This paper introduces SEDM, a Student Engagement Detection Model that leverages computer vision and deep learning to monitor student engagement in both classroom and online environments. SEDM combines YOLOv9 for face detection, HopeNet for head pose estimation, and DeepFace for emotion analysis, and it uses temporal smoothing and engagement buffer to minimize noise. A peer-discussion mode is created to ensure collaborative interactions are not misclassified as disengagement. SEDM was evaluated in both classroom and online settings, achieving accuracies of 98.6% and 97.1% respectively, which outperform or match existing models that rely only on head pose detection. SEDM's precise measurement of student engagement enables educators to identify drop-offs and refine their teaching strategies accordingly. The model is adaptable to various classroom settings and provides a scalable, data-driven approach to improving student learning experiences.
Achieving more reliable communication necessitates precise beam management, a process often challenged by the dynamic and unpredictable nature of mobile environments. This study proposes an advanced prediction methodology that integrates a beam sweeping technique to acquire user prototype trajectory information. The method leverages received signal strength (RSS) and interference signal to noise ratio (SINR) metrics collected over a defined period to construct a prototype trajectory that reflects the behavior of user equipment (UE) over time. To enhance the accuracy of user movement predictions, two distinct Kalman filter approachesDthe Extended Kalman Filter (EKF) and the Unscented Kalman Filter (UKF) are applied to the prototype trajectory and data from proximity sensors. These filters are used to predict the user's position and derive an average path representing the movement from one point to another. This comparative analysis enables the evaluation of prediction accuracy and its implications for adaptive beam steering in future. The findings demonstrate the potential of this approach that enable us to dynamically direct beams toward the user, thereby enhancing connectivity and overall network performance in mmWave-based systems. Such advancements are instrumental in addressing the challenges posed by high mobility and intermittent connectivity in nextgeneration wireless communication networks.
As modern semiconductor devices have become increasingly complex, guaranteeing their functional correctness and security assurance has become more challenging. Assertion-based verification (ABV) is a typical method to detect functional and security issues at the RTL stage of the design process. Properties - often expressed as SystemVerilog Assertions (SVAs) - are constructed to formally specify a design’s expected behavior by enabling automatic checks during simulation and formal verification. Although writing assertions enhances semiconductor assurance before fabrication, it requires expertise and is a time-consuming process, motivating the development of automated methods to facilitate the entire procedure. This article explores the various methodologies and best practices to enhance the automation of the SVA generation and presents insights into the challenges and opportunities in this area.
The rapid evolution of cyber threats demands innovative solutions for real-time detection and mitigation. This paper presents a novel hybrid AI-powered framework for detecting cyberattacks by leveraging a multi-dimensional dataset comprising 40,000 network records. The dataset integrates key features, including IP addresses, protocols, anomaly scores, attack signatures, and geo-location data, providing a rich foundation for threat analysis. Our approach combines machine learning techniques with advanced feature engineering to identify patterns indicative of malicious activities. By employing supervised learning algorithms and anomaly detection models, the proposed framework achieves high accuracy in differentiating between benign and malicious traffic. Furthermore, the framework incorporates metadata analysis to enhance interpretability and facilitate informed decision-making. Experimental results demonstrate the model’s ability to detect diverse cyber threats, including fileless malware, DDoS attacks, and protocol-based exploits, with improved precision and reduced false positives compared to existing solutions. The findings highlight the effectiveness of leveraging multi-dimensional data and AI-driven methodologies for proactive cybersecurity. This study underscores the potential of integrating comprehensive datasets and advanced analytics to enhance cyber defense mechanisms. The proposed framework sets the stage for future research in real-time threat detection and response, contributing to the broader goal of securing digital ecosystems.
This paper presents a detailed performance evaluation of various power quality (PQ) monitoring devices utilizing the Hydro-Quebec benchmark and a combination of statistical analysis tests and machine learning techniques, specifically Partial Least Squares Discriminant Analysis (PLS-DA). The study examines six devices, assessing their compliance with IEC standards 61000-4-30, 61000-4-15, and 61000-4-7. The evaluation was conducted using a proprietary reference system developed by Hydro-Quebec/IREQ. The devices underwent rigorous testing, including stationary and variable frequency assessments, as well as event detection for voltage dips, swells, and interruptions. The combination of statistical tests and machine learning provided a robust analysis, identifying the most accurate and reliable devices, closely matching the Hydro-Quebec benchmark. This study highlights the importance of integrating statistical analysis with advanced machine learning techniques like PLS-DA to enhance the evaluation and selection process of PQ monitoring devices, ensuring compliance with industry standards and optimizing reliability at Hydro-Quebec.
In this paper, we introduce a framework for energyefficient fire detection by combining a simple color-based approach with a deep-learning-based model. We show how the two detectors can complement each other, achieving strong accuracy with reduced computational load. The experiments and results illustrate the performance in terms of detection speed, accuracy, and the feasibility of running multiple instances on shared hardware.
Immediate Feedback Assessment Technique (R) (IFAT (R)) provides opportunities for students to answer a multiple-choice question until the correct answer is revealed. This process is claimed to promote self-learning while testing, reduce test anxiety, and provide teachers with partial knowledge recognition. Scratch card-based IFAT (R) was developed by Epstein and made available to the teaching community through a commercial channel. Despite several advantages, grading using IFAT (R) cards for a final exam of a typical first-year course requires several manual hours. To ease this tedious process, this paper discusses an automated approach for grading exams conducted using scratch card-based IFAT. An algorithm that uses the OpenCV image processing library that can detect scratched, unscratched, and varying degrees of partially scratched boxes with high accuracy has been developed and implemented to create a logical representation of the IFAT (R) card. Our tool demonstrated a high degree of precision (> 99 %) for an assessment involving 39 cards of 10 questions, each with 5 scratch boxes.
This paper addresses the challenge of accurately predicting lane change maneuvers in autonomous driving while quantifying prediction uncertainty. We present an LSTM-based model that predicts three classes: lane keeping, left lane change, and right lane change. By considering the position, velocity, and acceleration of multiple vehicles, we enhance the accuracy of our predictions, achieving an overall accuracy of 97.10% with our base LSTM model. Our analysis of different uncertainty sources, including Gaussian noise injection, Dropout noise, and deep ensembles, reveals varying levels of uncertainty, with mean standard deviations ranging from 0.0084 to 0.1094. This highlights the model’s sensitivity to different types of perturbations and the importance of uncertainty quantification for robust lane change prediction. This contribution enables more informed decisionmaking in ADAS and paves the way for safer autonomous driving.
This paper deals with the integration of graphonbased connectivity models with state-space representations to examine the dynamic behavior of systems as the number of nodes increases. A particular focus is placed on analyzing various graphon functions, such as constant, step, exponential, Gaussian, and periodic, to investigate their influence on system connectivity, scalability, and stability. Furthermore, robust control methodologies, specifically H-infinity control, is employed to enhance system performance under disturbances and uncertainties in graphon-based networks. This study provides a comprehensive framework that combines theoretical insights and practical applications, offering new directions for the design and control of large-scale dynamic networks.
Transmitted data face many challenges during transmission to achieve data security. Securing transmitted multimedia data can be obtained through images and videos encryption for confidentiality purposes. Hackers usually try to implement several attacks including differential attack to get the original plaintext image/ video frame through revealing the encryption key. Evaluation metrics are applied to test the robustness of the applied encryption algorithm against various attacks. Metrics are based on statistical measures such as number of pixels change rate (NPCR). Image contrast is not sensitive while changing pixel intensity few values and the resulting image may still reveal the details of the original image. For example, applying encryption using Caesar Cipher with a small key value such as 1 or 2. Nevertheless, the NPCR of the resulted image is 100% which is the ultimate goal. This is a considerable flaw and resulted from the core function of the NPCR equation. This research proposes an enhancement to the NPCR metric based on the difference of intensity values introducing enhanced NPCR (ENPCR). Linear and non-linear core functions were proposed and tested instead of the current binary core function to provide a better NPCR performance. The results presented show that approaching non-linear core functions are more accurate than utilizing a linear function. On the other hand, the processing time and complexity of a linear function makes it preferable.
Digital technologies are rapidly transforming healthcare by integrating networked devices, advanced analytics, telemedicine, and electronic health records. Intelligent systems have evolved from early rule-based medical expert systems to modern AI-driven Intelligent agent-based systems (software systems that autonomously reason, learn, and interact). These systems not only enhance operational efficiency and diagnostic accuracy but also support personalized and proactive patient care. This paper presents a historical overview of intelligent systems in healthcare, compares past and present systems through two detailed tables, and introduces current research and development that focuses on clinical decision making (CDM) in the age of AI. The evolution of intelligent systems in digital health from early medical expert systems like MYCIN to modern data-driven platforms illustrates significant technological progress. Contemporary intelligent systems demonstrate increased adaptability, scalability, and integration with clinical workflows. The introduction of advanced AI frameworks and the application of Explainable AI (XAI) techniques have further enhanced transparency and clinician trust. In particular, we discuss an AI-agent framework that integrates machine learning models with XAI techniques and human feedback to automate medical data analysis and report generation. We conclude with a discussion of the advantages, challenges, and future promise of these systems, emphasizing the importance of ethical leadership, effective security mechanisms and stronger cross-disciplinary collaboration.
Reducing cold-starts while improving resource utilization has been a persistent challenge in serverless computing services. This paper proposes a self-adaptive keep-alive window control mechanism that dynamically adjusts window sizes for different applications based on varying incoming request patterns. Simulation results using real Microsoft Azure Functions invocation trace demonstrate that, compared with the fixed-size window mechanism, the proposed approach significantly reduces cold-starts and decreases the request rejections. Furthermore, the new mechanism is straightforward to implement, lightweight, and robust to the initial window size configuration.
Tactile texture recognition is a crucial skill for humans, but it is challenging to emulate in robots. This is mainly due to the complexities of detecting and analyzing textures on uneven or irregular surfaces. This paper introduces a novel approach that leverages haptic surface reconstruction combined with reinforcement learning (RL) to enhance robotic tactile texture recognition. Our method involves an initial haptic surface reconstruction refined through RL to estimate contact points and surface normals accurately. A robot equipped with a multimodal tactile sensing module then uses this information to explore surfaces, collecting data instrumental for texture identification. For the texture recognition phase, we employ two types of recurrent neural networks (RNNs): one with a feature encoder and one without. Our findings demonstrate that the RL-refined trajectories significantly improve classification accuracy from 63.46% to 89.58%. Combining haptic feedback and reinforcement learning significantly improves robotic texture classification on uneven surfaces, irrespective of the classifiers employed.
This study introduces a novel approach to enhancing microgrid operations by incorporating demand response (DR) strategies with a hybrid optimization model that combines the strengths of the Imperialist Competitive Algorithm (ICA) and Particle Swarm Optimization (PSO). The approach aims to improve energy management, decrease the losses, stabilize voltage and frequency, and achieve a balanced alignment between energy supply and demand. A notable aspect of this work is the use of the IEEE 37-Bus test system, which provides a realistic environment to evaluate the hybrid ICA-PSO method. The findings show that this approach significantly outperforms traditional methods, delivering superior results in power losses reduction, voltage regulation, and frequency stabilization. These outcomes highlight the method’s practicality for real-world microgrid applications. The paper further explores the unique benefits of ICA and PSO, explaining the rationale for their integration and emphasizing how demand response enhances microgrid efficiency. It also suggests areas for future investigation, such as addressing uncertainties and integrating real-time data, to improve the adaptability and reliability of the proposed method in complex operational scenarios. The problem modeling has been done using MATLAB software. The results proved the effectiveness of the proposed model.
The use of satellite images in agriculture has always been limited due to the low resolution of the image acquisition equipment and the associated costs of their launches. Super Resolution (SR) is a technique that allows for image enlargement through software, which can aid in the efficient use of satellite images. In this work, we investigate the application of a generative SR model for enlarging 4 times medium-resolution satellite images (6m/pixel) in the agricultural context. The Real-ESRGAN model, an extension of ESRGAN (Enhanced Super-Resolution Generative Adversarial Networks), models the degradation process by considering higher-order terms, which better approximates atmospheric effects. The results obtained indicate perception gains ranging from 35.7% to 61.5% with the fine-tuning of the base model, compared to traditional models, as well as an equally good numerical approximation in the MSE, PSNR and SSIM metrics.
In recent years, object detection and recognition algorithms have achieved acceptable performance in normal weather conditions. However, these algorithms fail to provide the same results in adverse weather conditions which hampers autonomous vehicles widespread utilization. As a result, analyzing the performance of autonomous vehicle sensors in severe environments is very important. This paper evaluates the performance of camera sensors under diverse weather scenarios using the BDD100k dataset. Specifically, we compare the detection accuracy of models trained and tested on individual weather conditions, such as rain, snow, and clear weather including different times like day and night. In addition, a general model is trained on aggregated data to compare the performance of models specifically trained for a weather condition with a generalized model. Metrics such as mean Average Precision, Recall, and Precision are calculated and assessed for 3 types of object groups categorized as small, medium, and large objects. The results demonstrate that weather-specific models perform better in their respective weather conditions, particularly for small and medium objects. However, the general model provides consistent and robust performance across all weather scenarios, making it suitable for general-purpose applications. Additionally, the paper highlights the trade-offs between model generalization and specialization, time of the day, and the impact of object size on detection accuracy. These findings contribute to the development of more reliable and adaptable object detection systems for adverse weather environments.
Optimizing the operation of heating, ventilation, and air Conditioning (HVAC) systems is crucial for improving energy efficiency, enhancing user comfort, and maximizing system performance. This paper proposes a machine learning (ML)-based method that leverages environmental sensor data to predict high room occupancy, enabling more efficient HVAC operation. It integrates multiple ML techniques to enhance energy efficiency by minimizing HVAC operation in unoccupied spaces while maintaining occupant comfort. The methodology involves using sensor readings, such as temperature, humidity, light, and CO2 levels to improve occupancy detection accuracy. It exploits advanced techniques such as synthetic minority over-sampling technique (SMOTE)-Tomek for data balancing, eXtreme Gradient Boosting for robust classification, multi-layer artificial neural networks (ML-ANN) for capturing nonlinear relationships, and long short-term memory (LSTM) models for time-series forecasting. To further improve prediction accuracy, a probabilistic voting classifier is implemented by combining the predictions of the above individual models, leveraging a soft voting mechanism to generate more reliable occupancy forecasts. The results show that the proposed solution achieves highly competitive performance on a publicly available dataset.
In this research, we present SecPassGAN, a novel AI-driven framework designed for generating robust and unpredictable passwords. Our approach merges the generative capabilities of LSTM networks with the refinement and scrutiny of GANbased discrimination to craft passwords that are both secure and varied. The framework also features a built-in mechanism for realtime password strength evaluation and incorporates adversarial techniques to filter out weak or easily guessable patterns. Through extensive evaluation, SecPassGAN demonstrates a clear advantage over existing approaches, offering enhanced resistance to cracking attempts while preserving user accessibility. This work offers a practical perspective on leveraging AI to strengthen modern authentication systems.
This paper introduces a novel approach to manipulator end-effector path planning and obstacle avoidance by integrating Artificial Potential Fields with Reinforcement Learning. Unlike traditional methods where the agent learns to select the optimal action at each time step, our approach enables the agent to determine the optimal structure of the artificial potential field. This allows the manipulator to reach the goal without collisions while minimizing joint acceleration. Given any environment—defined by the robot’s starting position, goal, and obstacle location—the agent can rapidly construct an optimal potential field to guide the motion. Simulations using an experimentally validated model demonstrate the effectiveness of the proposed method. This strategy holds significant potential for path planning in dynamic environments.
Hybrid renewable energy systems (HRES), integrating photovoltaic (PV) farms, wind turbines, and battery storage, are revolutionizing modern power generation by offering sustainable and reliable energy solutions. However, their inherent complexity poses significant challenges in fault detection and protection, as traditional methods, such as over-current relays, often fall short, leading to system disruptions. This paper proposes an ML-based fault detection framework utilizing advanced machine learning algorithms and minimal hardware-one voltage sensor on the DC link and one per AC phase. A simulated hybrid system comprising a 500 kW PV farm, 250 kW wind farm, and 200 kW battery storage, connected to a 25 kV grid, evaluates the framework's efficacy. The results demonstrate 98.14% fault detection accuracy on the AC side and flawless accuracy on the DC side. By leveraging the robustness and precision of machine learning techniques, this framework enhances the reliability, scalability, and operational efficiency of HRES, ensuring robust fault management in modern energy systems.
