Projet de Fin d'Études : Mémoire d'Ingénieur

Permanent URI for this collection

Browse

Recent Submissions

Now showing 1 - 13 of 13
  • Item
    Deep Reinforcement learning for vehicule Platooning Optimization
    (Tassadit, 2025-01-21) Amani Chaimaa SELLAM
    Automated vehicle platooning has emerged as a significant method for improvi traffic efficiency, reducing fuel consumption, and enhancing road safety. This investigates the optimization of vehicle platooning using reinforcement learning techniques, specifically focusing on Deep Q-Networks (DQN) integrated with dueli networks and prioritized experience replay (PER). A two-layered approach is employed, where the first layer identifies joinable platoons and evaluates the benef joining them, and the second layer uses reinforcement learning to optimize merging, lane-changing, and acceleration strategies. The results of the simulation demonstrate that the proposed method significantly reduces travel time, fuel consumption, a improves the overall safety of the platooning process. Future work includes extending the model to mixed traffic environments with both autonomous and human-dri vehicles, as well as integrating predictive traffic models
  • Item
    Traffic Management Optimization in the Presence of Automated Vehicles
    (Tassadit, 2025-01-21) Cheima MEZDOUR
    Traffic congestion on freeways is a significant issue today due to the high volume of vehicles. This congestion impacts traffic flow and can lead to severe delays. Effective traffic management methods and solutions are essential to mitigate this problem and enhance safety for all road users. One specific issue in freeway traffic congestion is the problem of ramp metering. This occurs when vehicles merge from an entrance ramp onto the main freeway. Any malfunction or inefficiency in the traffic light system at these ramps can cause significant congestion and even lead to accidents. By implementing a sophisticated decision-making system that follows a real-world network, we can develop better policies to address this issue. This approach can optimize traffic flow, reduce congestion, and improve overall road safety. We proposed a deep q-learning algorithm to address this problem, which guarantees better learning policies and adapt to changing traffic conditions in real-time, learning optimal actions to take at different states of traffic flow in complex environments. Deep Q-Networks can be used across multiple ramps, so this can improve the scalabil ity of the model. So during this study, we will present the performance of this algorithm compared to traditional systems in improving the efficiency of traffic control.
  • Item
    Intelligent Video Recording Optimization using Activity Detection for ESTIN Surveillance Systems
    (Tassadit, 2025-01-21) MELIZOU Ouassila; TOUATI Hayet
    Surveillance systems often face the challenge of managing extensive amounts of footage, much of which is irrelevant, leading to inefficient storage and difficulty in event retrieval. This thesis addresses this issue by proposing an optimized video recording solution that focuses on activity detection. The proposed approach utilizes a hybrid method combining motion detection via frame subtraction and object detection using You Look Only Once model . This strategy aims to record only scenes with human activity, thereby reducing unnecessary footage and optimizing storage usage. The developed model demonstrates superior performance, achieving precision metrics of 0.855 for car detection and 0.884 for person detection, highlighting its effectiveness in enhancing the efficiency of surveillance systems. However, some limitations remain, such as false positives and false negatives in bad weather conditions like powerful winds.
  • Item
    Image Text Similarity using Deep Learning Object Detection and Word Spotting Approach
    (Tassadit, 2025-01-21) Billal MOKHTARI; Lilia MAHDID
    With the fast expansion of Deep Learning, multi-modal models have become increasingly popular for tasks requiring complex data inputs. Content gen eration—such as image, video, or text generation—as well as recent object detection and segmentation methods, frequently use Large Language Mod els (LLMs). This project focuses on enhancing image and text similarity measures, aiming to improve the CLIP (Contrastive Language-Image Pre training) method by examining the impact of object semantics on image descriptions. Our approach, named ODITS (Object Driven Image and Text Similarity), uses the CLIP model pre-trained with the ViT-B/32 architecture, which is subsequently fine-tuned for our specific purposes. We evaluated the performance of the fine-tuned model using modified metrics, selecting the optimal checkpoint based on precision to minimize false associations between descriptions and images. Our findings indicate that this optimal checkpoint is 10% more precise than the original checkpoint. The weights from this model will be integrated into ODITS’s shared components with CLIP, providing a robust starting point for further optimization. The research component of the ODITS model, including theoretical and preliminary analysis, is also discussed, providing insights into its potential and areas for future development.
  • Item
    investigation of Poisoning Effects and Defensive Measures on Reinforcement Learning-based Intrusion Response via Optimal Stopping
    (Tassadit, 2025-01-21) ALMAMMA Amir
    As cyber infrastructures become increasingly complex and integral to critical sectors like finance, healthcare, and national security, the need for advanced security systems continues to grow. Intrusion Detection System (IDS) have become a cornerstone in de tecting and mitigating a wide range of cyber threats. However, as attacks become more sophisticated, there is a pressing need for systems that not only detect intrusions but also autonomously deploy defensive actions in real-time without human intervention. This engineering report focuses on the development of poisoning strategies aimed at compromising the training process of Reinforcement Learning (RL)-based agents in active IDS. It also proposes defensive measures to mitigate the adverse effects of such poisoning attacks, including the implementation of a dynamic reward adjustment framework through a MetaAgent strategy. The effectiveness of these strategies was evaluated within the Cyber Security Learning Environment (CSLE) framework, exploring the impact of poisoning on intrusion length and agent performance. Furthermore, the report highlights the challenges faced during the experiments, partic ularly the limitations in computational resources, which constrained the ability to conduct large-scale experiments. Future work will explore alternative defensive strategies and fur ther investigate methods to enhance the robustness of RL-driven IDS
  • Item
    Seismic image segmentation based on Deep Learning for the Characterization of Hydrocarbon Reservoirs in the Oil & Gas Industry: Comparative study
    (Tassadit, 2025-01-21) Anis MOHAMMEDI
    Accurate identification and delineation of Salt Domes from seismic images play a critical role in geological studies and resource exploration. How ever, traditional methods often struggle with complex geological structures and require extensive manual intervention. This work addresses these chal lenges by proposing a deep learning-based approach for semantic segmen tation of Salt Domes. We introduce novel techniques including Transform ers for spatial context aggregation, U-Net for precise feature extraction, and VAE-Liquid layers for enhanced representation learning. Through rigorous experimentation and evaluation on real-world datasets, we demonstrate the effectiveness of our approach in automating and improving the accuracy of Salt Dome identification. This work contributes to advancing automated ge ological analysis, offering insights into subsurface structures vital for both exploration and hazard assessment
  • Item
    Predictive maintenance using Machine learning and sensor data
    (Tassadit, 2025-01-21) BELAYALI Rezkia; OURARI Tinhinane
    This thesis explores the relationship between Machine Learning (ML) and sensor data to achieve predictive maintenance (PdM) in Industry 4.0. The focus is on Deep Learning (DL) methods for developing a comprehensive predictive maintenance process using time series data from vibration sensors. The IMS (Center for Intelligent Maintenance Systems) bearing dataset serves as the foundation for this exploration. The process progresses from predicting the machine's next operational state, where the best result was achieved by an LSTM (Long Short-Term Memory) single model in the frequency domain, with an RMSE (Root Mean Square Error) of 0.0005, MAE (Mean Absolute Error) of 0.00004, and an R² (coefficient of determination) of 0.93. It then identifies change points indicative of anomalies. Finally, the goal is to predict the remaining useful life (RUL) of the machinery, where the best result was achieved using the hybrid DL model LSTM-CNN (Long Short-Term Memory - Convolutional Neural Networks), with an RMSE of 0.001 and an MAE of 0.0007. By implementing these techniques, the thesis aims to demonstrate the efficacy of ML and sensor data in establishing proactive maintenance strategies within the industry 4.0 framework, leading to significant cost savings and enhanced operational efficiency.
  • Item
    Text Mining for Predictive Maintenance Case Study: Cevital agro-industry
    (Tassadit, 2025-01-21) Asma BOUROUBA; Narima ALKAMA
    Achieving operational excellence is vital for maintaining competitiveness in today’s in dustrial landscape. Traditional maintenance strategies, both reactive and preventive, often fail to fully leverage the extensive data available. The advent of Industry 4.0, with its fo cus on data acquisition and analytics, has led to the development of predictive maintenance, enabling real-time insights into equipment health and proactive interventions. This thesis introduces an innovative approach to predictive maintenance by applying NLP techniques to textual maintenance logs. Our method utilizes BERTopic for embedding, K-means cluster ing for grouping intervention descriptions, and autoencoders for feature extraction to identify equipment degradation states. These states are then used to develop a Bayesian network based predictive model. Applied to Cevital, an Algerian agri-food conglomerate using Coswin 8i, our approach aims to enhance maintenance planning and operational efficiency. The clus tering model achieved a high silhouette score of 97%, and by integrating it with the Bayesian model, we attained an accuracy of 83%. This study highlights the potential of integrating advanced NLP and predictive analytics into maintenance management. This work underscores the transformative potential of integrating cutting-edge NLP tech niques and predictive analytics into traditional maintenance management practices. By har nessing textual data effectively, organizations can transition from reactive to proactive main tenance strategies, thereby minimizing downtime, reducing operational costs, and ultimately enhancing overall productivity and competitiveness in the industry.
  • Item
    Exploring the Integration of Visual Data for Enhancing the Accuracy and Reliability of Recommender Systems
    (Tassadit, 2025-01-21) TOUATI Yanis
    This dissertation presents a multi-model hybrid recommendation approach aimed at enhancing recommenda tion systems’ performance. The approach integrates content-based recommendation, collaborative filtering, and clustering techniques to exploit the strengths of each method. We discuss the architecture and moti vation behind our approach, along with its objectives, expectations, limitations, and potential challenges. The proposed model incorporates visual data processing through deep feature extraction techniques, as well as textual-based models utilizing word and sequence embedding learning. We also detail collaborative fil tering recommendation using similarity metrics, alongside clustering techniques including deep contrastive clustering. for further enhancement, we explore the hybridization of these methods to further improve rec ommendation accuracy. Experimental results and discussions are also displayed at the end of our paper, offering insights for future research in recommendation system development.
  • Item
    Recommender systems for multimodal transportation systems in smart cities
    (Tassadit, 2025-01-21) Madadi Mounia
    Transportation recommendation systems have been trending for quite some time due to their continu ous potential for improvement. Among the most interesting advancements are multimodal transportation recommendation systems, which provide suggestions for traveling from one location to another using a combination of available transportation modes. In this thesis, we present a multimodal transportation recommendation system that recommends trajectories to users based on their personal preferences. Our system consists of two main phases. The first phase involves trajectory generation, where we search for optimal trajectory combinations between the starting point and the destination using Particle Swarm Op timization, followed by post-processing on the trajectories. Once the trajectories are generated, we rank them using the RankNet model trained on previously selected user trajectories, employing a content-based approach.After testing our system, we observed that the generated trajectories were quite feasible and the recommendations were highly accurate.
  • Item
    Energy Efficiency in Wireless Sensor Networks
    (Tassadit, 2025-01-21) BOUTAOUI Nada
    Wireless Sensor Networks (WSNs) face significant challenges in terms of energy efficiency, requiring strategies that balance power consumption, data quality, and network reliability to maximize network lifetime. This issue becomes more complex with the growing adop tion of IoT applications and the dynamic nature of sensor deployment environments. In this context, energy-efficient solutions for WSNs have gained prominence, especially those incorporating machine learning and graph-based techniques. This thesis investigates energy optimization strategies for Wireless Sensor Networks (WSNs) in the context of IoT applications and dynamic deployment environments. We introduce Delta-GCN, a novel model that combines Graph Convolutional Networks (GCNs) with a customized Long Short-Term Memory (LSTM) layer to capture spatial and temporal dependencies in sensor networks. The model is integrated into a Federated Learning (FL) framework, enabling decentralized training while preserving data privacy. Our semi-supervised approach addresses the challenge of missing data in WSNs through data imputation, simultaneously optimizing energy consumption by reducing unnecessary sensor activity. Extensive empirical evaluations demonstrate that Delta-GCN significantly outperforms existing state-of-the-art methods across key performance metrics, including F2 score, precision, recall, energy consumption, and network lifetime. The research find ings underscore the potential of integrating GCNs and FL for developing scalable, energy efficient WSN deployments. This work contributes to the field by advancing energy man agement strategies applicable to real-world sensor networks, paving the way for more sus tainable and efficient IoT infrastructures.
  • Item
    Evaluation of Machine learning Models for Detecting Adversarial attacks on Anomaly Detection Oriented Dataset
    (Tassadit, 2025-01-21) Ahmed Yacine Bouchouareb
    This report evaluates the capability of machine learning models in detecting ad versarial attacks on a given dataset, with a test on the NSL-KDD dataset. The study’s objectives are twofold: first, to analyze the dynamics of the autoencoder’s reconstruction loss for normal, anomalous, and adversarial data points; second, to benchmark various candidate models, including Support Vector Machines (SVM), Decision Trees, and Naive Bayes, in detecting adversarial data crafted using Fast Gradient Sign Method (FGSM)[5] and Projected Gradient Descent (PGD)[10] techniques. Additionally, this research tests a feature engineering technique that considers the reconstruction loss as a vector[21], as suggested in recent literature. The results demonstrate that the reconstruction loss exhibits similar behavior between anomalous and adversarial examples, differentiating them from normal records in terms of mean and variance. Furthermore, the study reveals that the benchmarked models face significant challenges in detecting PGD attacks com pared to FGSM attacks.
  • Item
    COMPUTER VISION BASED INDUSTRIAL INSPECTION SYSTEM
    (Tassadit, 2025-01-21) Adel BELLAHCENE
    This review explores the cutting-edge applications of computer vision (CV) in industrial inspection. We highlight the limitations of traditional methods and showcase how CV, coupled with Machine Learning (ML) and Deep Learning (DL) techniques like Convolutional Neural Networks (CNNs), is revolutionizing defect detection and quality control. The review explores the recent advancements in real-time object detection with YOLO models, emphasizing their potential for high-speed production lines. We conclude by discussing the future of CV in industrial inspection, including integration with robotics and sensor fusion for intelligent and comprehensive inspection systems