École Marocaine des Sciences de l'Ingénieur

Due to the diverse storage and performance requirements of modern data-driven environments, choosing between relational and non-relational database systems is a complex decision, as each type offers distinct advantages and limitations depending on the application. A single database system therefore fails to deliver the necessary balance of latency, scalability, and resiliency expected by contemporary workloads. One effective solution is polyglot service persistence, an approach where multiple database engines coexist within the same system to leverage the individual strengths of each. In this paper, we introduce the development of a hybrid platform that unifies several database types—together with a distributed file store—through a micro-services architecture. This strategy yields an adaptable, horizontally scalable solution capable of satisfying heterogeneous data requirements while simplifying operational overhead. Our methodology centres on a framework that exposes a unified interface to all underlying datastores, so end users can manage relational, document, and file-based data without installing or configuring each engine individually; connection logic, driver management, and schema coordination are fully abstracted. In addition, it synthesises fully documented REST APIs for every generated schema, enabling seamless CRUD operations across all underlying stores. The hybrid system is engineered to optimise every database model’s forte: transactional consistency in relational stores, elastic growth for schema-less datasets in document stores, and built-in redundancy plus fault tolerance for large files in distributed storage. Although comprehensive benchmarks are scheduled for future work, preliminary integration tests confirm that the framework streamlines deployment workflows and accelerates application development by eliminating per-database setup tasks and management.

The proliferation of embedded computing systems and efficient deep learning models has opened up new possibilities for deploying intelligent surveillance systems in constrained environments. However, real-time detection, contextual understanding and explainability within the limitations of edge hardware remains a major challenge. Traditional license plate recognition solutions often rely on cloud infrastructure, leading to latency, dependency and privacy issues. In this work, we present a AI system designed for the recognition and interpretation of license plates of suspicious vehicles. The architecture integrates a YOLOv8-based object detector and EasyOCR for multilingual text extraction, with a lightweight, locally hosted language model (LLM) that generates human-readable justifications for matches identified in a preloaded database. To ensure ease of use in the field, a Streamlit-based dashboard provides real-time visualization, and interactive reporting. Our design eliminates the need for external computation offering a complete solution that combines perception and reasoning on a single Raspberry Pi device. Experimental results confirm the system’s ability to provide low-latency inference, accurate recognition and contextual explanations, establishing its suitability for deployment in security-critical environments such as border checkpoints, government facilities and smart city infrastructure.

Knowledge acquisition in engineering education, particularly in hands-on disciplines like electrical drawing, faces challenges such as prolonged learning durations, low engagement, and limited real-time competency assessment. Traditional methods rely on physical apparatus and static experimentation, which often fail to address these issues effectively. This paper proposes an XR pedagogical framework integrating augmented reality (AR) with the Conceive-Design-Implement-Operate (CDIO) standards to enhance learning experiences. The framework merges constructivist learning theories, social collaboration, and competency-based assessment to create immersive industrial training. To demonstrate its application, we developed a Minimum Viable Product (MVP) AR prototype. It features interactive 3D circuit simulations aligned with CDIO Standard 7 (experiential learning), self-assessment quizzes for competency tracking (Standard 11), and real-time collaborative troubleshooting workspaces (Standard 8). By simulating electrical drawing tasks in a risk-free virtual environment, the prototype enables learners to practice complex concepts, receive immediate feedback, and refine problem-solving skills—key competencies for modern engineering practice. The framework empowers educators to design AR-enhanced curricula bridging theoretical knowledge with industry-driven skill development. Preliminary insights suggest that AR integration enhances engagement, accelerates skill acquisition, and fosters critical thinking. The prototype’s interactive simulations reduce cognitive load by visualizing abstract electrical principles spatially, while collaborative modules mirror real-world engineering teamwork. This work contributes to the evolving discourse on immersive learning by aligning XR technologies with globally recognized educational standards.

Abstract Leaf blight is a prevalent and destructive disease that significantly impacts crop yield and agricultural productivity worldwide. This paper explores an advanced approach to leaf blight detection by leveraging transfer learning techniques to enhance classification accuracy. We employed AlexNet, MobileNet-V2, and VGG16Net, pre-trained on extensive image datasets and fine-tuned using a curated dataset comprising PlantVillage images and real-world agricultural photographs to ensure robustness against diverse environmental conditions. The models’ performance was rigorously evaluated, with AlexNet achieving an accuracy of 45.06%, MobileNet-V2 reaching 94.24%, and VGG16Net demonstrating the highest accuracy at 97.8%. The fine-tuning process incorporated data augmentation techniques, including age and color classification, to mitigate the phenotypic variability inherent in agricultural environments. This study underscores the effectiveness of transfer learning in automated leaf blight detection, highlighting the superior performance of VGG16Net in distinguishing diseased from healthy leaves. The findings contribute to advancing precision agriculture by providing scalable, data-driven solutions for early disease detection, thereby supporting proactive crop management and improving food security. relevant articles in literature searches, great care should be taken in constructing both.

This paper presents an enhanced algorithm for pattern recognition using an partial overlapping method of a 4x4 block and exact computation moments: Legendre and Gegenbauer. By employing a limited number of moments (order 0), the methods achieve high accuracy while maintaining a low computational cost. It also provides better recognition rates compared to conventional methods based on calculating moments on the entire image, which involve higher orders. Experimental results highlights the superior performance of the proposed method, particularly with Legendre moments, compared to existing techniques.

The integration of renewable energy sources, particularly photovoltaic (PV) and wind energy, presents significant challenges due to their inherent variability and nonlinear dynamics. This chapter introduces a novel hybrid energy system that combines PV and wind power, managed by an advanced control strategy that integrates model predictive control (MPC) with particle swarm optimization (PSO). The MPC&PSO approach is designed to optimize power output and ensure the stability of the system under dynamic environmental conditions.

In this present study, two mathematical models derived from the single-diode model of photovoltaic (PV) cells and modules are evaluated. Moreover, the two mathematical approaches are investigated for the reduced single-diode model and the reduced single-diode combined to Lambert W-function model. The Levenberg-Marquardt optimization algorithm is chosen due to its effectiveness in addressing the complex optimization problems encountered in PV cells modeling. The comparative analysis evaluated the accuracy and computational efficiency of the both models for three PV datasets obtained from the literature. The results demonstrated that the combination of the reduced single-diode with Lambert W-function model achieves a minimum RMSE value of 3.1627×10−4, compared to the reduced single-diode model one of 3.2066×10−4.