Portail HAL Ensta
Not a member yet
11080 research outputs found
Sort by
Modèles homogénéisés enrichis en présence de bords : Analyse et traitement numérique
No full textWhen we are interested in the propagation of waves in a periodic medium at low frequency (i.e. wavelength large compared to the period length), it is possible to model the periodic medium by an equivalent or effective homogeneous medium which has the same macroscopic properties. It is the homogenization theory that mathematically justifies this process. This process is very attractive because numerical calculations are then much less expensive (the small periodic structure has disappeared) and analytical calculations are again possible in certain configurations. The waves in the periodic medium and in the effective medium are very close from a macroscopic point of view except in the presence of boundaries or interfaces.Indeed, it is well known that the homogenized model is obtained by neglecting the boundary effects and consequently, it is much less precise at the boundaries of the periodic medium. When interesting phenomena appear at the edges of the middle (such as the propagation of plasmonic waves on the surface of metamaterials for example), it therefore seems difficult to trust the effective model.Returning to the homogenization process, we propose a homogenized model which is richer at the boundaries. The enriched homogenized model is as simple as the classical homogenized model far from the interfaces, only the boundary conditions change and take better account of the phenomena. We apply this model to an elliptical equation in the case of the simple geometry of the half-plane with Dirichlet or Neumann type conditions. From a numerical point of view, in addition to classic cell problems that appear in homogenization, periodic band problems must also be solved. Finally, we apply these results to the long time wave equation.Quand on s’intéresse à la propagation des ondes dans un milieu périodique à basse fréquence (i.e. la longueur d’onde est grande devant la période), il est possible de modéliser le milieu périodique par un milieu homogène équivalent ou effectif qui a les mêmes propriétés macroscopiques. C’est la théorie de l’homogénéisation qui justifie d’un point de vue mathématique ce procédé. Ce procédé est très séduisant car les calculs numériques sont beaucoup moins couteux (la petite structure périodique a disparu) et des calculs analytiques sont de nouveau possibles dans certaines configurations. Les ondes dans le milieu périodique et dans le milieu effectif sont très proches d’un point de vue macroscopique sauf en présence de bords ou d’interfaces.En effet, il est bien connu que le modèle homogénéisé est obtenu en négligeant les effets de bords et par conséquent il est beaucoup moins précis aux bords du milieu périodique. Quand les phénomènes intéressants apparaissent aux bords du milieu (comme la propagation des ondes plasmoniques à la surface des métamatériaux par exemple), il semble donc difficile de faire confiance au modèle effectif.En revenant sur le processus d’homogénéisation, nous proposons un modèle homogénéisé qui est plus riche aux niveaux des bords. Le modèle homogénéisé enrichi est aussi simple que le modèle homogénéisé classique loin des interfaces, seule les conditions aux bords changent et prennent mieux en compte les phénomènes. Nous appliquons ce modèle à une équation elliptique dans le cas de la géométrie simple du demi-plan avec des conditions de type Dirichlet ou Neumann. D’un point de vue numérique, en plus des problèmes de cellule classiques qui apparaissent en homogénéisation, des problèmes de bandes périodiques doivent également être résolus. Pour finir, nous appliquons ces résultats à l'homogénéisation de l'équation des ondes en temps long et en présence de bords
A boundary approach for set inversion
No full textInternational audienceIn the paper, we present a new interval-based set inversion algorithm which takes into account the continuity of the problem. In the case where the set Y to be inverted has some volume, we show that inverting the boundary ∂Y of Y is sufficient to reconstruct the preimage X=f−1(Y). The inversion of ∂Y separates the domain of f into two regions: one inside X and one outside. To detect which part is inside or outside, we show that we can retro-propagate the information coming from Y at negligible cost. The efficiency of the approach is illustrated on a localization problem
Prospects on Solving an Optimal Control Problem with Bounded Uncertainties on Parameters using Interval Arithmetics
No full textInternational audienceAn interval method based on Pontryagin's Minimum Principle is proposed to enclose the solutions of an optimal control problem with embedded bounded uncertainties. This method is used to compute an enclosure of all optimal trajectories of the problem, as well as open loop and closed loop enclosures meant to validate an optimal guidance algorithm on a concrete system with inaccurate knowledge of the parameters. The differences in geometry of these enclosures are exposed, and showcased on a simple system. These enclosures can guarantee that a given optimal control problem will yield a satisfactory trajectory for any realization of the uncertainties. Contrarily, the probability of failure may not be eliminated and the problem might need to be adjusted
DEEP SETS FOR GENERALIZATION IN RL
No full textThis paper investigates the idea of encoding object-centered representations in the design of the reward function and policy architectures of a language-guided reinforcement learning agent. This is done using a combination of object-wise permutation invariant networks inspired from Deep Sets and gated-attention mechanisms. In a 2D procedurally-generated world where agents targeting goals in natural language navigate and interact with objects, we show that these architectures demonstrate strong generalization capacities to out-of-distribution goals. We study the generalization to varying numbers of objects at test time and further extend the object-centered architectures to goals involving relational reasoning.International audienceThis paper investigates the idea of encoding object-centered representations in the design of the reward function and policy architectures of a language-guided reinforcement learning agent. This is done using a combination of object-wise permutation invariant networks inspired from Deep Sets and gated-attention mechanisms. In a 2D procedurally-generated world where agents targeting goals in natural language navigate and interact with objects, we show that these architectures demonstrate strong generalization capacities to out-of-distribution goals. We study the generalization to varying numbers of objects at test time and further extend the object-centered architectures to goals involving relational reasoning
RESOLUTION OF FLUID-STRUCTURE COUPLED PROBLEMS WITH FLOW USING THE BOUNDARY ELEMENT METHOD
No full textEffects of fluid-structure coupling on the dynamic behaviour of flexible airship can be modelled with a potential, incompressible, inviscid flow. A new formalism to study linear variations of the flow induced on the fluid-structure interface in a time dependent ambient flow is introduced. The features of the Boundary Element Method used to solve this problem numerically are exhibited. Numerical results of the linear model are compared with analytical and non linear numerical results, assessing the validity and the limitations of the approach
Acoustics and photo-identification provide new insights on killer whale presence and movements when interacting with longline fisheries in South East Australia
No full textInternational audienceRemoval of target catch from longlines by marine mammals, known as depredation, is a global issue creating animal welfare, socioeconomic and management concerns. The killer whale (Orcinus orca) is a key species of concern for longline depredation due to their global presence and ability to remove large quantities of caught fish. Currently, data on whale behaviours around fishing vessels, including timing of depredation and ability to follow vessels, is not completely understood. This lack of knowledge prevents a complete assessment of depredation both spatially and temporally, which is important to reduce the chance of underestimating depredation events. Our study utilised both photo-identification and acoustics as complementary approaches to investigate depredation on the blue-eye trevalla (Hyperoglyphe antarctica) demersal longline fishery in Australia. Of the 14 d when depredations were recorded acoustic detection of killer whales prior to visual confirmations occurred in 13 (93 %). Photo-identification revealed individuals repeatedly interacting with the vessel, sometimes over long distances (>1000 km), with increasing inter-depredation times as the vessel travelled further. These findings suggest that killer whales move to known fishing areas well before detection from surface monitoring occurs. While this study has revealed aspects of killer whale behaviours when encountering and interacting with longline fisheries, it has highlighted how traditional monitoring methods underestimate depredation and that this aspect needs to be further investigated
Cloud-connected flying edge computing for smart agriculture
No full textInternational audienceDue to recent advancements and the success of versatile mobile applications, more and more services around the globe are being moved to the cloud. As a result, its limitations have become more evident. The major issues that cloud-based applications face include large latency, bottlenecks because of central processing, compromised security, and lack of offline processing. The drawbacks of cloud computing are reduced by fog and edge computing, where data are processed near the places where it is generated—at network edges or fog nodes—most importantly in a distributed way. Smart agriculture is an approach based on the Internet of Things (IoT) where cloud computing is not an option as the internet is usually not available at remote sites. In addition, pure edge computing also is not practical, as most sensor nodes are very small and they do not have enough computing power. Intermediate fog computing also is not a good choice, as fixed fog nodes (getaway nodes) do not work well with high node fluctuation caused by bad weather or harsh conditions. Considering these issues and limitations, we have proposed the idea of flying edge computing where an unmanned aerial vehicle (UAV) acts as an edge-computing machine. This can be an ideal solution for smart agriculture, given the size and remoteness of many agricultural areas. This technique can be called “wind or breeze computing” because the data are “blown” or moved by the current of computing. The Flying-Edge offers fast deployment of edge facilities in challenging locations and it can be a major step to accomplish the goal of IoT-based smart agriculture
Integrated water vapour content retrievals from ship-borne GNSS receivers during EUREC<sup>4</sup>A
No full textInternational audienceIn the framework of the EUREC4A (Elucidating the role of clouds–circulation coupling in climate) campaign that took place in January and February 2020, integrated water vapour (IWV) contents were retrieved over the open tropical Atlantic Ocean using Global Navigation Satellite System (GNSS) data acquired from three research vessels (R/Vs): R/V Atalante, R/V Maria S. Merian and R/V Meteor. This paper describes the GNSS processing method and compares the GNSS IWV retrievals with IWV estimates from the European Centre for Medium-range Weather Forecasts (ECMWF) fifth reanalysis (ERA5), from the Moderate Resolution Imaging Spectroradiometer (MODIS) infrared products and from terrestrial GNSS stations located along the tracks of the ships. The ship-borne GNSS IWV retrievals from R/V Atalante and R/V Meteor compare well with ERA5, with small biases (−1.62 kg m−2 for R/V Atalante and +0.65 kg m−2 for R/V Meteor) and a root mean square (rms) difference of about 2.3 kg m−2. The results for the R/V Maria S. Merian are found to be of poorer quality, with an rms difference of 6 kg m−2, which is very likely due to the location of the GNSS antenna on this R/V prone to multipath effects. The comparisons with ground-based GNSS data confirm these results. The comparisons of all three R/V IWV retrievals with MODIS infrared products show large rms differences of 5–7 kg m−2, reflecting the enhanced uncertainties in these satellite products in the tropics. These ship-borne IWV retrievals are intended to be used for the description and understanding of meteorological phenomena that occurred during the campaign, east of Barbados, Guyana and northern Brazil. Both the raw GNSS measurements and the IWV estimates are available through the AERIS data centre (https://en.aeris-data.fr/, last access: 20 September 2020). The digital object identifiers (DOIs) for R/V Atalante IWV and raw datasets are https://doi.org/10.25326/71 (Bosser et al., 2020a) and https://doi.org/10.25326/74 (Bosser et al., 2020d), respectively. The DOIs for the R/V Maria S. Merian IWV and raw datasets are https://doi.org/10.25326/72 (Bosser et al., 2020b) and https://doi.org/10.25326/75 (Bosser et al., 2020e), respectively. The DOIs for the R/V Meteor IWV and raw datasets are https://doi.org/10.25326/73 (Bosser et al., 2020c) and https://doi.org/10.25326/76 (Bosser et al., 2020f), respectively
SDN Intent-based conformance checking: application to security policies
No full textInternational audienceWith the popularity of software defined networking architectures, the growing complexity of its use cases dictates the need for better auditability especially for security. In this paper, we aim at facilitating high-level management-plane policy configuration conformance auditing and their reflection in the data plane, to detect missing or spurious flow rules with respect to security policies. To this end, we propose an efficient conformance checking approach based on an intentional northbound interface as well as traces of management, control and data plane. Leveraging a proof-of-concept implementation of our approach, we compare its conformance-checking runtime and precision against a direct method on virtual topologies and find that it significantly improves scalability. We conclude by proposing directions for further enhancements extending the techniques presented herein