Informatique

Browse

Recent Submissions

Now showing 1 - 20 of 304
  • Item
    An Explainable Self-Supervised Learning-Based Ensemble Framework for Audio Deepfake Detection
    (Faculty of Sciences, 2026) RAHMANI ,Aya; TAGHANE ,Nedjoua; HAZMOUNE, Samira
    Recent advances in artificial intelligence have enabled the generation of highly realistic audio deepfakes through speech synthesis and voice conversion technologies. Although these technologies offer numerous beneficial applications, they also pose serious security, privacy, and misinformation risks, making reliable audio deepfake detection an increasingly important research challenge. To address this issue, this work proposes an audio deepfake detection framework that combines self-supervised learning (SSL) and ensemble learning. Three state-of-the-art SSL models: Wav2Vec 2.0, HuBERT, andWavLM, were fine-tuned to distinguish real from fake speech. Several ensemble strategies were investigated to exploit the complementary strengths of the individual models. Furthermore, the Gradient Saliency explainable AI technique was employed to enhance model interpretability and provide insights into the features influencing the detection decisions. The proposed framework was evaluated on the ASVspoof 2019 Logical Access dataset and further assessed on the DeepVoice dataset to examine its generalization capability. Experimental results demonstrated the effectiveness of SSLbased speech representations for audio deepfake detection. Among all evaluated methods, the stacking ensemble achieved the best performance, reaching an accuracy of 99.15%, an F1-score of 99.14%, and an EER of 0.38% on the ASVspoof 2019 LA dataset. Cross-dataset experiments on the DeepVoice dataset yielded an accuracy of 81.25% and an F1-score of 88.89%, confirming the robustness and generalization capability of the proposed approach. These findings show that combining self-supervised speech representations with ensemble learning provides an effective, robust, and interpretable solution for audio deepfake detection.
  • Item
    AI-POWERED WEB APPLICATION FOR HOTEL RESERVATION SYSTEM
    (Faculty of Science, 2026) CHORFI, DHIA EDDINE; LAHSASNA ,ADEL
    The digital transformation of the hospitality sector has made online reservation systems central to the way hotels operate, while artificial intelligence (AI) has become a decisive competitive factor through dynamic pricing and the automated analysis of guest opinion. These intelligent capabilities, however, remain largely fragmented, costly, or opaque, and so out of reach for the independent hotel owner. This thesis addresses that gap through the design and implementation of Hotelify, a complete, full-stack web application for hotel reservation in which AI is a built-in capability. The system unites the conventional functions of reservation and hotel management with two integrated models, namely a LightGBM model that predicts and recommends room prices in support of dynamic pricing, and a fine-tuned DistilBERT model, trained with focal loss to counter class imbalance, that analyses the sentiment of guest reviews. These capabilities are organised around three roles, those of the guest, the owner, and the administrator, and are presented transparently, so that the owner retains authority over the decisions that the system informs. The application was built with a React frontend, a Node.js and Express backend, a separate FastAPI AI service, and a MongoDB database, orchestrated with Docker Compose. The evaluation shows the system to be functionally complete and soundly engineered; the sentiment analysis model attains strong and balanced performance, while the price prediction model, despite strong offline metrics, is affected by a target leakage limitation that is identified as the foremost subject of future work. Overall, the work demonstrates how intelligent and explainable decision support can be brought, within a single integrated platform, into the service of the guests, owners, and administrators who together make up a hotel reservation platform.
  • Item
    Development of a Low-Cost Edge AI Pipeline for Waste Object Characterization and Mass Estimation
    (Faculty of Sciences, 2026) Karim ,Aouaouda; Adel, Lahsasna
    Waste characterization is an important step in modern waste management, recycling, and environmental monitoring. The central objective of this thesis is simple: we try to estimate the mass of waste objects using nothing more than a low-cost RGB camera. This objective is difficult because mass is not directly visible in an image; it must be inferred from object category, mask geometry, depth, material behavior, density, deformation, and calibration. This thesis presents a first full version of a low-cost Edge AI pipeline for waste object characterization and mass estimation. The system uses an RGB camera and an NVIDIA Jetson Nano target platform to perform object segmentation, monocular depth estimation, metric feature extraction, volume approximation, and hybrid mass prediction. We followed a practical machine-learning engineering workflow. A segmentation dataset was prepared by combining and adapting TACO, AquaTrash, and RealWaste. Because several sources did not provide instance masks, the work included manual annotation, detection-assisted annotation, Segment Anything based mask generation, and CVAT correction. In total, 7,648 images were annotated during the wider dataset preparation effort, and 6,460 images with 9,762 instances were retained in the final segmentation training manifest. An Ultralytics YOLO26m-seg model was fine-tuned with Optuna-assisted hyperparameter search, ZenML pipeline organization, MLflow registration, and ONNX export. The final evaluated model achieved mask mAP@50:95 of 0.5453 on validation and 0.5016 on test. For mass estimation, a hybrid residual model was trained on 745 object-level samples. The selected Extra Trees residual model reduced test MAE from 44.31 g for the physics baseline to 23.18 g, corresponding to a 47.69% improvement. The thesis documents the mathematical foundation, system design, training pipelines, deployment-oriented architecture, and remaining limitations. The results show that the proposed pipeline is technically feasible as a low-cost Edge AI approach, while also showing that robust mass estimation requires stronger physical calibration, real measured masses, and Jetson-specific latency validation in future work.
  • Item
    Évaluation de la Qualité Physico-chimique et Organoleptique du Colostrum Bovin et Caprin
    (Faculté des sciences, 2026) AZZOUZ, Ikram; AMROUS ,Sarra; BERROUAGE, Chaima; BOUROUBA ,Amina; NAKIB.L
    La présente étude avait pour objectif de comparer les caractéristiques physicochimiques et organoleptiques du colostrum bovin et caprin, d’évaluer leur aptitude à la transformation fromagère, ainsi que de caractériser les propriétés physicochimiques et sensorielles des fromages frais obtenus. À cette fin, trois échantillons de colostrum bovin et trois échantillons de colostrum caprin ont été prélevés dans des exploitations d’élevage situées dans la wilaya de Skikda. Les analyses physicochimiques ont porté sur la détermination du pH, de l’acidité titrable, de la densité, de la matière sèche, de la matière grasse et de la teneur en protéines. Les résultats ont montré que le colostrum bovin présentait un pH de 5,96 à 6,28, une acidité de 28 à 36 °D, une densité de 1,0959 à 1,2312, une matière sèche de 13,27 à 20,75 g/100 g, une matière grasse de 1,5 à 4,4 g/100 g et une teneur en protéines de 84,6 à 128,7 g/L. Le colostrum caprin présentait quant à lui un pH de 6,01 à 6,39, une acidité de 25 à 70 °D, une densité de 1,0017 à 1,1342, une matière sèche de 16,10 à 33,35 g/100 g, une matière grasse de 7 à 8 g/100 g et une teneur en protéines de 165,4 à 187,5 g/L. Globalement, le colostrum caprin était plus riche en matière sèche, en matière grasse et en protéines que le colostrum bovin. Les deux colostrums ont ensuite été transformés en fromage frais, dont les propriétés physicochimiques et organoleptiques ont été évaluées. Le fromage issu du colostrum caprin présentait un pH de 4,69 ± 0,02, une acidité de 0,60 ± 0,00 %, une teneur en protéines de 111,7 ± 0,14 g/L, une matière grasse de 21 g/100 g et une matière sèche de 49,86 ± 2,72 g/100 g, contre respectivement 4,26 ± 0,05, 0,85 ± 0,07 %, 62,6 ± 0,28 g/L, 8 g/100 g et 23,20 ± 1,44 g/100 g pour le fromage obtenu à partir du colostrum bovin. L’évaluation sensorielle a révélé une bonne acceptabilité des deux fromages, le fromage élaboré à partir du colostrum caprin ayant obtenu une meilleure appréciation globale (7,18 ± 1,56) que le fromage bovin (6,18 ± 2,14). Cette étude montre que les colostrums bovin et caprin présentent un potentiel intéressant pour la fabrication de fromage frais et constituent des matières premières prometteuses pour la valorisation des ressources laitières. Les résultats obtenus contribuent à une meilleure connaissance de leurs propriétés technologiques et ouvrent des perspectives pour le développement de nouveaux produits laitiers à forte valeur ajoutée
  • Item
    Design and Implementation of a RAG-Based Medical Report Data Extractor API
    (Faculty of Sciences, 2026) BOUCHERBAT ,Mohamed Elhacene; OUAZETA Ouail ,Chams Eddine; LAHSASNA ,Adel
    The digitalization of modern healthcare has generated vast amounts of unstructured clinical data, trapping valuable patient history within free-text medical reports. While Large Language Models (LLMs) offer unprecedented capabilities to analyze and synthesize this information, their direct application in clinical settings is severely limited by the risk of artificial hallucinations and strict patient data privacy regulations. To address this critical gap, this thesis presents the design, implementation, and evaluation of a privacy-preserving Retrieval-Augmented Generation (RAG) system tailored specifically for the medical domain. We propose a secure, multi-tier architecture that automatically ingests heterogeneous medical documents, extracts key clinical entities, and transforms unstructured text into highly searchable, localized vector representations. To guarantee data sovereignty and patient confidentiality, the system enforces strict mathematical isolation of patient records and employs an automated de-identification module to mask sensitive information before interacting with external generative models. Empirical evaluation demonstrates that the proposed architecture successfully mitigates LLM hallucinations by forcing the AI to generate responses strictly grounded in retrieved, patientspecific evidence. Ultimately, this research proves that integrating domain-specific natural language processing with secure RAG frameworks can safely unlock the value of unstructured medical data, providing clinicians with a reliable, privacy-compliant tool to support medical decision-making.
  • Item
    CodeWatchAI Based on Agentic AI Architecture
    (Faculty of Sciences, 2026) BOUKALKOUL ,Sarra; CHAOUCH ,Malak Fatima Zohra; Bounour, Fatima Zohra; Bougamouza ,Fateh
    The rapid growth of software development projects and the increasing complexity of modern codebases have made code review an essential activity for ensuring software quality, maintainability, and security. However, traditional manual code review processes remain time-consuming, inconsistent, and highly dependent on the availability and expertise of senior developers. Existing automated solutions often focus on isolated aspects of code analysis and lack a comprehensive understanding of the structural impact of code modifications. This project presents the design and development of CodeWatchAI, an AI-powered automated code review system based on the Agentic AI paradigm. The proposed solution combines multiple intelligent agents specialized in semantic analysis, security assessment, and syntax verification, operating in parallel to provide a comprehensive evaluation of Pull Requests. The system also introduces a Call Graph analysis mechanism to identify dependencies and assess the impact of code changes beyond simple file-level modifications. The proposed architecture relies on GitHub and GitLab webhooks for event triggering, n8n for workflow orchestration, PostgreSQL for data management, and Large Language Models (LLMs) for intelligent code analysis and decision support. The generated reports include detailed findings, risk classification levels, and merge recommendations designed to assist development teams and Team Leads in making informed decisions. The system is further supported by a management dashboard providing realtime visibility into review activity and system performance. The project was developed following a modular and scalable architecture suitable for real-world deployment and future commercialization as a Software-as-a-Service (SaaS) platform. Comparative analysis with existing solutions such as CodeRabbit, GitHub Copilot Code Review, Anthropic Code Review, and Qodo Merge demonstrates the added value of integrating Call Graph analysis, multi-agent collaboration, and decision-oriented reporting within a unified system. The results demonstrate that the proposed approach effectively improves review efficiency, reduces manual workload, strengthens security vulnerability detection, and provides more reliable merge decision support validated through comparative analysis and a functional prototype deployed on a real development workflow. Keywords: Agentic AI, Code Review, Multi-Agent Systems, Large Language Models, Call Graph Analysis, GitHub, GitLab, DevOps, Workflow Automation, Software Quality.
  • Item
    Segmentation et Caractérisation des Gliomes par IRM : U-Net/VGG16 et Clustering Radiomique Non Supervisé
    (Faculté des Sciences, 2026) BOULGUEMH ,Wafa; FILALI ,Radja; MAGROUNE.H
    Ce travail propose un système hybride de segmentation et de caractérisation des gliomes à partir d’images IRM T1 pondérées. L’architecture U-Net, couplée à un encodeur VGG16 pré-entraîné sur ImageNet, est entraînée sur 766 images du dataset LGG TCIA, atteignant un coefficient de Dice de 0,9152, un Jaccard de 0,8532 et une accuracy de 0,9972. Dix-huit caractéristiques radiomiques sont extraites, puis soumises à une réduction PCA et à un clustering GMM, révélant deux profils : le profil A (68,5 %, 510 tumeurs, surface moyenne 597 px) et le profil B (31,5 %, 234 tumeurs, surface moyenne 1595 px), avec un score de Silhouette de 0,4347. Malgré ses performances, l’approche reste limitée par l’usage d’une modalité unique (T1), l’absence de soustraction osseuse et le manque de validation clinique externe, ouvrant des perspectives vers des stratégies multimodales.
  • Item
    Étude comparative entre les approches d'apprentissage classique (SVM) et d'apprentissage profond (CNN) pour la classification des pathologies cérébrales par IRM
    (Faculté des Sciences, 2026) BOUTRIA, Manar; Azzouz, Daouia
    Les maladies cérébrales telles que la sclérose en plaques, la maladie d’Alzheimer et les accidents vasculaires cérébraux constituent aujourd’hui un enjeu majeur de santé publique. Elles affectent fortement la qualité de vie des patients et représentent une charge importante pour les systèmes de santé. Dans ce contexte, l’imagerie par résonance magnétique (IRM) joue un rôle essentiel dans l’exploration non invasive du cerveau grâce à sa capacité à fournir des images détaillées des structures anatomiques et des anomalies pathologiques. Cependant, l’analyse manuelle des images IRM reste une tâche complexe, longue et parfois sujette à des différences d’interprétation entre spécialistes. Avec les avancées récentes de l’intelligence artificielle, plusieurs méthodes automatiques ont été développées afin d’assister les médecins dans le diagnostic des maladies neurologiques. Ce mémoire s’intéresse à une tâche de classification automatique d’images IRM cérébrales en quatre catégories : cerveau Normal (sain / contrôle), sclérose en plaques (MS), maladie d’Alzheimer et accident vasculaire cérébral (AVC / Stroke), à partir d’images en coupe 2D. L’objectif principal est de comparer, les performances d’un réseau de neurones convolutif (CNN) fondé sur MobileNetV2 avec transfert d’apprentissage et fine-tuning, à celles d’une machine à vecteurs de support (SVM) à noyau gaussien (RBF) alimentée par des descripteurs extraits du même backbone convolutionnel. Le travail expérimental repose sur un pipeline de traitement, comprenant la collecte et l’équilibrage des données avec 380 images par classe, soit un total de 1 520 images. Des techniques d’augmentation de données, réalisées en partie hors ligne. Les deux approches utilisent le même prétraitement : redimensionnement des images en 160×160 pixels, normalisation via la fonction preprocess_input de MobileNetV2, et partitionnement 80% / 20% en ensembles d’entraînement et de validation, avec graine aléatoire fixée pour assurer la reproductibilité. Le modèle CNN a été entraîné en deux étapes : une première phase de gel du backbone suivie d’un fine-tuning partiel. Quant au modèle SVM, il exploite des vecteurs de caractéristiques extraits par MobileNetV2 figé, suivie d’une normalisation (StandardScaler) et d’une réduction de dimension par ACP (PCA) à 128 composantes. Les performances des modèles sont évaluées à l’aide de plusieurs métriques classiques en classification supervisée telles que l’accuracy globale, la précision, le rappel, le F1-score macro et par classe et les matrices de confusion 4×4, ainsi que les courbes ROC et les scores AUC sont également utilisés pour l’évaluation du CNN. Les résultats chiffrés finaux seront présentés après exécution complète des expérimentations. Toutefois, les travaux récents montrent généralement que les modèles CNN, offrent de très bonnes performances dans les tâches de classification d’images médicales, tandis que les SVM restent efficaces lorsque les caractéristiques extraites sont suffisamment pertinentes. Ce travail propose ainsi une comparaison rigoureuse entre deux approches largement utilisées en intelligence artificielle appliquée à l’imagerie médicale, discute les conditions dans lesquelles le CNN surpasse le SVM. Il met également en évidence certaines limites liées à la taille du jeu de données, au choix des paramètres et aux biais potentiels, et ouvre des perspectives vers l’utilisation de données 3D, des modèles plus récents et l’intégration de techniques d’explicabilité.
  • Item
    Smart Iris Biometric Identification System for Healthcare Management -EYEID
    (Faculty of Sciences, 2026) Chaima ,Messikh; Houda ,Hamrouche; Hanene ,Bourriche
    In contemporary institutional environments, maintaining secure, non-repudiable user authentication mechanisms is paramount to protecting highly sensitive database records. Traditional identity verification frameworks, which rely on physical tokens, knowledge-based passwords, or complex deep learning architectures, consistently expose systems to operational latencies, high hardware costs, and severe security vulnerabilities. To resolve these challenges, this thesis proposes and implements a lightweight, high-performance biometric identification system leveraging deterministic local feature extraction from human ocular tissue. The developed framework utilizes automated circular boundary segmentation to isolate the iris stroma, followed by the execution of the Oriented FAST and Rotated BRIEF (ORB) algorithm to extract invariant identity templates without the computational overhead of offline network training. Matching operations are conducted on-the-fly using bitwise logical operators to evaluate Hamming distance metrics against a structured local relational database. Empirical validation confirms that the proposed architecture compresses the entire computational pipeline spanning from raw optical acquisition to complete profile rendering into an exceptional sub-second execution latency of less than 0.5 seconds under a strict verification threshold (RECOGNITION THRESHOLD = 0.45). These results demonstrate that high-precision identity verification can be seamlessly achieved on standard low-cost workstations, eliminating system bottlenecks and proving the system’s practical readiness for secure, high-throughput institutional deployment.
  • Item
    Design and development of an integrated smart system to help Type 1 Diabetes patients in Algeria
    (Faculty of Sciences, 2026) Chemaa ,Sarra; Moussaoui, Hadil; Talhi ,Hadjer; Bougamouza ,Fateh
    Type 1 Diabetes (T1D) is a chronic autoimmune disease requiring lifelong insulin therapy and precise glycemic control to avoid severe complications such as Hypoglycemia and Diabetic Ketoacidosis. One of the main challenges in daily T1D management is determining the appropriate pre-meal insulin dose, which depends on accurate carbohydrate estimation combined with personalized clinical parameters such as the Insulin-to-Carbohydrate Ratio (ICR) and the Insulin Sensitivity Factor (ISF). However, manual carbohydrate counting remains time-consuming, error-prone, and difficult to apply consistently in everyday life. In this work, we propose DiabyLife, an AI-based mobile application designed to assist patients with T1D in estimating their pre-meal insulin dose directly from a photograph of their meal. The system relies on a DL model based on the MobileNetV2 architecture, trained using transfer learning to recognize food items and estimate their carbohydrate content. The estimated carbohydrates are then combined with the patient's personalized medical parameters to automatically compute the recommended rapid-acting insulin dose. Developed using Flutter and integrated with Firebase cloud services, the application also offers additional features such as blood glucose monitoring, meal history tracking, reminder management, and secure storage of medical data. Experimental results show that the optimized model achieved a validation accuracy of 95.4%, demonstrating the feasibility of using computer vision and DL to simplify carbohydrate counting and reduce insulin dosing errors. This study highlights the potential of combining AI with mobile health applications to improve the daily self-management of T1D and to support patients in achieving better glycemic control.
  • Item
    A Comparatif Study Of Data Storage In Cloud Computing And Blockchain Technologie
    (Faculty of Science, 2026) -KEDADRA ,Selma; LECHEHEB ,Naima; ALI GUECHI ,Farida
    Cloud Computing and Blockchain are two key technologies in modern information systems. While Cloud Computing provides scalable and flexible resources for data storage and processing, Blockchain offers a decentralized and secure environment that ensures data integrity and transparency. This dissertation presents a comparative study between the two technologies with a focus on security aspects. A practical application was developed using MongoDB Atlas and BigchainDB to evaluate their performance and security. The results show that Cloud Computing offers greater flexibility and efficiency, whereas Blockchain provides enhanced mechanisms for ensuring data integrity, traceability, and transparency. The study highlights the strengths and limitations of both technologies and demonstrates their potential contribution to building secure and reliable information systems.
  • Item
    Indexing Continuous Data in IoT Environment
    (Faculty of Sciences, 2026) Zahra, KHETTABI; Karima, KHETTABI
    The Internet of Things (IoT) has emerged as one of the most transformative paradigms in modern computing, generating continuous, high-volume, and multi-dimensional data streams from interconnected devices. With the rapid growth of such data, efficient indexing mechanisms have become essential, particularly in Fog Computing environments, where computational and memory resources are significantly more constrained than in centralized cloud infrastructures. The k-Nearest Neighbor (kNN) search is a fundamental operation widely used in IoT applications, including anomaly detection, intelligent monitoring, and real-time analytics. However, executing kNN queries over continuously growing data streams presents significant challenges, such as increased search latency, structural degradation of indexing schemes over time, and higher memory and energy consumption in resourcelimited environments. To address these challenges, this study proposes a hierarchical indexing architecture tailored for Fog-based IoT systems. The proposed approach consists of two complementary phases. In the first phase, data reduction is achieved using the BIRCH algorithm through the construction of a Clustering Feature (CF) tree. This step generates compact micro-clusters that summarize homogeneous regions of the data space, effectively reducing the overall computational complexity while preserving the intrinsic spatial structure of the dataset. In the second phase, an independent binary indexing tree is constructed for each micro-cluster. This separation between clustering and indexing ensures improved data homogeneity within each index structure, resulting in shallower trees and reduced traversal cost during kNN search operations. The proposed methodology exploits the cluster-based partitioning to restrict the search space from the global dataset to localized cluster-specific subspaces. Additionally, kNN queries can be executed in parallel across clusters, further enhancing search efficiency. Experimental results demonstrate that this hierarchical approach significantly reduces search time and computational overhead while maintaining scalability and adaptability to continuous data streams. These findings confirm that integrating cluster-based data reduction techniques with distributed per-cluster indexing constitutes an effective and scalable solution for real-time data indexing in fog-based IoT environments.
  • Item
    Artificial Intelligence Techniques for Chess Puzzle Generation and Automated Evaluation
    (Faculty of Science, 2026) Louahem M’sabah ,Mohamed Akram; Mokhtari ,Rezkallah; Touil, Ghassen
    Chess puzzles are a central tool for developing tactical skill, yet assembling large, high-quality collections by hand does not scale, and the difficulty of a newly created puzzle cannot be known until many players have attempted it. This dissertation addresses both problems through an automated, engine-based pipeline for generating and evaluating tactical chess puzzles. The system keeps generation separate from evaluation: candidate positions are obtained either by extracting them from real games in PGN form or by applying small, legal perturbations to seed positions, and every candidate then passes through the same verification and scoring stages. Verification combines a legality check, a shallow plausibility screen, and a recursive onlymove test that drives the Stockfish 16 engine with a two-line (MultiPV) search to confirm that the solution is forced and uniquely best at each move. Accepted puzzles are scored for uniqueness margin, counter-intuitiveness, novelty, and diversity, and are assigned a difficulty rating without any solving history, using a transparent heuristic together with an optional learned model built on thirteen history-free features. The pipeline is implemented as Tactica, a self-contained client-side web application built with React, chess.js, and a single-threaded WebAssembly build of Stockfish, in which puzzles are generated, solved, and stored entirely in the browser. The implementation is backed by thirty-eight automated tests, all passing, and by a measured generation run on the Lichess Elite database. The learned difficulty model, trained and tested on real Lichess puzzle ratings, reaches a mean absolute error of about 309 rating points and a Pearson correlation of about 0.72 with human difficulty, clearly outperforming both baselines. The main contribution is a reproducible, fully engine-based framework that unifies puzzle generation, soundness and uniqueness verification, and historyfree difficulty estimation.
  • Item
    Système Intelligent pour le Diagnostic précoce du diabète Basè sur les Algorithmes d'Apprentissage Automatique
    (Faculté des science, 2026) Khalfaoui ,Fairouz; Hamida ,Wissal; Lazhar ,Benoudina
    Le diabète est une maladie chronique dont le diagnostic précoce permet de réduire les complications graves. Ce mémoire présente le développement d’un système intelligent de diagnostic précoce du diabète basé sur l’apprentissage automatique. Cinq algorithmes de classification ont été étudiés : KNN, Arbre de décision, SVM, Random Forest et MLP, en utilisant la base de données Pima Indians Diabetes Database. Les données ont été prétraitées par nettoyage, normalisation et équilibrage des classes à l’aide de la technique SMOTE. Les modèles ont été évalués à l’aide de plusieurs métriques de performance telles que l’Accuracy, la Précision, le Recall, le F1-Score et l’AUC-ROC. Les résultats obtenus montrent que le modèle MLP offre les meilleures performances globales avec un F1-Score de 0,6720 et un AUC-ROC de 0,8076. Une interface graphique développée sous Tkinter permet une utilisation simple et efficace du système afin d’assister le diagnostic médical
  • Item
    Design and Development of a Configurable Framework for Soft Sensor Generation in Wastewater Treatment: A Case Study of BOD5 Prediction at the Sidi Merouane WWTP
    (Faculty of Science, 2026) BEN LOUCIF ,Nour El Houda; ZEGHDOUDI, Achouak; TOUIL ,Ghassen
    The monitoring of wastewater quality is essential for the effective operation of wastewater treatment plants and for the protection of receiving environments. Among the main indicators used to assess organic pollution, the five-day biochemical oxygen demand (BOD5) remains particularly important because it reflects the biodegradable organic load of wastewater. However, the conventional determination of BOD5 requires a five-day incubation period under controlled laboratory conditions. This delay limits its usefulness for timely operational decision-making and prevents its direct use for rapid process monitoring. Soft sensors provide a promising alternative by estimating difficult-to-measure variables from other physicochemical parameters that are more readily available. In wastewater treatment, machine learning models can learn the relationships between routinely measured variables and BOD5, thereby providing rapid estimates that may support monitoring and decision-making. Nevertheless, many existing studies focus on a single model developed for a particular dataset and treatment plant. Such solutions are often difficult to reproduce, adapt, compare, or extend to other target variables and operating conditions. This work proposes the design and implementation of a configurable framework for the development of data-driven soft sensors in wastewater treatment. The framework supports the main stages of the soft-sensor development process, including dataset preparation, preprocessing, experimental configuration, regression-model training, performance evaluation, model comparison, interpretation of results, and prediction using a selected model. A functional prototype was implemented to demonstrate these capabilities. The framework was validated through a case study involving BOD5 prediction at the Sidi Merouane wastewater treatment plant in Algeria. Historical physicochemical data were used to train and compare five regression models: Linear Regression, Random Forest Regression, Extra Trees Regression, XGBoost, and CatBoost. Three experimental scenarios were considered: the generation of a local BOD5 soft sensor, the comparison of candidate models, and the evaluation of different input-variable combinations. CatBoost achieved the best performance on the test set, with an R² of 0.9042, an RMSE of 1.5205 mg/L, and an MAE of 1.1270 mg/L. SHAP-based analysis indicated that COD, specific conductivity, and pH were among the most influential variables. vii The results demonstrate the feasibility of the proposed framework for developing BOD5 soft sensors from treatment-plant data. The prototype should be regarded as a decisionsupport and experimentation tool; it does not replace laboratory analysis or constitute a live operational deployment. Future work should focus on validation using multiple treatment plants, chronological data splits, periodic retraining, and integration with real monitoring systems.
  • Item
    Système de reconnaissance des gestes de la main pour le contrôle d’actions informatiques
    (Faculté des sciences, 2026) BOUSMAT ,Aya; ZAIR, Mohamed Walid; BOULNEMOUR, Imene
    L’interaction gestuelle constitue une alternative prometteuse aux périphériques traditionnels tels que le clavier et la souris. Dans ce contexte, ce mémoire présente la conception et la réalisation d’un système intelligent de reconnaissance des gestes de la main destiné au contrôle d’actions informatiques en temps réel. Le système proposé repose sur l’utilisation de la vision artificielle et de l’intelligence artificielle. Les points clés de la main sont extraits à l’aide de MediaPipe, puis exploités par deux modèles complémentaires : un classificateur SVM ( Support Vector Machine ) pour la reconnaissance des gestes statiques et un réseau de neurones BiLSTM (Bidirectional Long Short-Term Memory ) pour la reconnaissance des gestes dynamiques. Une interface graphique a également été développée afin de permettre une interaction intuitive avec l’utilisateur. Les expérimentations réalisées sur un jeu de données personnalisé de plus de 2200 échantillons ont montré des performances élevées, avec une précision de 97,3 % pour le modèle SVM et de 99,4 % pour le modèle BiLSTM. Les résultats obtenus démontrent la robustesse et l’efficacité de l’approche proposée pour le contrôle d’actions informatiques par gestes de la main.
  • Item
    Prédiction des maladies cardiovasculaires à l’aide du Deep Learning
    (FACULTÉ DES SCIENCES, 2026) Lounissi ,BOUDIS; Yousra ,SOLTANI; Seloua ,HADIBY
    Les maladies cardiovasculaires constituent une cause majeure de morbidité et de mortalité, ce qui rend la prédiction précoce du risque particulièrement utile pour la prévention, l’orientation du dépistage et l’aide à la décision médicale explicable. La problématique étudiée est d’évaluer la valeur pratique des modèles d’apprentissage profond par rapport aux modèles classiques pour des données cliniques tabulaires limitées et déséquilibrées, tout en tenant compte de la sensibilité, de la discrimination et de l’interprétabilité. La solution proposée repose sur un cadre comparatif et reproductible associant des modèles classiques et des architectures profondes adaptées aux données tabulaires. L’évaluation ne se limite pas à la performance prédictive : elle intègre également la sensibilité, l’équilibre entre les classes et une analyse explicable des prédictions afin de relier les sorties des modèles aux facteurs de risque cliniquement pertinents. Les résultats montrent que le Perceptron Multicouche (MLP) atteint la meilleure discrimination au même niveau que la régression logistique, tout en obtenant le meilleur équilibre global entre les classes, tandis que ResNetTab fournit la sensibilité la plus élevée, critère important pour le dépistage. Ces résultats positionnent les modèles profonds comme des alternatives prometteuses et complémentaires lorsque la priorité porte sur la sensibilité, l’équilibre des performances et l’analyse des prédictions, avec un potentiel renforcé par des données plus riches.
  • Item
    EduTwin AI A Learner-Centric Digital Twin and Retrieval-Augmented Generation platform for Personalized Higher Education
    (Faculty of Sciences, 2026) Ramdane ,Abdeldjalil; Guerriche Ala ,Abdel Moumene; Kerraoui Sarra
    The rapid evolution of digital education has exposed the limitations of traditional, "onesize- fits-all" learning management systems, highlighting an urgent need for intelligent, adaptive environments. This thesis presents EduTwin AI, an innovative platform designed to bridge the gap between static academic resources and the dynamic cognitive state of the learner. By natively integrating a Learner-Centric Digital Twin with a Retrieval-Augmented Generation (RAG) architecture, the system creates a continuous physical-digital feedback loop. The Digital Twin employs Bayesian Knowledge Tracing (BKT) to autonomously model and predict the student's concept mastery in real time. Simultaneously, the RAG pipeline acts as a strict pedagogical filter, ensuring that all interactions generated by the Large Language Model (LLM) are exclusively anchored to validated course materials, thereby eliminating AI hallucinations. Through adaptive assessments, personalized study paths, and a conversational tutor, EduTwin AI dynamically calibrates its complexity to match the student's Zone of Proximal Development. This work demonstrates the profound potential of combining real-time cognitive tracking with generative AI to deliver a truly personalized, secure, and academically rigorous learning experience.
  • Item
    conception et développement dune plateforme E-Commerce Multi-Vendeurs de Vêtements avec Essayage Virtuel par IA
    (Faculté de science, 2026) REHIM, Safa; TABANI ,Chams; BENOUDINA ,Lezhar
    La réalité augmentée constitue une innovation majeure dans le domaine du commerce électronique, car elle améliore considérablement l’expérience utilisateur en la rendant plus interactive et immersive. En effet, elle permet de visualiser les produits en trois dimensions dans un environnement réel et offre également la possibilité d’un essayage virtuel, notamment dans le secteur de la mode. Ainsi, elle réduit l’incertitude liée aux achats en ligne et renforce la confiance des consommateurs dans leurs décisions. De plus, elle contribue à diminuer les retours de produits grâce à un choix plus précis. Par ailleurs, elle rend l’expérience d’achat plus attractive et engageante, tout en procurant un avantage concurrentiel important aux entreprises. Enfin, elle améliore les performances commerciales en augmentant les ventes et la satisfaction des clients.
  • Item
    Intelligent Search Submarine Using Artificial Intelligence
    (Faculty of Science, 2026) SACI ,Nedjla; LARAFA, Anfal; BOULEHOUACHE,Soufiane
    The rapid development of Artificial Intelligence and embedded technologies has significantly contributed to the creation of intelligent systems capable of operating in complex environments. Underwater search is one of the fields that can benefit from these technologies, as underwater environments are often characterized by reduced visibility, communication difficulties, and limited human intervention. This thesis presents the design and development of a low-cost intelligent Remotely Operated Vehicle (ROV) prototype dedicated to underwater object search. The proposed system combines embedded electronics, remote control, computer vision, and Artificial Intelligence to assist the operator during underwater search missions. The ROV is remotely controlled by the operator, while the vision system analyzes images captured by the onboard camera and detects objects of interest using a YOLO-based object detection model. The entire system is built around a Raspberry Pi 5, which serves as the main processing unit and is responsible for camera management, motor control, video processing, and the execution of the Artificial Intelligence model. The Raspberry Pi 5 also manages communication between the ROV and the external control station. The current prototype is designed as a proof of concept, demonstrating the feasibility of integrating Artificial Intelligence into a low-cost underwater robotic platform for object search applications. This work highlights the potential of intelligent underwater systems to support search tasks and provides a foundation for future improvements, including enhanced communication methods, improved waterproofing, more powerful detection models, and advanced real-time processing capabilities.