• Definition and implementation of the end-to-end “radar data → ATC instruction” generation task • Rule-based parsing algorithm extracts 328,131 ATC instructions from radar trajectory data with 91.3% accuracy • TGCEATC model integrates Transformer temporal encoding and GNN spatial encoding with cross-attention fusion • Strict exact rate SER of 0.712 outperforms XGBoost by 38.5% • Establishes a paradigm combining rule-based data construction with deep learning for ATC automation research Predicting and generating air traffic control (ATC) instructions directly from real-time radar data remains an unsolved challenge, primarily due to the absence of paired “radar data–instruction” training datasets. Existing research has addressed only fragmented sub-tasks such as trajectory prediction or conflict detection, without achieving comprehensive end-to-end instruction generation. This study proposes a two-stage solution: first, a rule-based parsing algorithm reconstructs 328,131 ATC instructions from 30 days of radar trajectory data by detecting trajectory parameter variations, forming large-scale paired training samples; second, a Transformer-GNN Cross-attention Enhanced ATC (TGCEATC) model is developed to learn the mapping from air traffic situations to complete instructions via a multi-task architecture integrating Transformer-based temporal encoding and GNN-based spatial encoding. Experimental results demonstrate that all generated instructions possess practical ATC significance, with instruction type and control action F1 scores of 0.857 and 0.774, respectively, a strict exact rate (SER) of 0.712, and a mean absolute time deviation (MATD) of 14.7 s—representing a 38.5% SER improvement over XGBoost. This work defines and implements the end-to-end ATC instruction generation paradigm for the first time.
Zhao et al. (Sun,) studied this question.