Archive ouverte de l'ENAC
Not a member yet
3458 research outputs found
Sort by
Global Sensitivity Analysis: a novel generation of mighty estimators based on rank statistics
No full textErratum for Global Sensitivity Analysis: a novel generation of mighty estimators based on rank statistics. Fabrice Gamboa, Thierry Klein, Agnès Lagnoux, and Paul Rochet.International audienceWe propose a new statistical estimation framework for a large family of globalsensitivity analysis indices. Our approach is based on rank statistics and uses anempirical correlation coefficient recently introduced by Chatterjee [9]. We show howto apply this approach to compute not only the Cramér-von-Mises indices, whichare directly related to Chatterjee’s notion of correlation, but also first-order Sobolindices, general metric space indices and higher-order moment indices. We establishconsistency of the resulting estimators and demonstrate their numerical efficiency,especially for small sample sizes. In addition, we prove a central limit theorem forthe estimators of the first-order Sobol indices
Canonical foliations of neural networks: application to robustness
No full textDeep learning models are known to be vulnerable to adversarial attacks. Adversarial learning is therefore becoming a crucial task. We propose a new vision on neural network robustness using Riemannian geometry and foliation theory. The idea is illustrated by creating a new adversarial attack that takes into account the curvature of the data space. This new adversarial attack called the two-step spectral attack is a piece-wise linear approximation of a geodesic in the data space. The data space is treated as a (degenerate) Riemannian manifold equipped with the pullback of the Fisher Information Metric (FIM) of the neural network. In most cases, this metric is only semi-definite and its kernel becomes a central object to study. A canonical foliation is derived from this kernel. The curvature of transverse leaves gives the appropriate correction to get a two-step approximation of the geodesic and hence a new efficient adversarial attack. The method is first illustrated on a 2D toy example in order to visualize the neural network foliation and the corresponding attacks. Next, experiments on the MNIST dataset with the proposed technique and a state of the art attack presented in Zhao et al. (2019) are reported. The result show that the proposed attack is more efficient at all levels of available budget for the attack (norm of the attack), confirming that the curvature of the transverse neural network FIM foliation plays an important role in the robustness of neural networks
HARVIS: a digital assistant based on cognitive computing for nonstabilized approaches in Single Pilot Operations
No full textInternational audienceWith Single Pilot Operations, the work that was shared between crew members in the past will be assumed by one pilot. To maintain an adequate level of safety and support the single pilots' performance and decision-making in complex situations, we investigated a concept of cognitive computing algorithms and adaptive automation implemented in a digital assistant. The aim of the present work was to assess if an Artificial Intelligence (AI) assistant could support pilots' decision-making during a Non-Stabilized Approach (NSA). Method: A NSA assistant prototype based on pilots expertise and eye-tracking data was developed. We performed a human in the loop session with 7 professional pilots on ENAC A320 simulator and analyzed the impact of the assistant on operations, human performance and the safety feeling of pilots. Results: Generally, the assistant was appreciated and trusted by pilots. The system improved pilots' decision-making in order to perform a go-around. Aspects like trust, complacency, usability and pilots' physiological data monitoring were dominant discussion points. Discussion and conclusion: There were no evidence that the assistant reduced pilots' workload, but it helped pilots focusing a bit more on the 'wider picture' since part of the monitoring task could be delegated to the assistant. Understanding the reason of a go-around alert appeared to be a key point for pilots to build trust
On-site experiments and electromagnetic modelling to characterize Corona discharges in VHF-band
No full textInternational audienc
Co-Development Approach integrating Training into the Design Process
No full textInternational audienceSince the introduction of advanced automation technology within highly dynamic and complex systems, traditional approaches to training no longer appear to be adequate in preparing end-users of such technologies (Carroll and Olson, 1988). Advances in design processes have resulted in the creation of user-centered designs that employ an iterative design process that follows a cyclical pattern, where the prototype is designed, created and evaluated, throughout multiple loops. Once the final design solution is validated, a training session is organized to prepare the end-user for the Human in the loop Simulation. Since the training sessions and the preparation of these training sessions are conceived after the completion and validation of the design; any issues with the design often result in extended training sessions to compensate for human factors that were overlooked in the initial design. Nonetheless, such enhanced training “cannot and should not be a fix for bad design” (Sarter, Woods & Billings, 1997.p.1936). This research intended to consider the training from the onset of the design in order to avoid transferring the burden of a poor design onto the training. This was not fully possible in the framework of the SAFEMODE project, where this research was conducted, as some design iterations had already taken place prior to the implementation of the proposed training considerations. Instead, a training approach involving future users as part of the system design, hereafter referred to as “Co-Development”, was proposed and tested during the creation of a safety net or alert for air-traffic controllers. Building on this added value to user-centered design, we investigated: To what extent does the co-development of the training with end-users improve the training, the design, and the acceptability of the designed alert?This paper presents this Co-Development approach and the outcome of its first implementation; particularly in terms of its impact on the training, the design, and the acceptability of the alert. Specifically, this system design approach allows for human factor issues to be identified and hence corrected early in the design stage, thus yielding a potential impact on the acceptability and usability of the alert, while simultaneously affecting safety through an improved user-centered design experience. This paper also proposes a future research direction that consists of a joint training and design approach that considers training right from the onset of the first iteration of the design. As a result, the training is conceived jointly and evolves simultaneously with each design loop, further improving the design, and thus reducing the burden placed onto the training due to design shortfalls and impossible design challenges that are seemingly improbable to meet
Experimental Evaluation of Panel-Method-Based Path Planning for eVTOL in A Scaled Urban Environment
No full textInternational audienceIn this study, previously proposed panel method based path planning for electric vertical take off and landing vehicles in urban environments is tested in a high fidelity simulation environment and with real-life drones in an indoor flight arena. Panel method is a numerical tool, borrowed form fluid dynamics domain, that can generate collision free paths for multiple vehicles in environments with arbitrarily shaped obstacles while guaranteeing obstacle avoidance and convergence to global minima with little computational load. In this study, panel method based path planning is further improved with introduction of novel safety source element that enables a safety perimeter around obstacles without losing convergence guarantee. Furthermore, path planning capability of panel method for electric vertical take off and landing vehicles in urban environments is demonstrated with hardware experiments in a scaled urban environment. Experiment results indicate that panel method is a promising tool for path planning in urban environments
Probabilistic proofs of large deviation results for sums of semiexponential random variables and explicit rate function at the transition
No full textInternational audienceAsymptotics deviation probabilities of the sum S n = X 1 + · · · + X n of independent and identically distributed real-valued random variables have been extensively investigated, in particular when X 1 is not exponentially integrable. For instance, A.V. Nagaev formulated exact asymptotics results for P(S n > x n) when x n > n 1/2 (see, [13, 14]). In this paper, we derive rough asymptotics results (at logarithmic scale) with shorter proofs relying on classical tools of large deviation theory and expliciting the rate function at the transition
Study and development of an AI assistant for future Moon and Mars stations
No full textInternational audienceFollowing the Global Exploration Roadmap (GER) defined by the International Space Exploration Coordination Group(ISECG), the Spaceship FR team from CNES, the French Space Agency, wishes to contribute to the development oftechnologies extending human reach toward space, notably for the development of Moon and Mars bases. Theoperation and sustainability of such structures in stressful isolation conditions constitute a high level technological and human challenge. To relieve the high mental load of the astronauts, the solution could be an artificial intelligenceassistant that would supervise the automation of the base as well as monitor and maintain the mental health of the crewthrough a cognitive approach of the human-computer interaction. AI4U is the system at the crossroad betweencomputer sciences and human factors, aspiring to take on the task. Besides, its organisation and automation skills, thisartificial intelligence interacts with the astronauts with an intuitive and natural interface. It can support their work and leisure activities with an empathic approach, recognising their current mental state by using face and voice recognition. This paper will present the different steps of the development of AI4U, the challenges encountered and the next steps until its usage in a Moon outpost.<br /
New advances in Multidisciplinary Design Optimization with Gaussian Process for Eco-design Aircraft
No full textInternational audienceModern aircraft design increasingly requires the simultaneous optimization of performance, cost, and environmental impact. In this study, we present new advances in Multidisciplinary Design Optimization (MDO) for eco-design aircraft by leveraging Gaussian process (GP) surrogate models. Our approach develops a novel GP-based framework that efficiently approximates high-fidelity multidisciplinary simulations, integrating a customized kernel formulation to capture the complex interdependencies among aerodynamic, structural, and eco-design variables. The proposed methodology not only accelerates the optimization process but also enhances the reliability of sustainability assessments. Results on representative design cases illustrate significant improvements in both computational efficiency and environmental performance, paving the way for more sustainable aircraft design practices
