Telecommunication
Permanent URI for this collectionhttps://dspace.univ-boumerdes.dz/handle/123456789/3080
Browse
133 results
Search Results
Item Automated detection of diabetic retinopathy using transfer learning with deep CNN models(University M’hamed Bougara : Institute of Electrical and Electronic Engineering (IGEE), 2025) Messiheddine, Amira; Daamouche, AbdelhamidRecently, the integration of artifi cial intelligence in the medical fi eld has increased dueto technological advancements and the availability of large-scale medical data. As diabetes is growing around the world, changes in the retina of diabetic patients cause a condition called diabetic retinopathy (DR), which can lead to vision loss if not properly managed. Early detection of diabetic retinopathy helps in the treatment of this disease and thus prevents blindness. Deep learning made it possible to develop automated models capable of detecting and classifying DR from retinal images. This project focuses on utilizing deep learning pretrained architectures, including ResNet50, Effi cientNetB3, and DenseNet121, for multi-class classifi cation ofDR using publicly available datasets such as the APTOS and DDR datasets. At fi rst, the study focused on a 5-class classifi cation, and through diff erent experiments, the best model achieved an overall accuracy of approximately 82% and 83% for both datasets. However, to address some limitations and simplify the task, a 3-class classifi cation approach was adopted, in which the performance of the models has improved, highlighting the impact of class imbalance on the results of our study. The fi ndings of this project demonstrate the potential of deep learning in automating DR diagnosis and helping practitioners make more accurate decisions. Keywords: Retinal imaging, Fundus images, diabetic retinopathy detection, APTOS, DDR, Data cleaning, Pre-processing, deep learning, convolutional neural networks, transfer learning, ResNet50, Effi cientNetB3, DenseNet121.Item 5G NR OFDM multi numerologies interference mitigation(University M’hamed Bougara : Institute of Electrical and Electronic Engineering (IGEE), 2025) Lahlah, zehor; Bendaimi, Aicha; Challal, MouloudThis work examines the importance of multi-numerologies in fifth-generation (5G) using orthogonal frequency-division multiplexing (OFDM). The impact of multi-numerologies on coexistence issues, latency, and spectrum efficiency is highlighted in this work. A discussion of potential future research directions following a comparative study of previous efforts is also presented. The proposed work presented is a hybrid mitigation approach that combines time domain mitigation by zero padding WOLA filtering and frequency domain mitigation using a fixed guard band. The work discusses each case separately, then combines them and studies the overall improvements and the possible trade-offs. The hybrid ZP-WOLA mitigation approach reached a 78% - 91% BER reduction percentage across the 0 - 30 dB SNR range in the ideal case (under AWGN), while under moderate flat Rayleigh fading it achieved BER reductions of 37% - 72% over the same range.Item Darsy : an E-learning Platform Using WebRTC for real-time communication(University M’hamed Bougara : Institute of Electrical and Electronic Engineering (IGEE), 2025) Tlemsani, Oussama; Tabti, Mohamed Nourelislam; Benmoussa, YahiaThis thesis presents the design and development of an integrated e-learning platform that addresses the growing need for accessible, interactive, and real-time educational environ-ments. The objective of the proposed solution is to seamlessly connect students and teachers, regardless of their geographical location, through a unifi ed web-based system. Unlike traditional platforms that rely on external video conferencing tools, this solution embeds real-time communication as a core feature of the learning experience. At the heart of the platform lies WebRTC, a modern technology that enables low-latency, peer-to-peer video and audio communication directly in the browser. WebRTC off ers several built-in features critical to online education, including screen sharing, camera and microphone management, and secure encrypted channels—all without the need for plugins or additional software. These capabilities create a smooth, interactive classroom experience that mimics face-to-face instruction. Beyond communication, the platform supports full course management—from enrollment to content delivery, real-time chat, assignment submission, and grading. The system architecture follows a decoupled monolithic model, ensuring scalability and modularity. React and Vite power the frontend interface, while Django REST and Channels provide backend services, APIs, and real-time WebSocket support. Agora.io further enhances the WebRTC layer with scalable media handling for multi-user sessions. This project demonstrates how tightly integrated real-time communication can transform e-learning into a more dynamic, immersive, and inclusive experience for both students and educators. Keywords: WebRTC, E-learning, Agora.io, Django, React, Redis, Postgress, Tailwind, Locust.Item Large language models (LLMs) exploration on pubMedQA and deployment(University M’hamed Bougara : Institute of Electrical and Electronic Engineering (IGEE), 2025) Amine, Chaima; Cherifi, DalilaMaking LLM research accessible to everyone faces a big problem: fi ne-tuning these models requires so much computational power that most researchers and institutions simply can’t afford to participate in cutting-edge development. This creates a divide where only well-funded organizations can actually work with and improve LLM capabilities, while others are limited to pre-trained models that often miss the mark for their particular use cases. Optimization techniques like 4-bit quantization, LoRA, and aggressive frameworks like Unsloth claim they can solve this problem, but the thing is that current research only looks at how much computational power these methods save, not whether the optimized models actually work well in practice. This brings us to a critical question that nobody has really answered yet: when you optimize an LLM to run on limited hardware, which capabilities do you lose first, and how does this affect the model’s ability to handle domain-specific tasks like medical question answering? Our thesis systematically evaluate these trade-offs by looking at computational efficiency, task performance, and whether the models actually work in real-world scenarios. We used Llama 3.1 8B for medical question answering on the PubMedQA dataset, comparing base versus instruction-tuned models and testing recent advancement Zero-shot and few-shot learning as well as aggressive optimization against standard fi ne-tuning methods. To demonstrate the practical impact of our work, we deployed the selected model using a web interface, making it accessible for real-world testing and exploration. What we found was pretty clear: aggressive optimization techniques often make models computationally cheaper but practically useless. Unsloth’s optimization reduced computational requirements signifi cantly but made the fi ne-tuned model completely unreliable for medical question answering, while classical methods kept the model working well despite costing more resources. The idea that we can have both computational effi ciency and good model perfor- mance at the same time just isn’t true, and while optimization techniques are promising for researchers with limited resources, we still need balanced approaches that prioritize keeping themodel functional.Item Middlepoint selection for segment routing : a deep reinforcement learning approach(University M’hamed Bougara : Institute of Electrical and Electronic Engineering (IGEE), 2025) Elahouel, Imad; Boutellaa, ElhoucineThe modern telecommunications landscape, characterized by 5G infrastructure and network slicing, presents significant challenges in efficient traffic management, where traditional routing protocols like Open Shortest Path First (OSPF) and Equal-Cost Multi-Path (ECMP) often result in suboptimal traffic distribution, congestion,and underutilized paths, despite the potential of Segment Routing (SR) and Software-Defined Networking (SDN ) technologies that currently demand extensive manual configuration and lack dynamic automation . This thesis introduces an innovative solution integrating Segment Routing Traffic Engineering with Deep Reinforcement Learning (DRL), framing network optimization as a Markov Decision Process where autonomous agents learn optimal routing strategies through environmental interac-tion, with a focus on automated middlepoint selection for segment routing paths to minimize Maximum Link Utilization (MLU) while upholding routing efficiency and network constraints. Employing Proximal Policy Optimization (PPO) enhanced with action masking to handle connectivity constraints, the DRL agent processes network state representations including link utilization and traffic demands to select optimal intermediate nodes, enabling intelligent, automated decision-making that eliminates manual configuratio nneed san ddeliver sadaptive, millisecond-level routing decisions. Comprehensive experimental validation across multiple network topologies reveals substantial performance gains, with the approach achieving 14.4% MLU improvement over ECMP on NSFNet and 17.4% on GEANT2, underscoring the efficacy of DRL- SR integrationfor automated traffic engineering, particularly in 5G network slicing scenarios requiring differentiated Quality of Service guarantees.Item Design, simulation, and performance assessment of MPLS core networks(University M’hamed Bougara : Institute of Electrical and Electronic Engineering (IGEE), 2025) Baouali, Chems Eddine; Benzaoui, MessaoudaWith the continuous growth of the Internet and the increasing diversity of user demands, service providers face the challenge of delivering fast, reliable, and secure communication across large-scale networks. Multi-Protocol Label Switching (MPLS) has become a cornerstone technology in modern backbone networks because it goes beyond traditional IP routing by combining flexibility, efficiency, and support for advanced services. This project explores the design and simulation of an MPLS-based core network using GNS3, focusing on how Internet Service Providers (ISPs) can manage complex customer requirements while ensuring performance and scalability. Provider Edge (PE) routers are considered as the key control points, enabling customer isolation through Virtual Routing and Forwarding (VRF), resolving IP address overlaps with external BGP (eBGP), and ensuring seamless communication across multiple domains. While, Label Distribution Protocol (LDP) is used to establish label-switched paths. The findings show that MPLS is not only a powerful forwarding technique but also a strategic enabler for ISPs to deliver advanced services with efficiency and reliability. Future work will emphasize the deeper integration of QoS and MPLS-TE, providing a more adaptive and intelligent backbone capable of meeting the ever-growing demands of next-generation networks.Item Advanced classification methods applied to ECG signals(University M’hamed Bougara : Institute of Electrical and Electronic Engineering (IGEE), 2025) Merdji, Fadia; Belkadi, Mohamed AmineThis work presents an advanced automated ECG classificatio nsyste mtha taddresse scritical challenges in cardiac arrhythmia diagnosis by integrating sophisticated signal processing with state-of-the-art deep learning techniques. Beginning with a thorough examination of cardiac electrophysiology, ECG waveform characteristics, and noise artifacts, the research establishes a robust foundation for subsequent algorithmic development. The study systematically evaluates traditional machine learning methods (SVM, KNN, DNN) and modern deep learning archi- tectures, culminating in an innovative multi-stage framework featuring optimized Butterworth filtering ,hybri dFourie rwavele tfeatur eextraction ,an d aspecialize d2 DCN Ndesign .Exten-sive validation on the MIT-BIH Arrhythmia Database demonstrates exceptional performance, achieving 98.60% classificatio naccurac yan d99 %precisio nfo rcritica larrhythmia swhil ere-vealing important insights about minority class recognition challenges. The work makes three key contributions: (1) a comprehensive theoretical and methodological framework for ECG analysis, (2) significan tperformanc eimprovement sthroug hmultimoda lfeatur efusion ,an d(3) practical guidelines for clinical implementation. Future research directions focus on real-time processing optimization, attention mechanism integration, and multi-modal data fusion to en-hance diagnostic capabilities across diverse healthcare environments, representing a substantial advancement in intelligent cardiac monitoring technology with the potential for significant clinical impact.Item Implementation and evaluation of conjugate gradient-based MIMO detectionin sionna for 5G and 6G-like scenarios(University M’hamed Bougara : Institute of Electrical and Electronic Engineering (IGEE), 2025) Tahir, Imene; Saoudi, Amina; Smaili, NessrineAs wireless communication systems progress toward higher spectral effi ciency and support for massive connectivity, robust detection techniques for Multiple-Input Multiple-Output systems have become increasingly important. While traditional linear detectors such as Zero-Forcing and Linear Minimum Mean Square Error are simple to implement, their performance can degrade in large-scale scenarios due to high computational cost or limited accuracy. To overcome these challenges, the Conjugate Gradient algorithm has gained attention as an effi cient iterative method for solving linear systems without requiring matrix inversion, making it highly suitable for massive MIMO detection. In this work, we implement and evaluate a CG-based linear MIMO detector using Sionna , an open- source link-level simulation library developed on TensorFlow. The system is assessed using performance metrics such as Bit Error Rate versus Energy per Bit to Noise Power Spectral Density across different modulation schemes and channel conditions. The results confirm that CG provides competitive detection performance with reduced complexity, making it a promising candidate for scalable and machine learning-oriented wireless system design.Item Performance study of hybrid OFDM/OTFS system over doubly selective channel(University M’hamed Bougara : Institute of Electrical and Electronic Engineering (IGEE), 2025) Dihia, Amani; Razaoui, Feriel; Smaili, NesrineAs the demand for fast and reliable wireless communication grows particularly in high-mobility environments, traditional modulation schemes are limited by Doppler shifts. Orthogonal Frequency Division Multiplexing (OFDM) performs effectively in low mobility scenarios, but experiences performance degradation under dynamic conditions. On the other hand, Orthogonal Time Frequency Space (OTFS) modulation offers greater robustness in high mobility channels due to its operation in the Delay Doppler domain, but it introduces greater complexity and does not outperform OFDM in low mobility channels. This work proposes a hybrid system that dynamically estimates user velocity and switches between OFDM and OTFS to maintain optimal performance. MATLAB simulations evaluating the Bit Error Rate (BER) versus Signal-to-Noise Ratio (SNR) show that OTFS outperforms OFDM at higher speeds, while OFDM remains more effi cient in static or low mobility settings. The adaptive solution ensures effective communication under various mobility conditions and supports the waveform fl exibility for future wirelessnetworks.Item Design and implementation of a single phase grid tied inverter(University M’hamed Bougara : Institute of Electrical and Electronic Engineering (IGEE), 2025) Asfirane, Nihad; Badi, Lotfi; Bentarzi, HamidWith the rapid spread of renewable energy sources, particularly solar energy, photovoltaic (PV) systems which employ solar cells to convert solar energy into electrical power become widely used nowadays. However, PV panels generate DC electricity whereas electrical grids and household appliances operate on AC electricity. So gridtied inverters are used as an interface between PV panels and the electrical grid, ensuring effi cient power conversion and synchronization with grid parameters. This report focuses on the design and implementation of a single-phase grid-connected inverter, providing a comprehensive analysis of its control strategy, hardware design, and implementation process.The control strategy, implemented on the LAUNCHXL-F28379D development board, consists of a proportional-integral (PI) controller for the DC-link voltage regulation, a proportional-resonant (PR) controller for current control, and a phase-locked loop (PLL) to ensure precise synchronization with the grid. An LCL fi lter is used to enhance power quality and minimize harmonics in the output current and voltage, effectively reducing high-frequency switching components before connecting the inverter to the grid. The proposed system is designed to comply with grid codes, ensuring stable operation, high power injection effi ciency, and minimal dis- tortion in compliance with power quality standards. It is important to note that this work is limited to the design and control of the inverter stage only, without addressing the complete PV system.