https://dspace.univ-guelma.dz/jspui/handle/123456789/32 2026-10-16T15:29:12Z https://dspace.univ-guelma.dz/jspui/handle/123456789/19776 Titre: Towards intelligent routing approaches for field service operations Auteur(s): Bouchbout, Yehya Résumé: Real-world logistics and field-service operations are considerably more demanding than the textbook static Vehicle Routing Problem. Two classes of complexity that remain largely open in the multi-depot setting are priority constraints, the hard operational requirement that certain customers must be served regardless of cost, and dynamism, the arrival of customer requests after the initial routing plan has been constructed. A third difficulty concerns the solvers themselves: the performance of metaheuristics depends strongly on their control parameters, yet standard practice applies a single global configuration across all depots, disregarding the structural heterogeneity of sub-problems. This thesis addresses all three challenges through a sequence of contributions grounded in real Algerian logistics data. The first contribution is a foundational case study of Algérie Poste’s single depot mail delivery in Guelma. A hybrid Ant Colony Optimisation and Variable Neighbourhood Search (ACO+VNS) solver, in two hybridisation strategies, reduces total travel distance by up to 16.17% over the company’s manual routing; a Bayesian tuned extension (BO-ACO+VNS) then makes the solver self-configuring: in the published study it gave 10.5% shorter routes with 66.3% less computation than the hybrid with literature-default parameters, and a stricter controlled re-run confirms 3.6% shorter routes at 31% less total computing time. This study provides the metaheuristic and the automated tuning reused in the field-service chapters, and leads to the multi-depot, priority-sensitive model that follows. The second and central contribution introduces the Multi-Depot VRP with Priority Requests (MDVRP-PR), the first MDVRP formulation in which priority is enforced as a lexicographic admissibility constraint: solutions leaving any priority customer unserved are structurally inadmissible, regardless of the travel-time savings they might offer. A three-stage algorithmic pipeline is proposed: (i) a hybrid DBSCAN-Hierarchical clustering algorithm for topology-aware, capacity-balanced depot assignment; (ii) an Enhanced Max-Min Ant System (MMAS) with a priority-first construction rule and lexicographic solution selection; and (iii) a Boundary Relocate post-optimisation operator for cross-depot correction. The framework is validated on a real Algérie Télécom field-service instance (4 depots, 55–150 customers) over 2290 controlled runs, achieving 100% priority feasibility at a route-time overhead below 5% and improving customer coverage by 16.1 percentage points over current manual practice in about 28 seconds. The third contribution extends the static MDVRP-PR to the dynamic setting. RBDI (Regret-Based Dynamic Insertion) is a single-pass insertion module that handles customer requests arriving after route construction, combining: a cross depot k=3 regret framework that evaluates insertion opportunities across all depots simultaneously; an arrival-time-aware feasibility check that eliminates the “insertion into-the-past” artifact of static-snapshot heuristics; a priority-first selection rule that inserts every priority customer whenever the fleet can feasibly accommodate it; and post-insertion local search (2-opt + Or-opt) for route quality recovery; an anticipatory capacity reservation then protects route slack for priority customers who have not yet arrived. Evaluated on five Cordeau MDVRP benchmarks and the real-world Algérie Télécom dataset, RBDI serves 94–99% of dynamic requests and, through a formal never-strand guarantee, serves every reachable priority customer (priority-customers-served above 94% on the Cordeau benchmarks even under heavy dynamism); it operates 350–1700× faster than an ALNS re-optimisation baseline at a route-time cost of up to 15%, matching or beating that baseline on priority feasibility on six of eight instances. The fourth contribution addresses automated algorithm configuration. PBO (Pooled Bayesian Optimisation) fits a single Gaussian Process surrogate over the joint space of depot feature vectors and MMAS hyperparameters, pooling observations across structurally heterogeneous depot sub-problems. PBO is evaluated on 15 Cordeau benchmarks and the Algérie Télécom instances through six controlled experiments spanning tuning effectiveness, sample efficiency, zero-shot cross-instance transfer, feature ablation, and real-world deployment. Results show that Bayesian tuning cuts route time by up to 18.1% over literature defaults, with the gain concentrated in large per-depot sub-problems: it is inert below 40 customers per depot and reaches 6–18% only on the largest; that pooling makes tuning robust to over-tuning, reaching its plateau within 3–5 evaluations per depot; and that a configuration pooled offline transfers zero-shot to unseen instances, recovering about 92% of the gain of tuning on the instance itself. On the real instances, service time dilutes the gain to about 1% even above the size threshold. An ablation study shows that the benefit comes entirely from pooling observations across depots in the sharedGP; the depot features themselves bring no gain and even reduce cross-instance transfer. The four contributions thus go from an application study, to a new problem formulation and its static solver, to a dynamic extension, and finally to automated algorithm configuration for multi-depot field-service routing. 2026-10-05T00:00:00Z https://dspace.univ-guelma.dz/jspui/handle/123456789/19662 Titre: Towards the Improvement of the Healthcare System in Algeria Based on Internet of Things Technologies Auteur(s): AOUAMRIA, Selma Résumé: This thesis investigates the transition toward proactive smart healthcare, in which continuous sensing, near-real-time analytics, and clinically actionable indicators support earlier detection and long-term management of chronic diseases. This shift is especially relevant to Algeria, given persistent structural constraints, geographic disparities in access to care, and the growing burden of chronic illness. The thesis positions the Internet of Health Things (IoHT) as a practical foundation for remote monitoring and early warning, while highlighting key barriers to adoption, including heterogeneous data sources, interoperability gaps, and stringent privacy requirements that limit centralized data sharing. To address these challenges, the thesis leverages artificial intelligence and deep learning to extract predictive signals from clinical and IoHT data. It further argues that single-model approaches may lack robustness across datasets and care settings; accordingly, Ensemble Learning is adopted to enhance predictive stability and generalization. The first contribution proposes an Ensemble Deep Learning framework for early risk prediction, particularly for chronic conditions such as diabetes. The framework integrates multiple deep models through probabilistic decision fusion based on weighted soft voting and is evaluated using clinically meaningful metrics to reduce medically consequential misclassification. The second contribution introduces a secure and privacy-preserving federated learning framework for ECG classification under non-IID data distributions. The framework combines permissioned blockchain-based coordination with IPFS-based decentralized storage and proposes FedLA-Prox, an aggregation strategy that jointly incorporates label-aware aggregation, to mitigate class imbalance and majority-class dominance, and proximal regularization, to reduce client drift. This design achieves smoother convergence, improved fairness toward minority classes under severe and moderate heterogeneity, and no performance degradation in near-IID settings, while maintaining feasible coordination overhead in the system-level evaluation. 2026-09-19T00:00:00Z https://dspace.univ-guelma.dz/jspui/handle/123456789/19089 Titre: Data Indexing in smart grid networks Auteur(s): BRAHMIA, Abdelbacet Résumé: The transformation of conventional electrical networks into smart grids has led to the large-scale deployment of smart meters, sensors, and Internet of Things (IoT) devices. These technologies continuously generate massive volumes of operational data related to electricity consumption, environmental conditions, and grid behavior. Managing and analyzing this data efficiently has become a major challenge for modern energy systems. In particular, smart grid data are mainly represented as time-series data, where electricity consumption evolves over time and is influenced by multiple environmental and contextual factors. Traditional data management and indexing techniques are not well suited for handling such large, dynamic, and multidimensional datasets. This thesis addresses the problem of efficient organization and retrieval of smart grid time-series data. The objective is to design indexing structures capable of supporting fast data retrieval, scalable storage, and real-time analysis while preserving the relationships between temporal patterns and electricity consumption behavior. First, a statistical analysis of environmental and temporal variables influencing electricity consumption is conducted using Principal Component Analysis (PCA) and Analysis of Variance (ANOVA). This analysis identifies the most relevant features affecting electricity demand and provides a statistical foundation for constructing multidimensional indexing structures. Second, a multidimensional indexing approach based on feature engineering and KD-Tree structures is proposed. Environmental variables are normalized, and temporal attributes are encoded using cyclic transformations to preserve their periodic nature. The resulting multidimensional representation allows observations to be organized in a unified feature space, enabling efficient similarity search and faster retrieval compared to traditional sequential scanning. Third, a hybrid tree-based indexing model is introduced for large-scale smart grid time-series data. This structure integrates temporal segmentation and symbolic representation using Piecewise Aggregate Approximation (PAA) and iSAX concepts to reduce dimensionality while maintaining important consumption patterns. The hierarchical organization improves scalability and supports efficient search operations. Finally, a new indexing structure called the Temporal Consumption Binary Tree (TCB-Tree) is proposed. This model organizes smart grid data simultaneously along temporal and consumption dimensions using horizontal and vertical segmentation strategies. The structure supports time-based, consumption-based, and hybrid queries within a unified framework, enabling efficient data retrieval in large smart grid datasets. Experimental evaluations demonstrate that the proposed indexing approaches improve retrieval efficiency, scalability, and query performance when compared to traditional methods. These results highlight the importance of combining feature engineering, multidimensional representation, and specialized indexing structures for efficient smart grid data management. 2026-06-28T00:00:00Z https://dspace.univ-guelma.dz/jspui/handle/123456789/19088 Titre: IoT-Based Traceability and AI-Enhanced Techniques: Case Studies on Livestock Monitoring and Grain Management in the Food Supply Chain Auteur(s): BENYOUNES, Ali Résumé: The increasing complexity of the global food supply chain requires intelligent traceability systemscapable of ensuring transparency, efficiency, and sustainability across its critical stages. Amongthese stages are production and storage. This thesis evaluates the traceability enhancement using integrated IoT-AI frameworks, illustrated through two complementary case studies: livestockmonitoring and grain storage management. On the animal production side, the work introduces CowStrus, an automated estrus detectionsystem designed for dairy farms. The proposed pipeline combines data acquisition using fixedcameras and drones, the construction of a near top-view dairy cow dataset, and transfer learning with YOLOv8 for cow body-part detection. A selected interaction features used within arule-based decision tree to classify sexual behaviors, followed by a scoring mechanism for reliableestrus detection. Complementary contributions include the integration of multi-scale attentionmechanisms within YOLOv5 and DeepLabv3+, improving detection and segmentation performance, particularly for small and overlapping parts such as the tail. These advances further enabled the development of an early detection system for vaginal infections through tail movement analysis. On the vegetal side, the thesis analyzes the limitations of existing IoT-based grain storage systems and proposes a hierarchical Edge-Fog-Cloud architecture for distributed and scalablemonitoring. In addition, a self-supervised learning pipeline is introduced for grain quality classification, reducing annotation dependency while improving generalization capability. Together, these contributions demonstrate how the strategic integration of IoT infrastructures and AI-enhanced perception techniques can strengthen traceability, improve decision-making, and advance intelligent management across key stages of the food supply chain. 2026-06-28T00:00:00Z