Auto-ATC：基於擴散變換器技術的自動空中交通管制系統

Abstract

中交通管制員（Air Traffic Controllers, ATCOs）在高壓環境中執勤，其人為錯誤可能導致嚴重後果。為減輕其工作負擔並提升運營安全性，我們提出了一套創新系統 AeroGuide，採用 Diffusion-Transformer 模型精準預測單架飛機的未來軌跡。該系統利用 OpenSky Network 提供的 ADS-B 數據進行訓練，生成準確的軌跡預測，並無縫整合至 GlideControl，後者是一種基於規則的演算法，可將預測結果轉化為具體的空中交通管制指令。為驗證系統效能，我們對 OpenScope（一款開源空中交通管制模擬器）進行改造，構建了基於東京國際機場現實場景的測試環境，對 AeroGuide 和 GlideControl 的結合進行綜合評估。模擬結果顯示，在超過 90,000 秒（略多於一天）的運行時間中，系統穩定運作，並取得卓越的性能表現：調整後的降落率平均值達 97.08%，中位數為 97.36%；降落率平均值為 95.07%，中位數為 95.27%。我們的主要貢獻包括：(1) 使用 OpenSky Network的 ADS-B 數 據 集 進 行 模 型 訓 練 與 評 估，(2) 開 發 基 於 Diffusion-Transformer 的AeroGuide 軌跡預測模型，(3) 設計將預測結果轉化為可操作指令的 GlideControl，(4) 改造 OpenScope 為可擴展的模擬器以進行系統集成與性能測試。本研究強調了數據驅動方法在支持空中交通管理與提升複雜現實場景運營成果方面的潛力。

Air traffic controllers (ATCOs) operate in high-stress environments where human error can have significant consequences. To alleviate their workload and enhance operational safety, we propose AeroGuide, a novel system that leverages Diffusion-Transformer models to predict aircraft trajectories with high precision, one aircraft at a time. Using the OpenSky Network’s ADS-B data as the training dataset, our model generates accurate future trajectory predictions and integrates seamlessly with GlideControl, a rule-based algorithm that translates these predictions into actionable air traffic control commands. To validate our system, we modified OpenScope, an open-source air traffic control simulator, creating a testing environment where the combination of AeroGuide and GlideControl was evaluated under realistic scenarios at Tokyo International Airport. Results collected over 90,000 seconds of simulated time (slightly over one day) demonstrate stable performance across experimental runs. The adjusted landed rate achieved a mean of 97.08% and a median of 97.36%, while the landed rate reached a mean of 95.07% and a median of 95.27%. Our key contributions include: (1) the utilization of the OpenSky Network ADS-B dataset for model training and evaluation, (2) the development of AeroGuide, a Diffusion-Transformer-based model for trajectory prediction, (3) the design of GlideControl for translating predictions into actionable commands, and (4) the adaptation of OpenScope as a scalable simulator for system integration and evaluation. This study underscores the potential for data-driven approaches to support air traffic management and enhance operational outcomes in complex real-world settings.