Search

Two main factors, including regression accuracy and adversarial attack robustness, of six trajectory prediction models are measured in this paper using the traffic flow management system (TFMS) public dataset of fixed-wing aircraft trajectories in a specific route provided by the Federal Aviation Administration. Six data-driven regressors with their desired architectures, from basic conventional to advanced deep learning, are explored in terms of the accuracy and reliability of their predicted trajectories. The main contribution of the paper is that the existence of adversarial samples was characterized for an aircraft trajectory problem, which is recast as a regression task in this paper. In other words, although data-driven algorithms are currently the best regressors, it is shown that they can be attacked by adversarial samples. Adversarial samples are similar to training samples, however, they can cause finely trained regressors to make incorrect predictions, which poses a security concern for learning-based trajectory prediction algorithms. It is shown that although deep-learning-based algorithms (e.g., long short-term memory (LSTM)) have higher regression accuracy with respect to conventional classifiers (e.g., support vector regression (SVR)), they are more sensitive to crafted states, which can be carefully manipulated even to redirect their predicted states towards incorrect states. This fact poses a real security issue for aircraft as adversarial attacks can result in intentional and purposely designed collisions of built-in systems that can include any type of learning-based trajectory predictor.

Best paper award in the trajectory prediction track; International audience; Ground-based aircraft trajectory prediction is a major concern in air traffic control and management. A safe and efficient prediction is a prerequisite to the implementation of new automated tools. In current operations, trajectory prediction is computed using a physical model. It models the forces acting on the aircraft to predict the successive points of the future trajectory. Using such a model requires knowledge of the aircraft state (mass) and aircraft intent (thrust law, speed intent). Most of this information is not available to ground-based systems. Focusing on the climb phase, we train neural networks to predict some of the unknown point-mass model parameters. These unknown parameters are the mass and the speed intent. For each unknown parameter, our model predicts a Gaussian distribution. This predicted distribution is a predictive distribution: it is the distribution of possible unknown parameter values conditional to the observed past trajectory of the considered aircraft. Using this distribution, one can extract a predicted value and the uncertainty related to this specific prediction. Using a physical model like BADA, this distribution could be used to derive a probability distribution of possible future trajectory ([1]). This study relies on ADS-B data coming from The OpenSky Network. It contains the climbing segments of the year 2017 detected by this sensor network. The 11 most frequent aircraft types are studied. The obtained data set contains millions of climbing segments from all over the world. Using this data, we show that despite having an RMSE slightly larger than previously tested methods, the predicted uncertainty allows us to reduce the size of prediction intervals while keeping the same coverage probability. Furthermore, we show that the trajectories with a similar predicted uncertainty have an observed RMSE close to the predicted one. The data set and the machine learning code are publicly available.

International audience; Ground-based aircraft trajectory prediction is a major concern in air traffic control and management. Focusing on the climb phase, we predict some of the unknown point-mass model parameters. These unknown parameters are the mass and the speed intent. This speed intent is parameterized by three values (cas 1 , cas 2 , $M$ ). These missing parameters might be useful to predict the future trajectory of a climbing aircraft. In this work, an ensemble of neural networks uses the observed past trajectory of the considered aircraft as input and predicts a Gaussian Mixture Model (GMM) modeling the joint distribution of (mass, cas 1 , cas 2 , $M$ ). Ideally, this predicted distribution will be close to a conditional distribution: the distribution of possible (mass, cas 1 , cas 2 , $M$ ) values given the observed past trajectory of the considered aircraft. This study relies on ADS-B data coming from The OpenSky Network. It contains the climbing segments of the year 2017 detected by this sensor network. The obtained data set contains millions of climbing segments from all over the world. Using this data, we show that using the proposed predictive model instead of a regression model brings almost as much information as using a regression model instead of a simple mean. The data set and the machine learning code are publicly available.

International audience; Ground-based aircraft trajectory prediction is a major concern in air traffic control and management. A safe and efficient prediction is a prerequisite to the implementation of new automated tools. In current operations, trajectory prediction is computed using a physical model. It models the forces acting on the aircraft to predict the successive points of the future trajectory. Using such a model requires knowledge of the aircraft state (mass) and aircraft intent (thrust law, speed intent). Most of this information is not available to ground-based systems. This paper focuses on the climb phase. We improve the trajectory prediction accuracy by predicting some of the unknown point-mass model parameters. These unknown parameters are the mass and the speed intent. This study relies on ADS-B data coming from The OpenSky Network. It contains the climbing segments of the year 2017 detected by this sensor network. The 11 most frequent aircraft types are studied. The obtained data set contains millions of climbing segments from all over the world. The climbing segments are not filtered according to their altitude. Predictive models returning the missing parameters are learned from this data set, using a Machine Learning method. The trained models are tested on the two last months of the year and compared with a baseline method (BADA used with the mean parameters computed on the first ten months). Compared with this baseline, the Machine Learning approach reduce the RMSE on the altitude by 48 % on average on a 10 minutes horizon prediction. The RMSE on the speed is reduced by 25 % on average. The trajectory prediction is also improved for small climbing segments. Using only information available before the considered aircraft takeoff , the Machine Learning method can predict the unknown parameters, reducing the RMSE on the altitude by 25 % on average. The data set and the Machine Learning code are publicly available.

International audience; Predicting aircraft trajectories is a key element in the detection and resolution of air traffic conflicts. In this paper, we focus on the ground-based prediction of final descents toward the destination airport. Several Machine Learning methods – ridge regression, neural networks, and gradient-boosting machine – are applied to the prediction of descents toward Toulouse airport (France), and compared with a baseline method relying on the Eurocontrol Base of Aircraft Data (BADA). Using a dataset of 15,802 Mode-S radar trajectories of 11 different aircraft types, we build models which predict the total descent length from the cruise altitude to a given final altitude. Our results show that the Machine Learning methods improve the root mean square error on the predicted descent length of at least 20 % for the ridge regression, and up to 24 % for the gradient-boosting machine, when compared with the baseline BADA method.

International audience; In this paper, we apply Machine Learning methods to improve the aircraft climb prediction in the context of ground-based applications. Mass is a key parameter for climb prediction. As it is considered a competitive parameter by many airlines, it is currently not available to ground-based trajectory predictors. Consequently, most predictors today use a reference mass that may be different from the actual aircraft mass. In previous papers, we have introduced a least square method to estimate the mass from past trajectory points, using the physical model of the aircraft. Another mass estimation method, based on an adaptive mechanism, has also been proposed by Schultz et. al. We now introduce a new approach, where the mass is considered as the response variable of a prediction model that is learned from a set of example trajectories. This Machine Learning approach is compared with the results obtained when using the BADA (Base of Aircraft Data) reference mass or the two state-of-the-art mass estimation methods. In these experiments, 9 different aircraft types are considered. When compared with the baseline method (resp. the mass estimation methods), the Machine Learning approach reduces the RMSE (Root Mean Square Error) on the predicted altitude by at least 58 % (resp. 27 %) when assuming the speed profile to be known, and by at least 29 % (resp. 17 %) when using the BADA speed profile except for the aircraft types E145 and F100. For these types, the observed speed profile is far from the BADA speed profile.

International audience; In this paper, we apply Machine Learning methods to improve the aircraft climb prediction in the context of groundbased applications. Mass and speed intent are key parameters for climb prediction. As they are considered as competitive parameters by many airlines, they are currently not available to groundbased trajectory predictors. Consequently, most predictors today use reference parameters that may be quite different from the actual ones. In our most recent paper ([1]), we have demonstrated that Machine Learning techniques provide a mass estimation significantly more precise than two state-of-the-art mass estimation methods. In this paper, we apply similar techniques to the speed intent. We first build a set of examples by adjusting CAS/Mach speed profile to each climb trajectory in our database. Then, using the adjusted values (ccas; cM) in this database, we learn a model able to predict the (cas;M) values of a new trajectory, using its past points as input. We apply this technique to actual Mode-C radar data and we consider 9 different aircraft types. When compared with the reference speed profiles provided by BADA, the reduction of the speed RMSE ranges from 36 % to 79 %, depending on the aircraft type. Using the predicted mass and speed profile, BADA is used to compute the predicted future trajectory with a 10 minute horizon. When compared with BADA used with the reference parameters, the reduction of the future altitude RMSE ranges from 45 % to 87 %.

International audience; Ground-based aircraft trajectory prediction is a major concern in air traffic control and management. A safe and efficient prediction is a prerequisite to the implementation of automated tools that detect and solve conflicts between trajectories. This paper focuses on the climb phase, because predictions are much less accurate in this phase than in the cruising phase. Trajectory prediction usually relies on a point-mass model of the forces acting on the aircraft to predict the successive points of the future trajectory. The longitudinal acceleration and climb rate are determined by an equation relating the modeled power of the forces to the kinetic and potential energy rate. Using such a model requires knowledge of the aircraft state (mass, current thrust setting, position, velocity, etc.), atmospheric conditions (wind, temperature) and aircraft intent (thrust law, speed intent). Most of this information is not available to ground-based systems. In this paper, we improve the trajectory prediction accuracy by learning some of the unknown point-mass model parameters from past observations. These unknown parameters, mass and thrust, are adjusted by fitting the modeled specific power to the observed energy rate. The thrust law is learned from historical data, and the mass is estimated on past trajectory points. The adjusted parameters are not meant to be exact, however they are designed so as to improve the energy rate prediction. The performances of the proposed method are compared with the results of standard model-based methods relying on the Eurocontrol Base of Aircraft DAta (BADA), using two months of radar track records and weather data.