Doctoral Dissertations

Permanent URI for this collection

https://repository.univ-msila.dz/handle/123456789/3603

Browse

Now showing 1 - 20 of 93

Open Access
DEVELOPMENT AND IMPLEMENTATION OF AN INTELLIGENT SYSTEM FOR PREDICTING ALZHEIMER’S DISEASE
(University of M'Sila, 2025) Zakaria MOKADEM
Alzheimer’s Disease (AD) is a progressive neurodegenerative disorder and the leading cause of dementia, primarily affecting individuals over the age of 65. Patients experience a gradual decline in cognitive functions, including memory, judgment, and functional abilities. The increasing preva- lence of AD underscores the urgent need for accurate and eﬀicient computer-aided diagnosis (CAD) methods. While most AD diagnostic studies focus on neuroimaging data, its high cost and limited accessibility restrict its use to urban populations with advanced medical facilities, introducing poten- tial bias in machine learning (ML) models. Additionally, behavioral and psychological symptoms, which are critical for AD diagnosis, are often overlooked. To address these gaps, this study inves- tigates two alternative diagnostic approaches—neuropsychological assessments and plasma protein biomarkers—using ML techniques. In the first approach, we evaluate the predictive performance of ML models using neuropsy- chological assessment data collected through low-cost, first-line diagnostic tools. Both binary and multiclass classification techniques are applied to analyze five neuropsychological assessments for AD detection. The findings highlight the potential of neuropsychological data in early AD diagnosis and emphasize the advantages of ML-based approaches in clinical decision-making. In the second approach, we explore plasma protein biomarkers for AD diagnosis. We apply Sequential Backward Feature Selection (SBFS) and Analysis of Variance (ANOVA) to extract sig- nificant proteins from a dataset of 146 plasma proteins collected from 566 individuals, including both AD patients and healthy controls. Five ML models—Decision Tree, Random Forest, Extremely Randomized Trees, Extreme Gradient Boosting, and Adaptive Boosting—are used for initial clas- sification, followed by XGBoost and AdaBoost for final validation. Both approaches demonstrate the feasibility of low-cost, accessible diagnostic methods for AD detection. Neuropsychological assessments provide valuable cognitive and behavioral insights, while plasma protein biomarkers offer a promising non-invasive alternative to traditional techniques. These findings support the integration of ML-driven approaches in AD diagnosis, paving the way for scalable and affordable screening strategies that can benefit a broader population.
Open Access
Conception et analyse d ’une antenne tag RFID pour les applications UHF
(University of Msila, 2025-06-30) Meriem MOKRANI; Maroua BAOUCHE; ENCA/Elhadi KENANE; CO.ENCA/Mohamed SAHED
UHF RFID (Ultra High Frequency, 860-960 MHz) antennas are the most essential components of medium- and long-distance RFID systems. These antennas work by electromagnetic coupling, enabling them to read from several meters away without contact. Their design is typically based on dipole antennas or printed dipole antennas, optimized for the embedded chip for smooth data transmission. These antennas are widely used in logistics, inventory management and industry due to their high speed and ability to read multiple tags simultaneously. In this work, an RFID antenna has been designed using full-wave 3D simulation software, where we modeled a Monza 3 embedded chip using lumped elements. To reach the final model of our proposed Tag RFID antenna, many design techniques such as meanders, omega loop, and tapered meanders are used in the design the antenna. The obtained results are very acceptable, which is confirmed by several parameters such as the reflection coefficient S11, impedance Z11 vs frequency, and the antenna pattern.
Open Access
DEVELOPMENT AND IMPLEMENTATION OF AN INTELLIGENT SYSTEM FOR PREDICTING ALZHEIMER’S DISEASE
(University of Mohamed Boudiaf - M’sila, 2025-07-02) Zakaria MOKADEM; enca/DJERIOUI Mohamed
Alzheimer’s Disease (AD) is a progressive neurodegenerative disorder and the leading cause of dementia, primarily affecting individuals over the age of 65. Patients experience a gradual decline in cognitive functions, including memory, judgment, and functional abilities. The increasing prevalence of AD underscores the urgent need for accurate and efficient computer-aided diagnosis (CAD) methods. While most AD diagnostic studies focus on neuroimaging data, its high cost and limited accessibility restrict its use to urban populations with advanced medical facilities, introducing potential bias in machine learning (ML) models. Additionally, behavioral and psychological symptoms, which are critical for AD diagnosis, are often overlooked. To address these gaps, this study investigates two alternative diagnostic approaches—neuropsychological assessments and plasma protein biomarkers—using ML techniques. In the first approach, we evaluate the predictive performance of ML models using neuropsychological assessment data collected through low-cost, first-line diagnostic tools. Both binary and multiclass classification techniques are applied to analyze five neuropsychological assessments for AD detection. The findings highlight the potential of neuropsychological data in early AD diagnosis and emphasize the advantages of ML-based approaches in clinical decision-making. In the second approach, we explore plasma protein biomarkers for AD diagnosis. We apply Sequential Backward Feature Selection (SBFS) and Analysis of Variance (ANOVA) to extract significant proteins from a dataset of 146 plasma proteins collected from 566 individuals, including both AD patients and healthy controls. Five ML models—Decision Tree, Random Forest, Extremely Randomized Trees, Extreme Gradient Boosting, and Adaptive Boosting—are used for initial classification, followed by XGBoost and AdaBoost for final validation. Both approaches demonstrate the feasibility of low-cost, accessible diagnostic methods for AD detection. Neuropsychological assessments provide valuable cognitive and behavioral insights, while plasma protein biomarkers offer a promising non-invasive alternative to traditional techniques. These findings support the integration of ML-driven approaches in AD diagnosis, paving the way for scalable and affordable screening strategies that can benefit a broader population
Open Access
Etude de l’influence des couches tampons pour les cellules solaires multi-jonctions
(University of M'Sila, 2025-05-20) BOUZIDI Amina
Dans cette thèse, nous avons étudié le comportement et l’optimisation des cellules solaires tandem CIGS/Si par simulation numérique, en nous concentrant spécifiquement sur l’impact de différentes couches tampons (CdS, CdZnS, ZnS, ZnSTe, Zn(S,O)) dans la cellule supérieure en CIGS. Après avoir établi les bases théoriques de la modélisation et de la simulation à l’aide du logiciel SCAPS-1D, nous avons évalué l’effet de l’épaisseur de l’absorbeur et de la couche tampon sur les principaux paramètres photovoltaïques, à savoir la densité de courant de court-circuit (JCC), la tension en circuit ouvert (VCO), le facteur de forme (FF) et le rendement global (η). Les résultats ont montré que l’utilisation de couches tampons à large bande interdite, en particulier le Zn(S,O), améliore la transmission optique, réduit les pertes par recombinaison et conduit à des rendements plus élevés. Notamment, un rendement proche de 26 % a été atteint dans des conditions optimisées. Ce travail met ainsi en évidence l’importance cruciale du choix de la couche tampon dans l’optimisation des performances des cellules tandem, et ouvre des perspectives pour le développement de dispositifs à haut rendement et respectueux de l’environnement.
Open Access
Etude de l’influence des couches tampons pour les cellules solaires multi-jonctions
(UNIVERSITE MOHAMED BOUDIAF - M’SILA, 2025-06-11) BOUZIDI Amina; ENCA/BOUCHAMA Idris
Dans cette thèse, nous avons étudié le comportement et l’optimisation des cellules solaires tandem CIGS/Si par simulation numérique, en nous concentrant spécifiquement sur l’impact de différentes couches tampons (CdS, CdZnS, ZnS, ZnSTe, Zn(S,O)) dans la cellule supérieure en CIGS. Après avoir établi les bases théoriques de la modélisation et de la simulation à l’aide du logiciel SCAPS-1D, nous avons évalué l’effet de l’épaisseur de l’absorbeur et de la couche tampon sur les principaux paramètres photovoltaïques, à savoir la densité de courant de court-circuit (JCC), la tension en circuit ouvert (VCO), le facteur de forme (FF) et le rendement global (η). Les résultats ont montré que l’utilisation de couches tampons à large bande interdite, en particulier le Zn(S,O), améliore la transmission optique, réduit les pertes par recombinaison et conduit à des rendements plus élevés. Notamment, un rendement proche de 26 % a été atteint dans des conditions optimisées. Ce travail met ainsi en évidence l’importance cruciale du choix de la couche tampon dans l’optimisation des performances des cellules tandem, et ouvre des perspectives pour le développement de dispositifs à haut rendement et respectueux de l’environnement.
Open Access
Deep Learning Approach for Multimodal Biometric Recognition System
(University of M'Sila, 2025-04-21) NADIR Cheyma
Multimodal biometric recognition has emerged as a powerful approach to enhance recognition performance by leveraging multiple data sources. This thesis investigates unimodal and multimodal systems, focusing on finger vein (FV) and palmprint recognition, which are highly secure and reliable modalities due to their unique characteristics and resistance to forgery. Finger vein patterns, located beneath the skin, are invisible to the eye and difficult to replicate, making them ideal for high-security applications such as access control and financial transactions. Similarly, palmprint recognition relies on the distinct patterns of lines, wrinkles, and textures on the palm, which remain consistent throughout an individual’s life, even among identical twins. However, both modalities face challenges: finger vein recognition is affected by variations in lighting and image quality, while palmprint systems struggle with illumination changes, rotation, and scale variations, particularly in contactless environments. To address these challenges, we propose a series of experiments. First, we develop unimodal systems for each modality, evaluating their performance using transfer learning with three CNN models (VGG16, VGG19, and MobileNetV2). The best-performing model, MobileNetV2, is selected to extract more relevant features, and we fine-tune the top layers of the model. Additionally, we enhance the images using preprocessing techniques such as histogram equalization and contrast-limited adaptive histogram equalization (CLAHE). To account for potential issues like injuries or dirt, we propose leveraging information from both the left and right instances of each modality. This system is further evaluated by replacing the classifier with machine learning classifiers. All experiments are conducted under three protocols: Protocol 1 (P1), which uses 70% of the dataset for training and the remaining 30% for testing; Protocol 2 (P2), which uses 50% for training and the remaining 50% for testing; and Protocol 3 (P3), which uses 90% for training and the remaining 10% for testing. For finger vein recognition, we utilize the SDUMLA and FV-USM datasets, achieving first-rank accuracies of 99.57%, 97.58%, and 99.76% for SDUMLA and 99.90%, 99.60%, and 99.80% for FV-USM. For palmprint recognition, we employ the CASIA and IIT Delhi datasets, achieving accuracies of 99.13%, 99.69%, and 100% for CASIA and 99.46%, 98.50%, and 99.35% for IIT Delhi. Next, we preprocess finger vein and palmprint images to enhance system performance. For multimodal recognition, we propose a method to isolate the region of interest (ROI) from raw images, improve image quality using techniques such as CLAHE and data augmentation, and extract features from both modalities. These features are concatenated and fed into machine learning classifiers. To further improve performance, we apply score-level fusion using the outputs of MobileNetV2 and the classifiers. Experiments are conducted using the SDUMLA-HMT dataset for finger vein recognition and the IIT Delhi dataset for palmprint recognition. Three protocols (P1, P2, and P3) are evaluated. For finger vein recognition, first-rank accuracy rates reach 97.88%, 98.27%, and 99.06% for the left finger vein and 97.41%, 96.23%, and 97.64% for the right finger vein across the three protocols. Palmprint recognition achieves first-rank accuracies of 98.35%, 96.70%, and 99.06% for the left palmprint and 99.06%, 95.30%, and 99.50% for the right palmprint. The P3 protocol consistently demonstrates the highest performance. The proposed multimodal fusion approach achieves perfect recognition accuracy (100%) at both feature and score levels, significantly outperforming unimodal methods across all protocols. This superior performance is measured in precision, recall, F1 score, AUC, R1, and R5 metrics, with the fusion system consistently achieving optimal results. These findings underscore the effectiveness of our multimodal approach in enhancing the accuracy and reliability of biometric recognition systems.
Open Access
Development of a Leak Detection and Localization System Based on Parametric Models
(University of M'Sila, 2025) MEFTAH Sabir
In the core of this thesis, three methodologies were developed to address water leakage in the water distribution networks (WDNs). The first one was conducted on the development and the demonstration of a mathematical model for leak localization. It depends on the laws of fluid mechanics. Two crucial parameters that can’t be calculated, we should optimize. For that we use an evolutionary metaheuristic method, which is the biogeography-based optimization (BBO) method that has three main stages (migration, mutation, and an optional one, which is elitism). After obtaining the two unknown parameters by optimization and using the physical characteristics of the transportation pipe (length, diameter, etc.), the flow rate measurements at the two ends of the pipe, we can define the exact position of the leak. With the development of technology and the transition from the analog world to the digital world, as well as the exploitation of signal processing functions, correlators have emerged. The latter rely for their operations on two acoustic sensors placed on fire hydrants at a distance of 500 m to 1 km. The microphones or hydrophones not only transmit the leak signals but also pick up surrounding noises. The received radio frequency signals will be subjected to signal processing functions to confirm the presence of the leak and its location relative to one of the sensors. Their disadvantages lie in false alarms caused by environmental noise, thus causing the destruction of infrastructure. In addition, they require a qualified workforce. The researchers are oriented to exploit the vibration sensors and the analysis of the transient phenomena that occur. Road traffic and daily work share noises that are added to the useful signal, always causing anomalies in the infrastructure. The problem is to think of an effective and inexpensive way to solve the problem of leaks. The second contribution addresses noise and false alarms in a pressure leak detection system. Using a custom-built laboratory prototype, pressure signals were collected and denoised with a Savitzky-Golay filter. Leak localization was achieved through time-difference calculations of signal arrivals at high-precision transmitters, validated against known leak positions in a zigzag-shaped HDPE pipe network. The third contribution enhances the detection process using a larger experimental prototype. Pressure signals from leaks of varying sizes were processed using discrete wavelet transform (DWT) and Donoho thresholding for noise removal. Reconstructed signals were analyzed for quality metrics such as SNR, NCC, and MSE. Time differences in signal arrivals, combined with pressure wave velocity, allowed accurate leak localization, validated across diverse leak scenarios. This research advances the precision and robustness of leak detection methodologies, providing practical, cost-effective solutions for WDN maintenance and management.
Open Access
Applications of machine learning and deep learning in healthcare: Breast cancer case
(University of M'Sila, 2025) BEGHRICHE Tawfiq
Breast Cancer (BC) presents a significant global health challenge, highlighting the importance of timely and accurate diagnosis to improve patient outcomes. This thesis explores advanced machine learning (ML) and deep learning (DL) techniques to enhance breast cancer diagnosis, focusing on classification and detection tasks. It addresses data variability, class imbalance, and clinical applicability by developing robust diagnostic models using ultrasound imaging and clinical datasets, such as the Wisconsin Breast Cancer Dataset (WBCD) and the Coimbra dataset. Three primary frameworks for breast cancer detection are developed: (1) An ML-based approach emphasizing feature selection and class imbalance mitigation using techniques such as SMOTE and KNNOR while evaluating seven algorithms, including Support Vector Machine (SVM) and Deep Neural Networks (DNN). (2) An advanced DL-based method utilizing modified transfer learning with pre-trained convolutional neural networks (CNNs) like ResNet50 and MobileNetV2, reused on the Breast Ultrasound Images (BUSI) dataset for improved tumor classification. (3) A hybrid model combining MobileNetV2, DenseNet121, and InceptionV3 to extract features from ultrasound images, refined using LASSO-based feature selection before classification. Models performance is evaluated through metrics like accuracy, sensitivity, specificity, precision, recall, and F1-score, demonstrating the potential of ML and DL to enhance the accuracy and clinical relevance of breast cancer diagnostics, aiding radiologists and clinicians.
Open Access
Deep learning-based medical data analysis for disease prediction and classification
(University of M'Sila, 2025-05-06) Moustari Mohamed Abderaouf
The fundus images of patients with Diabetic Retinopathy (DR) often display nu- merous lesions scattered across the retina. Current methods typically utilize the entire image for network learning, which has limitations since DR abnormalities are usually localized. Training Convolutional Neural Networks (CNNs) on global images can be challenging due to excessive noise. Therefore, it's crucial to enhance the visibility of important regions and focus the recognition system on them to improve accuracy. This thesis investigates two tasks; the first one is a novel two-branch attention-guided con- volutional neural network (AG-CNN) with initial image preprocessing for DR classi- fication. The AG-CNN initially establishes overall attention to the entire image with the global branch and then incorporates a local branch to compensate for any lost dis- criminative cues. The second task is improving diabetic retinopathy classification by combining handcrafted and deep features. We extract LBP, HOG, and GLCM to cap- ture texture patterns and use DenseNet-121 for deep feature extraction. The fusion of these features enables a more comprehensive representation of the retinal images, en- hancing the model’s ability to discriminate between different severity levels of diabetic retinopathy. We conduct extensive experiments using the APTOS 2019 DR dataset for both tasks.
Open Access
Contribution à l’étude des propriétés optoélectroniques des semiconducteurs à base des éléments chalcogènes : Application photovoltaïque
(University of M'Sila, 2024-11-20) SERAI Housseyn
Actuellement, les recherches en cours se concentrent sur l'avancement des cellules solaires utilisant des éléments plus abondants et moins toxiques. Plusieurs facteurs déterminent l'efficacité des cellules solaires dans la conversion de la lumière solaire en énergie. Les matériaux photovoltaïques avec des bandes interdites optimales pour la conversion photovoltaïque, des coefficients d'absorption élevés, une stabilité dans diverses situations, une rentabilité et un impact environnemental minimal sont privilégiés. Les matériaux à base de chalcogénures de cuivre, en particulier les composés quaternaires tels que les CZTS, ont été étudiés de manière approfondie en raison de leurs propriétés électriques favorables, de leur évolutivité rentable et de leur présence abondante dans l'écorce terrestre. Cette étude a pour objectif d'examiner la structure de type kesterite des composés Cu2BeSnS4, Cu2BeSnSe4 et Cu2BeSnTe4 utilisant la théorie de la fonctionnelle de la densité (DFT) et la méthode de l'onde plane augmentée linéaire à potentiel complet (FP-LAPW). L'étude démontre que les composés Cu2BeSnS4 et Cu2BeSnSe4 sont des semi-conducteurs avec des gaps d’énergies directs au point Γ, tandis que Cu2BeSnTe4 présente un gap indirect (Γ→X). Les propriétés électroniques et optiques de ces matériaux indiquent leur utilité potentielle dans diverses applications optoélectronique. Ainsi, les résultats fournissent des informations précieuses sur les applications photovoltaïques possibles de ces composés. En outre, nous avons exploré la possibilité de moduler ces propriétés dans le cadre de la proposition et de la conception de nouvelles cellules solaires utilisant ces composés comme base.
Open Access
Big Data and Artificial Intelligence for Improving the Performance and Efficiency of Large-Scale Grid-Connected PV Power Plant
(University of M'Sila, 2024-11-17) Amiri Ahmed Faris
This thesis presents reliable methods for fault detection and diagnosis in Photovoltaic (PV) systems. The first method proposes a two step approach for developing a reliable PV model and constructed a fault detection procedure using Random Forest Classifiers (RFCs). The first step involves identifying the unknown parameters of the One Diode Model (ODM) using the Modified Grey Wolf Optimization (MGWO) algorithm and simulating the PV array t o extract maximum power point (MPP) coordinates. The second step involves developing two RFCs: one for fault detection and another for fault diagnosis. The second method uses the Sandia Array Performance Model (SAPM) for accurate photovoltaic system si mulation. Parameters are extracted with the Artificial Bee Colony (ABC) algorithm to optimize and reduce errors between measured and simulated data. Additionally, deep learning is employed by combining Convolutional Neural Networks (CNN) and Bidirectional Gated Recurrent Units (Bi GRU) to analyze dynamic system power outputs at the MPP for fault detection and diagnosis with high precision. The third work develops a predictive modeling method for PV generation using rigorous feature selection, outlier re moval, and hyperparameter tuning. The method is implemented in a MATLAB interface to predict key parameters and evaluate system performance. The efficiency of these methods is evaluated using real data from actual PV systems.
Open Access
Piégeage de la lumière pour le contrôle de l’absorption dans les cellules solaires photovoltaïques
(University of M'Sila, 2025-02-19) CHAHMI Maroua
En résumé de notre travail, des simulations optoélectriques couplées ont été réalisées pour étudier les propriétés optiques et électriques des cellules solaires à base de pérovskite PSC. Tout d'abord, l'effet des matériaux utilisés des couches HTL et ETL sur les caractéristiques optiques de la cellule a été étudié. Les pertes de lumière dans une architecture planaire ont été également étudiées, et une cellule solaire à base de nanostructures sur un substrat d’Or et on a changé avec d'autres métaux a été proposée pour atténuer ces pertes. Les observations ont révélé que l'utilisation de PDMS au lieu de verre dans la structure proposée diminue les pertes de réflexion globales et l’effet du grating pour piéger la lumière. En outre, le phénomène de piégeage de la lumière dans la couche active entraînerait une amélioration de l'absorption et de la diffusion SPR en champ lointain. L'absorption de la pérovskite dans les structures proposées lorsqu’on utilise un cristal photonique 1D (Ge/SiO2) représentées par trois structures différentes avec variation des épaisseurs de ce cristal, et les PSC présentent une efficacité considérable. Les caractéristiques courant-tension ont révélé des améliorations de JCC et VCO. Le rendement de conversion de puissance du PSC est passé de 14,03 % à 26,27 %. Les données de simulation décrivent que le PCE été augmenté jusqu'à 90 % par rapport à la structure planaire est la preuve d'une augmentation significative de l'efficacité.
Open Access
Commande d'une Machine Polyphasée en Modes Normal et Dégradé
(University of M'Sila, 2024-04-18) FODIL MALIKA
Operation at low speed and high torque can lead to the generation of strong ripples in the speed, which can deteriorate the system. To reduce the speed oscillations when operating at the low speed of a five-phase asynchronous motor. In this work, we propose a control method based on metaheuristic optimization algorithms such as Grey Wolf (GWO) and Particle Swarm Optimization (PSO), to adjust the parameters of proportional-integral (PI) controllers. Proportional-integral controllers are commonly used in control systems to regulate the speed and current of a motor. The controller parameters, such as the integral gain and proportional gain, can be adjusted to improve the control performance. Specifically, reducing the integral gain can help reduce the oscillations at low speeds. The proportional-integral controller is insensitive to parametric variations, but by the use of meta-heuristic optimization strategies, we may choose gains wisely, and the system becomes more reliable. The obtained results show that the hybrid control of the five phases IM offers high performance in the permanent and transient states. In addition, with this proposed strategy controller, disturbances do not affect machine performance.
Open Access
Study, fabrication and characterization of thin films for CZTS solar cells
(University of M'Sila, 2024-05-09) Ali-Saoucha Salim
This study primarily investigates CZTS thin-film solar cells. We employ numerical modeling through Silvaco Atlas and AMPS-1D software to analyze CZTS solar cells, including substrate, bifacial, and tandem configurations. A notable aspect of our research involves proposing an innovative design for CZTS bifacial solar cells. We specifically examine the influence of the barrier height at the p-CZTS/n-TCO interface in the structure ZnO:Al/CdS/CZTS/TCO/SLG. The efficiency results reveal 15.8% under front-side illumination and 4% under back-side illumination. In the subsequent section, we utilize the Density Functional Theory (DFT) method to scrutinize the structure, electronic, and optic properties of the ZnS1-xOx alloy. Shifting our focus to experimentation, we employ the Sol-Gel spin-coating method to fabricate pure ZnO, sulfur-doped ZnO, and ZnS thin films. The spin-coated films are characterized by analyzing their structural, morphological, and optical properties through techniques such as X-ray diffraction (XRD), Atomic Force Microscopy (AFM), and UV-VIS spectroscopy.
Open Access
Piégeage de la lumière pour le contrôle de l’absorption dans les cellules solaires photovoltaïques
(UNIVERSITE MOHAMED BOUDIAF - M’SILA, 2025-05-26) CHAHMI Maroua
En résumé de notre travail, des simulations optoélectriques couplées ont été réalisées pour étudier les propriétés optiques et électriques des cellules solaires à base de pérovskite PSC. Tout d'abord, l'effet des matériaux utilisés des couches HTL et ETL sur les caractéristiques optiques de la cellule a été étudié. Les pertes de lumière dans une architecture planaire ont été également étudiées, et une cellule solaire à base de nanostructures sur un substrat d’Or et on a changé avec d'autres métaux a été proposée pour atténuer ces pertes. Les observations ont révélé que l'utilisation de PDMS au lieu de verre dans la structure proposée diminue les pertes de réflexion globales et l’effet du grating pour piéger la lumière. En outre, le phénomène de piégeage de la lumière dans la couche active entraînerait une amélioration de l'absorption et de la diffusion SPR en champ lointain. L'absorption de la pérovskite dans les structures proposées lorsqu’on utilise un cristal photonique 1D (Ge/SiO2) représentées par trois structures différentes avec variation des épaisseurs de ce cristal, et les PSC présentent une efficacité considérable. Les caractéristiques courant-tension ont révélé des améliorations de JCC et VCO. Le rendement de conversion de puissance du PSC est passé de 14,03 % à 26,27 %. Les données de simulation décrivent que le PCE été augmenté jusqu'à 90 % par rapport à la structure planaire est la preuve d'une augmentation significative de l'efficacité.
Open Access
Ab-initio and artificial intelligence based methods for materials physical properties prediction
(University of M'Sila, 2024) BOUZATEUR Inas
Creating novel and inexpensive compounds that can meet both current and projected needs is becoming increasingly important due to the rapid speed of industrial change. However, traditional methods of rationally finding new materials with a specific set of features have grown challenging and costly because of the rise in material structural and functional complexity. This crucial fact has paved the way for the application of intelligence methods in this field. This dissertation focuses on two fundamental axes of research: The first focus on developing an intelligent prediction and identification technique for various material properties, and the second is on using the DFT theory to calculate and analyze the lattice parameter and band gap energy properties of the proposed compound, Ba2BiTaS6. To achieve this goal, we have addressed several points: First, we came up with a new model that uses artificial neural networks (ANN) and the particle swarm optimization algorithm (PSO) to get rid of problems with local minima in ANN models while keeping the quality of the fitting. By predicting the band gap properties, this method speeds up the search for new chalcopyrite materials in photovoltaic solar cells with better resolution. The model has two separate parts: an ANN sub-system that makes predictions using low-resolution training data; and an "error model" sub-system that was added to deal with resolution quality issues and show uncertainty in the primary model. Furthermore, we presented a comparative analysis of optimization algorithms to understand and quantify ANN's performance in guiding the search process towards better solutions over all feasible solutions. We used an efficient technique based on ANN-PSO and fuzzy logic-PSO to predict the lattice constants of pseudo-cubic and cubic perovskites. We predicted the lattice parameter of double perovskite compounds using the extreme learning machine (ELM). Finally, we used the FP-LAPW method in the WIEN2k environment, which is based on the DFT theory, to calculate and analyze the lattice parameter and the band gap energy properties of Ba2BiTaS6
Open Access
Development of a Leak Detection and Localization System Based on Parametric Models
(University Mohamed Boudiaf - M’sila, 2025-05-25) MEFTAH Sabir
In the core of this thesis, three methodologies were developed to address water leakage in the water distribution networks (WDNs). The first one was conducted on the development and the demonstration of a mathematical model for leak localization. It depends on the laws of fluid mechanics. Two crucial parameters that can’t be calculated, we should optimize. For that we use an evolutionary metaheuristic method, which is the biogeography-based optimization (BBO) method that has three main stages (migration, mutation, and an optional one, which is elitism). After obtaining the two unknown parameters by optimization and using the physical characteristics of the transportation pipe (length, diameter, etc.), the flow rate measurements at the two ends of the pipe, we can define the exact position of the leak. With the development of technology and the transition from the analog world to the digital world, as well as the exploitation of signal processing functions, correlators have emerged. The latter rely for their operations on two acoustic sensors placed on fire hydrants at a distance of 500 m to 1 km. The microphones or hydrophones not only transmit the leak signals but also pick up surrounding noises. The received radio frequency signals will be subjected to signal processing functions to confirm the presence of the leak and its location relative to one of the sensors. Their disadvantages lie in false alarms caused by environmental noise, thus causing the destruction of infrastructure. In addition, they require a qualified workforce. The researchers are oriented to exploit the vibration sensors and the analysis of the transient phenomena that occur. Road traffic and daily work share noises that are added to the useful signal, always causing anomalies in the infrastructure. The problem is to think of an effective and inexpensive way to solve the problem of leaks. The second contribution addresses noise and false alarms in a pressure leak detection system. Using a custom-built laboratory prototype, pressure signals were collected and denoised with a Savitzky-Golay filter. Leak localization was achieved through time-difference calculations of signal arrivals at high-precision transmitters, validated against known leak positions in a zigzag-shaped HDPE pipe network. The third contribution enhances the detection process using a larger experimental prototype. Pressure signals from leaks of varying sizes were processed using discrete wavelet transform (DWT) and Donoho thresholding for noise removal. Reconstructed signals were analyzed for quality metrics such as SNR, NCC, and MSE. Time differences in signal arrivals, combined with pressure wave velocity, allowed accurate leak localization, validated across diverse leak scenarios. This research advances the precision and robustness of leak detection methodologies, providing practical, cost-effective solutions for WDN maintenance and management.
Open Access
Deep learning-based medical data analysis for disease prediction and classification
(University of Mohamed Boudiaf - M’sila, 2025-05-25) Moustari Mohamed Abderaouf
Les imagesdufondd’oeildespatientsatteintsderétinopathiediabétique(RD)présen- tent souventdenombreuseslésionsdisperséessurlarétine.Lesméthodesactuelles utilisent généralementl’imageentièrepourl’apprentissageduréseau,cequiprésente des limitespuisquelesanomaliesdelaRDsontgénéralementlocalisées.Laformation de réseauxneuronauxconvolutionnels(CNN)surdesimagesglobalespeutêtrediffi- cile enraisondubruitexcessif.Parconséquent,ilestcruciald’améliorerlavisibilité des régionsimportantesetdeconcentrerlesystèmedereconnaissancesurellespour améliorer laprécision. Cette thèseétudiedeuxtâches;lapremièreestunnouveauréseauneuronalcon- volutionnelguidéparl’attentionàdeuxbranches(AG-CNN)avecprétraitementinitial de l’imagepourlaclassificationdelaRD.L’AG-CNNétablitd’abordl’attentionglob- ale surl’imageentièreaveclabrancheglobale,puisintègreunebranchelocalepour compenser leséventuelsindicesdiscriminantsperdus. La deuxièmetâcheconsisteàaméliorerlaclassificationdelarétinopathiediabétique en combinantdescaractéristiquesartisanalesetprofondes.NousextrayonsLBP,HOG et GLCMpourcapturerlesmotifsdetextureetutilisonsDenseNet-121pourl’extraction de caractéristiquesprofondes.Lafusiondecescaractéristiquespermetunereprésenta- tion pluscomplètedesimagesrétiniennes,améliorantainsilacapacitédumodèleà distinguer lesdifférentsniveauxdegravitédelarétinopathiediabétique. Nous menonsdesexpériencesapprofondiesenutilisantl’ensemblededonnéesAP- TOS2019DRpourlesdeuxtâches.
Open Access
Applications of machine learning and deep learning in healthcare: Breast cancer case
(Université Mohamed Boudiaf - M’sila, 2025-05-25) BEGHRICHE Tawfiq
Breast cancerrankssecondinfatalityamongwomen.Conventionaldiagnosticmethodsaretime-consuming,exhausting,and expensive,potentiallyleadingtodelayedtreatmentormisdiagnosis.Machinelearning(ML)anddeeplearning(DL)methods haveshownoutstandingpotential.However,theyfacechallengeslikefeatureidentificationandselection,dataimbalance, and parameteroptimization.Unlikemostknownsolutionsthathaveexploredthesestagesforbreastcancerprediction,either separately orinlimitedcombinations,ourapproachsimultaneouslytacklesthesecriticalissuesusingmulti-stageoptimization architecture. Threewell-establishedtechniques,namelycorrelationanalysis-basedfeatureselection(CFS),LASSOregres- sion, andmutualinformation(MI),areusedforFS.Databalancingisperformedusingbothoversamplingandundersampling techniques, includingthesyntheticminorityoversamplingtechnique(SMOTE),k-nearestneighboroversampling(KNNOR), and randomundersampling(RUS).Finally,hyperparameteroptimization(HPO)iscarriedoutbyadoptingvariousmethods including gridsearch,randomsearch,Bayesianoptimization,andsemi-automatictomaximizetheclassificationperfor- mance ofsevenrenownedMLalgorithms(logisticregression,decisiontree,randomforest,supportvectormachine,Naïve Bayes, k-nearestneighbor,andeXtremegradientboosting),andadeepneuralnetwork(DNN).Throughtheexperiments carried outonfourpubliclyavailabledatasets,includingWisconsindiagnosticbreastcancer(WDBC),Wisconsinbreast cancer dataset(WBCD),Wisconsinprognosticbreastcancer(WPBC),andBreastcancercoimbra(BCC),theobtainedresults clearly demonstratethesuperiorityoftheproposedmethodoverthestate-of-the-artmethods.
Open Access
La segmentation d’images par les modèles de Markov cachés (HMM) pour l’identification biométrique
(Université Mohamed Boudiaf - M'sila, 2025-05-25) Djalab Abdelhak