Few-shot flight trajectory prediction in the multi-mode transfer learning structure

Abstract

Trajectory prediction (TP) is critical to the aircraft flight safety especially in few-shot situations, for example, flight of private or company aircraft, take-off or landing on aircraft carriers, etc., however most existing high-accuracy TP methods rely heavily on big data. For achieving few-shot TP performance, this paper proposes a few-shot TP structure consisting of two modules: the three-mode TP for the source domain and the three-mode transfer learning for the target domain. In the first module, instead of learning evolutionary characteristics of the whole flight process by a Bi-directional Long Short-Term Memory Network (Bi-LSTM), three Bi-LSTMs corresponding to climbing mode, declining mode, and whole-flight mode are designed, respectively, to extract different flight evolutionary rules. Particularly, the whole-flight mode can improve TP in transition processes between different flight modes. Additionally, outputs from the above three Bi-LSTMs are integrated via Convolutional Neural Networks (CNNs) driven by the source domain big data. In the second module, to avoid few-shot TP performance degradation caused by the negative transfer, transfer learning is carried out between matched flight modes, then the fine-tuning way is applied to improve these new models, and finally the linear combination instead of the CNN is designed to fuse outputs in few-shot situations. Compared with many existing methods, the proposed few-shot TP structure not only learn time-varying non-linear features but also extract multiple flight models separately based on transferred weights. Two kinds of real-world few-shot flight trajectories from Xiamen Airport to Shanghai Airport, and from Xi’an Airport to Beijing Airport are taken as examples to demonstrate the effectiveness of the proposed few-shot TP structure.