Timothé Krauth

Modèle génératif profond pour l'estimation de probabilité de collision en vol: Il est essentiel de calculer la probabilité de collisions aériennes pour optimiser le trafic aérien tout en maintenant de hauts standards de sécurité. Cette nécessité s'est accentuée dans les années 1960 avec l'augmentation du trafic aérien commercial transatlantique. Initialement, les modèles analytiques tels que ceux de Reich et Anderson-Hsu étaient les références pour évaluer les risques de collision en vol, mais s’avèrent peu adaptés aux espaces aériens complexes autour des aéroports.Les méthodes basées sur les données, et en particulier les simulations de Monte-Carlo, sont devenues une alternative prometteuse pour l'évaluation du risque de collision. Elles offrent une grande flexibilité grâce à des hypothèses simplifiées, les rendant adaptables à divers contextes. Toutefois, les simulations de Monte-Carlo classiques se révèlent peu efficaces pour estimer les probabilités d'événements rares, nécessitant un grand nombre de trajectoires d'avions et des ressources computationnelles importantes. Cette thèse propose un modèle de risque de collision basé sur les simulations de Monte-Carlo, utilisant un modèle de génération de trajectoires pour pallier ces limitations dues aux événements rares. Ces méthodes génératives reproduisent fidèlement les distributions de trajectoires observées tout en intégrant les incertitudes dues à des facteurs externes. Trois axes de recherche principaux sont définis : (i) le développement d'une méthode de génération de trajectoires, (ii) la construction d'un modèle de risque de collision basé sur les méthodes de Monte-Carlo utilisant des trajectoires synthétiques, et (iii) l'amélioration de l'interprétabilité des estimations du risque de collision.La génération d'échantillons synthétiques nécessite l'estimation de la distribution des données observées pour garantir une distribution identique des nouveaux échantillons. C'est particulièrement important pour les trajectoires aériennes, où le modèle doit refléter les incertitudes causant des écarts par rapport aux trajectoires nominales. Nous avons d’abord utilisé des méthodes d'apprentissage statistique traditionnelles pour estimer des trajectoires aériennes complexes en deux dimensions. Malgré la réduction de la dimensionnalité du problème, les méthodes conventionnelles peinent à estimer les distributions en grande dimension. Nous avons alors exploré l’utilisation des autoencodeurs variationnels pour une estimation plus fine de la densité de probabilité. Convenablement adaptés aux applications de séries temporelles multivariées, les autoencodeurs variationnels se révèlent efficaces pour estimer la distribution de trajectoires aériennes complexes.En utilisant la méthode de génération développée, nous estimons alors le risque de perte de séparation induit par les procédures de décollage et d'atterrissage de l'aéroport de Paris-Orly à l'aide de simulations de Monte-Carlo. L'emploi d'une méthode de génération de trajectoires se révèle prometteur, permettant de créer l'équivalent de 20 ans de trajectoires aériennes à partir de seulement deux mois d'observations. Toutefois, les contraintes inhérentes aux méthodes classiques de Monte-Carlo ne sont pas réellement surmontées mais simplement différées par la production d'un ensemble de trajectoires de taille arbitrairement grande.Le travail final de cette thèse unifie les cadres de l'autoencodeur variationnel et de la quantification de l'incertitude. Il démontre comment les autoencodeurs variationnels peuvent construire des distributions d'entrée adaptées pour les algorithmes de quantification de l'incertitude, améliorant la fiabilité des simulations de Monte-Carlo grâce au subset simulation et l'explicabilité de l'estimation de la probabilité de collision en vol par l'analyse de sensibilité. Plus généralement, nous avons montré que l’autoencodeur variationel représente un outil prometteur à associer aux problèmes de quantification d’incertitudes.

The reaction time to a deconfliction situation refers to the interval between detecting a potential loss of separation and taking corrective action. For air traffic controllers, it represents the lead time between identifying a conflict and issuing a deconfliction instruction to a pilot. For pilots, it corresponds to the delay between receiving an ATC clearance and initiating the associated maneuver. Because both processes are influenced by human factors, these response times constitute a significant source of uncertainty in Air Traffic Management. While the controller’s reaction time is particularly difficult to estimate since the exact moment at which a conflict is cognitively detected cannot be directly inferred from operational data, this paper focuses on quantifying pilot response times. To this end, we propose a methodology that combines Natural Language Processing techniques with surveillance and flight plan data. ATC–pilot voice communications are transcribed using a fine-tuned Automatic Speech Recognition model, and aircraft callsigns are identified through Named Entity Recognition. The transcriptions are then matched with corresponding flights in ADS-B surveillance data. Using flight plans, we identify lateral deconfliction maneuvers and align them temporally with the preceding ATC clearances to estimate the elapsed time between instruction and execution. Because the approach depends on a sequence of emerging algorithms whose robustness is still evolving, the study focuses on identifying the conditions under which the methodology performs reliably, highlighting its current limitations and associated data challenges, and proposing ways to overcome them.

Reliable short-term aircraft trajectory prediction is essential for safety and efficiency in Air Traffic Management (ATM). This work introduces a generative framework for probabilistic 4D trajectory forecasting based on Conditional Flow Matching (CFM), a recent deep generative modeling approach that combines stable likelihood-based training with efficient sampling. The model is trained on historical ADS–B data from the OpenSky Network to predict aircraft motion over a 60 s horizon, conditioned on the preceding 60 s of observations. The model generates ensembles of realistic future trajectories that capture the inherent uncertainty of aircraft motion and enable probabilistic assessment of potential conflicts. As an application, we estimate the probability of mid-air collision during a loss-of-separation event using Monte Carlo simulation over the generated trajectories, providing a quantitative risk measure. The results demonstrate that flow-based generative modeling offers a principled foundation for uncertainty-aware trajectory prediction and safety analysis in ATM.

Accurate pre-flight fuel planning is essential to ensure that an aircraft carries enough fuel for a safe flight, while avoiding unnecessary weight that reduces efficiency. However, at certain airports, frequent shortcuts on arrival can lead to systematic discrepancies between the nominal STAR (Standard Terminal Arrival Route) distances used for fuel planning and the actual distances flown. This, in turn, can result in aircraft carrying unnecessary excess fuel. To address this issue, some airports have begun to publish expected STAR distances for specific procedures in the Aeronautical Information Publication, allowing operators to plan fuel more accurately. This paper examines the impact of providing expected STAR distances by analysing one year of ADS-B data from Geneva, Munich and Rome Fiumicino airports. The study compares the observed distances flown with the full and expected STAR distances and presents econometric models to identify factors influencing the actual flown distances. In addition, fuel calculations are presented to estimate the potential benefits of publishing expected distances at airports that do not currently provide this information. The results show that for most of the STARs analysed, significant differences between observed flight distances and full STAR distances exist. However, these discrepancies are mitigated by the availability of published expected distances at Munich and Rome. The econometric model highlights consistent influencing factors such as STAR shape, shortcut potential, peak traffic hours, and weather conditions, although some effects are more location-specific. The fuel savings analysis suggests that adopting this practice at more airports could significantly reduce unnecessary fuel burn and associated emissions. Overall, this paper increases the understanding of how publishing expected STAR distances can improve fuel planning accuracy, operational efficiency, and environmental sustainability.

Probability density function estimation with weighted samples is the main foundation of all adaptive importance sampling algorithms. Classically, a target distribution is approximated either by a non-parametric model or within a parametric family. However, these models suffer from the curse of dimensionality or from their lack of flexibility. In this contribution, we suggest to use as the approximating model a distribution parameterised by a variational autoencoder. We extend the existing framework to the case of weighted samples by introducing a new objective function. The flexibility of the obtained family of distributions makes it as expressive as a non-parametric model, and despite the very high number of parameters to estimate, this family is much more efficient in high dimension than the classical Gaussian or Gaussian mixture families. Moreover, in order to add flexibility to the model and to be able to learn multimodal distributions, we consider a learnable prior distribution for the variational autoencoder latent variables. We also introduce a new pre-training procedure for the variational autoencoder to find good starting weights of the neural networks to prevent as much as possible the posterior collapse phenomenon to happen. At last, we explicit how the resulting distribution can be combined with importance sampling, and we exploit the proposed procedure in existing adaptive importance sampling algorithms to draw points from a target distribution and to estimate a rare event probability in high dimension on two multimodal problems.

The analysis and prediction of go-arounds, also referred to as missed approaches, is an active field of research due to the go-around’s impact on safety and the disruption of the traffic flow at airports. The advent of open-source aircraft trajectories available to researchers has increased the level of interest in the field. This paper introduces a publicly available dataset containing metadata of almost 9 million landings and 33,000 go-arounds. The dataset is based on observations from the OpenSky Network and includes data from 176 airports in 44 countries observed in the year 2019. After downloading the data, a go-around classification was performed and the quality was assessed. The usefulness of the dataset is illustrated with two novel example applications. The first example shows how the go-around rate for a runway can be modeled by using a quasi-binomial generalized linear model, while the second example compares the go-around rates for a number of airport–airline pairs. The introduced dataset is significantly larger than the data used so far in the analysis of go-arounds and provides the opportunity to develop novel use cases. This dataset frees researchers from having to collect and process large amounts of data and instead lets them focus on the analysis. The authors are convinced that this large dataset will stoke the creativity of the research community and facilitate interesting and novel applications.

Aircraft trajectory generation is a high stakes problem with a wide scope of applications, including collision risk estimation, capacity management and airspace design. Most generation methods focus on optimizing a criterion under constraints to find an optimal path, or on predicting aircraft trajectories. Nevertheless, little in the way of contribution has been made in the field of the artificial generation of random sets of trajectories. This work proposes a new approach to model two-dimensional flows in order to build realistic artificial flight paths. The method has the advantage of being highly intuitive and explainable. Experiments were conducted on go-arounds at Zurich Airport, and the quality of the generated trajectories was evaluated with respect their shape and statistical distribution. The last part of the study explores strategies to extend the work to non-regularly shaped trajectories.