Archive ouverte de l'ENAC
Not a member yet
3458 research outputs found
Sort by
Machine Learning et Interfaces 2D/3D pour automatiser,interagir avec et visualiser la configuration d'un MicroscopeÉlectronique en Transmission.
No full textThe aim of this thesis is to respond to the need for automation and understanding of the electron microscope by its operator. Indeed, a transmission electron microscope is a complex instrument to operate, which is also difficult to get a coherent mental representation of, especially for a novice microscopist. To address this issue of electron microscope accessibility, this thesis focuses on automating microscope alignment using new machine learning techniques such as CNN or RL. On the other hand, we are interested in the use of interactive and immersive visualization interfaces to enable the operator to visualize how modifications to the TEM configuration impact the electron beam within it, and thus facilitate the construction of a mental model of the microscope. the construction of a mental model of the tool that is consistent with reality and more direct than the changes visible on the image. To this end, we have developed a modular TEM simulation and two visualization interfaces, one 2D and one 3D for (Meta Quest, Android and Microscoft Hololens).L'objectif de cette thèse est de répondre au besoin d'automatisation et de compréhension du microscope électronique par son opérateur. En effet, un microscope électronique à transmission est un instrument complexe à utiliser, dont il est également difficile d'avoir une représentation mentale cohérente, en particulier pour un microscopiste novice. Pour répondre à ce problème d'accessibilité du microscope électronique, cette thèse se concentre sur l'automatisation de l'alignement des microscopes en utilisant de nouvelles techniques d'apprentissage automatique telles que CNN ou RL. D'autre part, nous nous intéressons à l'utilisation d'interfaces de visualisation interactives et immersives pour permettre à l'opérateur de visualiser l'impact des modifications de la configuration du TEM sur le faisceau d'électrons qui s'y trouve, et ainsi faciliter la construction d'un modèle mental du microscope. La construction d'un modèle mental de l'outil cohérent avec la réalité et plus direct que les changements visibles sur l'image. Pour ce faire, nous avons développé une simulation modulaire du MET et deux interfaces de visualisation, l'une 2D et l'autre 3D pour (Meta Quest, Android et Microscoft Hololens
Star-Burst paradigm: implementation of an "invisible" dry-EEG reactive BCI
No full textInternational audienc
SMARTS Solution for Airspace Design and Configuration
No full textInternational audienceThis paper presents the SMARTS solution for airspace design and dynamic configuration in European air traffic management. The approach addresses the critical need for increased airspace capacity through intelligent sector management using artificial intelligence and optimization algorithms. The SMARTS solution comprises three main components: Basic Volume Design; Sector Design employing mixed integer linear programming for optimal sector boundaries; and Dynamic Configuration using graph-based optimization for near real-time configuration planning. Validation was conducted at Madrid Area Control Center (ACC) using 2023 traffic data. The results show that the SMARTS solution successfully balances air traffic controller workload, improves capacity utilization, and introduces a local resilience KPI for monitoring robustness toward a reliable framework for future European airspace architecture. The system delivers operationally viable configurations that align with real-world air traffic management needs while maintaining flexibility for various operational scenarios.</div
Downlink Optimization for Direct-to-Satellite IoT with LEO Satellites and LoRaWAN
No full textInternational audienceDirect-to-satellite communication systems for the Internet of Things, particularly those based on low-Earth-orbit satellite constellations, are emerging as a transformative solution to achieve global connectivity. However, ensuring efficient and reliable downlink communication from satellites to ground-based IoT devices remains a significant challenge due to intermittent satellite visibility, short contact durations, limited bandwidth, device energy constraints, and high network density. Unlike prior studies that primarily focus on uplink optimization, this work proposes a downlink-aware optimization framework that integrates satellite dynamics, LoRaWAN MAC constraints, and energy-aware scheduling. The framework accounts for physical-layer limitations, satellite visibility modeling, time-slot feasibility, and realistic system parameters consistent with LEO satellite operations. Simulations demonstrate that the proposed downlink-aware optimization framework improves the packet delivery ratio from 0.41 (achieved under random scheduling) to 0.96, while reducing the average energy consumption per successful transmission by approximately 55%. These results highlight the efficiency of the proposed NSGA-II-based scheduling approach and provide an initial pattern that points toward potential scalability, compared to conventional non-optimized methods, demonstrating its promise for next-generation satellite-enabled IoT networks
Hierarchical Planning Applied to the Preliminary Design of CubeSats: Nanospace Study
No full textInternational audienceubeSat design has been already studied and formalized but knowledge representation remains a challenge. The management of human‐learnt knowledge during the process is not an aspect that is often spoken about. This paper discusses the proposal of integrating a hierarchical planning approach with a model‐based one to the open‐source Nanospace framework, a web‐based application for concurrent engineering during the preliminary design phase of CubeSats. Hierarchical planning aids to introduce commonly tacit human expertise, an aspect that the preliminary design of CubeSats can benefit from. The proposed integration could allow a faster design convergence and faster inspection of candidate architectures The Nanospace framework itself may benefit from an approach bringing in model‐based efforts and hierarchical planning facilitate knowledge representation and reuse
Regional-Scale Multi-Modal Service Optimization for Innovative Air Mobility
No full textInternational audienceBy 2050, ninety percent of European travelers are expected to complete door-to-door journeys within 4 hours, requiring expanded transport infrastructure coverage, particularly in rural areas where 19.2% of the population resides. Innovative Air Mobility (IAM), which enhances connectivity with minimal infrastructure, offers a promising solution. Building on this potential, this paper presents a passenger-centric IAM service optimization framework that models the multi-modal network as a time-expanded graph and sequentially optimizes passenger journeys and IAM dispatching strategies using A* search and dynamic programming. A simulation based on the Catalan Pyrenees validates the framework, demonstrating its effectiveness in generating optimal IAM dispatch strategies and supporting multi-modal schedule coordination. Based on the findings, this paper offers recommendations for future IAM development, including per-vehicle capacity planning, pricing strategies to avoid direct competition with existing transport modes, and other operational considerations.</div
La géométrie des réseaux de neurones : une approche de la robustesse sous l'angle des feuilletages riemanniens.
No full textIn this thesis we employ the tools of geometry and statistics to shed light on the relationship between data and neural network predictions. In particular, we take inspiration from the field of information geometry combines precisely these two approaches. We see the neural network's output as the parameter of a probability distribution. By using the Data Information Matrix (DIM), a variation of the Fisher Information Matrix (FIM), we investigate the network's input/output relationship and reveal its understanding of the data structure. This statistical framework yields a (degenerate) Riemannian metric that we use to analyze the geometry of the data. In particular, we lean on a foliation arising from the kernel of the DIM to conduct our study of the low dimensional data in the high dimensional input space. We first employ the language of Cartan moving frames to compute the connection and curvature of the DIM in a efficient way, and give some explanations on the response of a neural network through experiments. Unfortunately, in most practical cases of machine learning, the DIM has a non constant rank and is non smooth, making it difficult to yield a well defined foliation. To tackle this issue, we prove that for usual neural network architectures, this only happens in a nowhere dense set. Besides, we investigate these singularities as they may teach us about distances between datasets and efficiency of knowledge transfer. Finally, we apply this new geometrical framework given by the DIM to the analysis of the robustness of neural networks. We show that the curvature of the transverse to the kernel leaves can be utilized to improve adversarial attacks, indicating that the geometry of the data is key in the robustness of machine learning algorithms.Dans cette thèse, nous utilisons les outils de la géométrie et de la statistique pour mettre en lumière la relation entre les données et les prédictions des réseaux neuronaux. En particulier, nous nous inspirons du domaine de la géométrie de l'information qui combine précisément ces deux approches. Nous considérons la sortie du réseau neuronal comme le paramètre d'une distribution de probabilité. En utilisant la matrice d'information des données (DIM), une variante de la matrice d'information de Fisher (FIM), nous étudions la relation entrée/sortie du réseau et analysons sa compréhension de la structure des données. Ce cadre statistique produit une métrique riemannienne (dégénérée) que nous utilisons pour analyser la géométrie des données. En particulier, nous nous appuyons sur un feuilletage issu du noyau de la DIM pour mener notre étude des données de faible dimension dans l'espace d'entrée de haute dimension. Nous utilisons d'abord le langage des repères mobiles de Cartan pour calculer la connexion et la courbure de la DIM de manière efficace, et nous donnons quelques explications sur la sortie d'un réseau de neurones à travers des expériences numériques. Malheureusement, dans la plupart des cas pratiques d'apprentissage automatique, la DIM a un rang non constant et n'est pas lisse, ce qui rend difficile l'obtention d'un feuilletage bien défini. Pour résoudre ce problème, nous prouvons que pour les architectures de réseaux de neurones habituelles, cela ne se produit que dans un ensemble dense nulle part. En outre, nous étudions ces singularités car elles peuvent nous renseigner sur les distances entre les ensembles de données et sur l'efficacité du transfert de connaissances. Enfin, nous appliquons ce nouveau cadre géométrique donné par la DIM à l'analyse de la robustesse des réseaux de neurones. Nous montrons que la courbure de l'espace transverse aux feuilles du noyau peut être utilisée pour améliorer les attaques adversaires, ce qui indique que la géométrie des données est un élément clé de la robustesse des algorithmes d'apprentissage automatique
SMARTS solution for airspace design and configuration
No full textInternational audienceThis paper presents the SMARTS solution for airspace design and dynamic configuration in European air traffic management. The approach addresses the critical need for increased airspace capacity through intelligent sector management using artificial intelligence and optimization algorithms. The SMARTS solution comprises three main components: Basic Volume Design; Sector Design employing mixed integer linear programming for optimal sector boundaries; and Dynamic Configuration using graph-based optimization for near real-time configuration planning. Validation was conducted at Madrid Area Control Center (ACC) using 2023 traffic data. The results show that the SMARTS solution successfully balances air traffic controller workload, improves capacity utilization, and introduces a local resilience KPI for monitoring robustness toward a reliable framework for future European airspace architecture. The system delivers operationally viable configurations that align with real-world air traffic management needs while maintaining flexibility for various operational scenarios
Experiment to test one of the incompleteness of quantum mechanics
No full textFor nearly 100 years, the incompleteness of quantum formalism and the probabilistic nature ofmeasurement have been the subject of ongoing debate, with no interpretation achieving unanimousagreement.In double-slit interference experiments, standard quantum theory does not take particle sizeinto account, which is not the case in de Broglie’s double-solution theory. We use the large sizeof Rydberg atoms to propose an experiment to test the incompleteness of the standard quantumformalism.We present a variation on the double-slit experiment performed with Rydberg sodium atoms,in which a grating of very narrow slits is added between the two slits. Rydberg atoms are toobig and cannot pass through the slits of the grating. We show with numerical simulations thatthe transmission densities in the standard interpretation and in the double-solution interpretationgive very different results (a dark band appears in the center of the pattern). Experimentalimplementation now seems possible and would be a crucial test between these two interpretations
Évaluation multidisciplinaire de l'impact de l'augmentation de la température et du vent de face sur les performances au décollage des aéronefs.
No full textInternational audienceThe projected rise in global temperatures and evolving wind patterns are expected to increasingly challenge aviation operations, particularly during the take-off phase, when aircraft performance is highly sensitive to ambient atmospheric conditions. In this study, a multidisciplinary framework is developed to quantify the combined impact of temperature rise and headwind variability on take-off field length (TOFL) at 60 major global airports. Climate projections are drawn from a multi-model ensemble of 26 bias-corrected members across six CMIP6 models, using the CDFt bias correction method with historical ERA5 reanalysis as reference. The correction is validated over the historical period and applied to future scenarios (SSP1-2.6 and SSP5-8.5), with daily maximum temperature and near-surface wind data used as inputs to a physics-based aircraft performance model. This model incorporates a semi-empirical TOFL formulation, thermodynamic thrust estimation, regulatory climb constraints, and wind effects. Results indicate a global mean TOFL increase of 4.5% under SSP1-2.6 and 9.5% under SSP5-8.5 by the end-of-century period (2075-2100), with regional maxima exceeding 20% at airports in tropical and subtropical climates. The frequency of inoperable days?defined as days when conditions prevent safe take-off at a given take-off weight?is projected to rise substantially, particularly at hot 1 Title Page and high-altitude airports, with some locations experiencing more than a doubling in such events under SSP5-8.5. This integrated, ensemble-based assessment underscores the operational risks posed by climate change and provides a foundation for adaptive strategies in aircraft performance standards and long-term airport infrastructure planning