Abstract

International audience; Trajectory prediction with Closest Point of Approach (CPA) concept is a fundamental element of aircraft Conflict Detection (CD) problem. Conventional motion-based CPA prediction model generally assumes that aircraft is flying in straight line with constant speed. But due to environment uncertainties and ground speed changes, this conventional method frequently lacks accuracy in the real world with a high rate of false alarms and missed detections. In this paper, we introduce a novel automated data-driven CD framework with Machine Learning (ML) for 3D CPA prediction in a lookahead time of less than 20 minutes. Firstly, a 3D CPA model with cylindrical norm is proposed as the baseline. Then, data preparation with Mode-S observation data in France is explained, including data collection and data processing, to convert raw Mode-S data to the close-to-reality dataset. Furthermore, feature engineering is applied to build up a feature set with 16 features. Finally, four prevailing ML models are used to predict the time, horizontal distance and vertical distance of CPA in 3D airspace. CD is conducted based on the predicted values. The prediction and CD results show that all proposed ML models outperform the baseline model. Especially, GBM and FFNNs could strongly enhance the performance of CD.

Abstract

Closest Point of Approach (CPA) is one of the main problems in aircraft Conflict Detection (CD). It aims to find out the minimum distance and the associated time between two aircraft on the same altitude with crossing traffic. Conventional CPA prediction model generally assumes that the speed and heading of the aircraft are constant. But the uncertainties in real operations lead to the inaccuracy of CPA prediction. In this paper, we introduce a novel CD framework with Machine Learning (ML) methods. It aims to improve the CPA prediction accuracy with the help of real trajectory data. The new model contributes to not only reduce the number of fault short-mid term conflict alert for air traffic controllers but also support the implementation of future free flight concept, so as to reduce fuel consumption and emission. In our study, we firstly propose a data processing method to generate a close-to-reality simulation data from Mode-S observations. Then, feature engineering is used to transform the raw data into suitable features, which will enable the ML models to make predictions with high- performance. Six prevailing ML methods (MLR, SVM, FFNNs, KNN, GBM, RF) are used to predict the CPA time and distance. Their prediction results are compared with the conventional CPA model (baseline). The simulation results demonstrate that the GBM is the best prediction model both in CPA prediction and conflict detection. However, the results also prove that not all ML models outperform the baseline CPA model. Suitable ML methods can greatly enhance the accuracy of conflict detection. Refereed/Peer-reviewed

Abstract

Closest Point of Approach (CPA) is one of the main problems in aircraft Conflict Detection (CD). It aims to find out the minimum distance and the associated time between two aircraft on the same altitude with crossing traffic. Conventional CPA prediction model generally assumes that the speed and heading of the aircraft are constant. But the uncertainties in real operations lead to the inaccuracy of CPA prediction. In this paper, we introduce a novel CD framework with Machine Learning (ML) methods. It aims to improve the CPA prediction accuracy with the help of real trajectory data. The new model contributes to not only reduce the number of fault short-mid term conflict alert for air traffic controllers but also support the implementation of future free flight concept, so as to reduce fuel consumption and emission. In our study, we firstly propose a data processing method to generate a close-to-reality simulation data from Mode-S observations. Then, feature engineering is used to transform the raw data into suitable features, which will enable the ML models to make predictions with high-performance. Six prevailing ML methods (MLR, SVM, FFNNs, KNN, GBM, RF) are used to predict the CPA time and distance. Their prediction results are compared with the conventional CPA model (baseline). The simulation results demonstrate that the GBM is the best prediction model both in CPA prediction and conflict detection. However, the results also prove that not all ML models outperform the baseline CPA model. Suitable ML methods can greatly enhance the accuracy of conflict detection Refereed/Peer-reviewed