction compared to experimental approaches. Key features derived from collected data include oxidation state, electronegativity, coordination number, ionic radius, and density. Several machine learning (ML) algorithms: AdaBoost Regressor, gradient boosting regressor, support vector regressor, CatBoost Regressor, and k-neighbor regressor, were implemented using Python. The research process involved data collection, preprocessing (feature scaling, fusion, reduction, and selection), model training and testing with 5-fold cross-validation, and hyperparameter optimization to achieve optimal results. Among the tested models, CatBoost Regressor yielded the best performance, achieving a coefficient of determination (R2) of 69.34%, a mean absolute error (MAE) of 23.1%, and root-mean-square error (RMSE) of 29.49%, demonstrating its effectiveness in predicting chalcogenide perovskite bandgaps.
Real-time payment platforms operating at scale face an unforgiving operational reality: even brief outages translate directly into failed transactions, regulatory exposure, and eroded customer trust. Database replication and failover automation have matured considerably over the past two decades, yet a troubling blind spot remains. Recovery frameworks built for general-purpose distributed systems were never designed with settlement finality in mind, and that design omission leaves payment operators exposed to split-brain scenarios that generic high-availability tooling cannot reliably prevent. This paper addresses that omission head-on through a self-healing database fabric purpose-built for payment rail environments. The proposed autonomous resilience fabric architecture (ARFA) operates across three coordinated layers: a continuous monitoring layer that harvests telemetry from compute, storage, and network subsystems; a decision layer that fuses rule-based heuristics with an ensemble of isolation forests, recurrent neural networks, and gradient boosting classifiers to separate genuine fault conditions from transient noise; and a deterministic action layer that executes recovery procedures anchored to explicit settlement finality constraints. In fault injection trials covering node crashes, network partitions, replication lag, and performance degradation, the architecture cut average recovery times by 88% against manual baselines, restoring service in roughly 8 seconds rather than the 180 seconds that human-driven remediation typically requires. False positive rates held below 2% across all failure categories, and the system achieved a 98% recovery success rate. Taken together, these results make a practical case that autonomous resilience and regulatory compliance reinforce rather than conflict with each other when the regulatory constraints are designed in from the start.
Deploying deep learning models for autonomous driving on resource-constrained edge devices, such as the Raspberry Pi, presents significant challenges due to strict limitations on inference latency and memory capacity. To address these constraints, this study conducts a comprehensive comparative evaluation of lightweight convolutional neural networks (CNNs) optimized for dual-output regression of steering angle and driving speed. We benchmark a task-specific end-to-end baseline (NVIDIA CNN) against representative classification-oriented architectures—including MobileNet, ShuffleNet, EfficientNet, GhostNet, and SqueezeNet—all reformulated for this regression task. Experiments were conducted on a physical Raspberry Pi-based autonomous RC car platform to assess prediction accuracy, inference speed, and real-world closed-loop driving stability using quantitative metrics such as the normalized jerk ratio. Experimental results demonstrate a clear trade-off: while GhostNetV1 0.5x achieved the highest regression accuracy with a Total R2 score of 95.8% and MobileNetV1 recorded a competitive MAE of 1.95, they failed to provide stable control due to severe high-frequency steering jitter. Conversely, the NVIDIA CNN proved to be the most practical solution for general edge deployment, achieving the lowest inference latency of 61.1 ms (16.4 FPS) and a minimal memory footprint of 2.78 MB, ensuring stable autonomous navigation (1.50xjerk ratio). Furthermore, ShuffleNetV2 0.5x emerged as the superior architecture for trajectory precision, recording the lowest weighted MAE of 1.60. These findings underscore that theoretical accuracy does not guarantee real-world drivability on embedded systems, providing practical guidelines for hardware-aware model selection in edge-based autonomous driving.
This study investigates the interplay between learning engagement, interaction, and outcomes within the context of gamified massive open online courses (G-MOOCs). By synthesizing literature on MOOCs, gamification, and user engagement, the research identifies significant correlations among these variables. Utilizing a structural equation model partial least squares (SEM-PLS) approach, the study analyzes data from a survey of Bachelor of Computer Science students at a technical and vocational education and training (TVET) public university. Results indicate that both learning engagement and interaction significantly influence learning outcomes, with optimal results achieved when both factors are high. These findings highlight the potential of gamification to enhance educational experiences and suggest directions for future research in gamified learning environments.
K-means++ is developed to improve the performance of k-means when choosing a starting centroid. However, both algorithms in clustering still select an initial centroid randomly. Randomly selecting initial centroids has the potential to produce unstable clusters. This paper proposes a deterministic centroid initialization method called variance-k-means++, which utilizes statistical properties—mean and variance—to generate pseudo-centroids and derive initial centroids. The method aims to improve clustering stability and reduce the number of iterations. For the initial study, we used low-dimensional data to conduct the experiment series. Then, we employed two baseline methods for benchmarking, k-means and k-means++. The results show that variance-k-means++ outperformed the baseline method on average. Evaluating in Davies-Bouldin Index (DBI) and convergence analysis, we obtained DBI values at 0.756 and 0,771 for vertical and horizontal variance k-means++ with Iris dataset. At the same time, baseline methods have 0.802 and 0.830 for k-means++ and k-means, respectively. In convergence analysis, the results are 5.158 for vertical and 5.474 for horizontal, while baseline methods are 9.000 and 8.842. The primary contribution of this study lies in its achievement of minimizing the number of iterations while enhancing cluster stability.
Modern control systems must operate reliably in the presence of modeling uncertainties and external disturbances, conditions under which conventional fixed-gain controllers often exhibit performance degradation. This paper proposes a residual reinforcement learning framework for disturbance-resilient pitch-rate control of an aircraft longitudinal model. A classical proportional-integral-derivative (PID) controller is employed as a stabilizing baseline, while a deep deterministic policy gradient (DDPG) agent learns a bounded residual control signal to compensate for unmodeled dynamics and external perturbations. To promote favorable transient behavior, the learning process incorporates transient-aware and reference-model-based reward shaping, while actuator constraints are enforced within the environment dynamics. Simulation results demonstrate that the proposed residual controller achieves a superior balance between response speed, overshoot, and tracking accuracy compared with both the standalone PID controller and a pure DDPG-based controller. In particular, the residual architecture significantly reduces overshoot and tracking error while preserving fast transient response and providing robust disturbance rejection under large pitching moment disturbances. These results indicate that residual reinforcement learning offers a practical and effective approach for enhancing robustness and performance in safety-critical flight control applications.
Early detection of depression is crucial for minimizing its adverse effects on mental and physical health. Recent advancements in natural language processing facilitate the large-scale analysis of social media texts to identify depressive tendencies. Our study introduces a novel approach by integrating a genetic algorithm for hyperparameter tuning, optimizing the classification performance beyond conventional methods. We provide a comprehensive comparison of vectorization techniques, including term frequency-inverse document frequency (TF-IDF), Word2Vec, and a fine-tuned bidirectional encoder representation from transformers (BERT) model specifically adapted to our dataset. Using a dataset of 7,731 entries, we implemented standard pre-processing steps such as stop word removal and lemmatization before vectorizing the text. Five machine learning algorithms—decision tree, logistic regression, random forest, gradient boosting, and support vector machine—were evaluated, with hyperparameter tuning performed using a genetic algorithm. The highest accuracy (95.99%) and F1-score (95.91%) were achieved with the combination of fine-tuned BERT, support vector machine, and genetic algorithm optimization. This study demonstrates the advantages of integrating BERT fine-tuning with genetic optimization, outperforming traditional TF-IDF and Word2Vec approaches in depression detection.
Robust and precise finger gesture recognition using surface electromyography (sEMG) is essential for developing intuitive prosthetic control systems. However, sEMG signals are inherently stochastic and non-stationary, posing significant challenges for high-accuracy classification in fine-grained movements. This study proposes an optimized 1D convolutional neural network (1D-CNN) framework for classifying 20 distinct fine-grained finger gestures using raw sEMG data from an 8-channel wearable Myo Armband sensor. Unlike traditional methods that rely on manual feature engineering, the proposed 1D-CNN performs end-to-end learning to automatically extract temporal features. The research specifically investigates the impact of temporal windowing strategies, ranging from 400 to 750 ms, on model performance. Experimental results demonstrate that the optimized 1D-CNN achieves a peak test accuracy of 94.4% with a 550 ms window size, demonstrating the model’s robustness across complex gesture classes and significantly outperforming the baseline principal component analysis- support vector machine (PCA-SVM) method which only attained 73.0% accuracy. While the model achieved perfect classification (100%) for index, middle, and little finger movements, a performance drop was observed in thumb recognition (50%) due to muscular crosstalk from deeper anatomical layers. These findings indicate that the integration of optimized windowing and 1D-CNN architectures provides a highly reliable solution for complex large-scale gesture recognition, offering a robust foundation for the next generation of multi-functional prosthetic hands.
Radar-based gesture recognition has emerged as a robust alternative to vision-based systems, particularly in environments where lighting and privacy pose challenges. This study presents a simulation approach for recognizing hand gestures to control unmanned aerial vehicles (UAVs) using radar signals. Five discrete gestures, i.e., TakeOff, Land, MoveForward, TurnLeft, and stop, were defined and modeled in MATLAB to generate synthetic radar signals. From each sample, four time-frequency domain features were extracted: duration, maximum amplitude, dominant frequency, and root mean square (RMS). A dataset of 500 samples (100 per class) was classified using three supervised learning models: support vector machine (SVM), k-nearest neighbors (k-NN), and decision tree. The k-NN classifier achieved the highest accuracy of 96%, demonstrating the feasibility of lightweight classifiers for gesture recognition using low-complexity features. These results highlight the potential of radar-based interfaces to replace traditional remote controls in UAV operation. The proposed simulation framework contributes to the development of intuitive, non-contact human-machine interaction systems.
Wearable Internet of Things systems increasingly depend on inertial measurement units (IMUs) to capture human motion, yet continuous high-frequency sensing, on-device processing, and long-range (LoRa) communication impose significant energy and latency challenges for battery-powered devices. This study formulates a practical scheduling framework that optimizes IMU sampling, quaternion-based feature extraction, and transmission decisions within the wearable/LoRa architecture. The framework operates in discrete time windows of W=0.5−1 s, within which sensing, processing, and communication decisions are updated at the window level to balance energy consumption and responsiveness. The method models energy consumption, accuracy degradation at lower sampling rates, and communication constraints to define feasible operating modes and determine optimal configurations under varying activity levels. An empirical accuracy–frequency mapping and component-wise energy model support both offline optimization and lightweight online scheduling. The results show that the proposed framework can balance accuracy, responsiveness, and battery life by dynamically shifting between high-performance, balanced, and low-power surveillance states. This scheduling strategy extends operational lifetime while preserving motion-detection reliability and ensuring timely event transmission. The findings demonstrate the importance of energy-aware IMU management in long-range wearable systems and provide a foundation for adaptive sensing strategies in real-world deployments.
To enhance the performance of long range wide area network (LoRaWAN), a routing algorithm and a novel medium access control (MAC) layer protocol are required. In addition to addressing scalability and security issues, the protocol seeks to improve communication efficiency, dependability, and power consumption. It presents a dynamic routing method that reduces energy consumption by utilizing machine learning processes, adaptive routing tactics, and route optimization approaches. Simulations in a range of deployment situations are used to assess the suggested solutions. These results imply that the suggested protocol and routing scheme have the potential to greatly enhance the sustainability, energy efficiency, and performance of LoRaWAN-based Internet of Things networks. The effectiveness of the proposed solutions is evaluated through extensive simulations across diverse deployment scenarios. The results demonstrate that the proposed MAC protocol achieves a throughput of 350 bps, outperforming conventional protocols that typically reach only 220 bps. Latency is reduced to 50 ms from 85 ms, energy consumption is decreased to 2.5 joules from 4.5 joules, and the packet delivery ratio (PDR) is improved to 95%, compared to 75% in existing approaches. These findings highlight the potential of the proposed protocol and routing scheme to significantly enhance the performance, energy efficiency, and sustainability of LoRaWAN-based IoT networks.
