Study on the Standardization Method of Radiotelephony Communication in Low-Altitude Airspace Based on BART

Boyuan Han ^1* Weijun Pan ¹ Peiyuan Jiang ²

• ¹ Civil Aviation Flight University of China, Guanghan, China
• ² University of Electronic Science and Technology of China, Chengdu, Sichuan Province, China

The research and application of air traffic control automation systems in the field are constrained by the scarcity and low quality of radiotelephony communication data, especially in complex low-altitude airspace where communication data is less standardized. This study propose a deep reinforcement learning model based on BART, named BART-Reinforcement Learning (BRL), for standardizing non-standard control instructions encountered during training or operational scenarios. Leveraging the BART pre-trained language model, the BRL model is optimized through transfer learning and reinforcement learning techniques. This study evaluated the model's performance across various control datasets, including training flight data, civil aviation control operation data, and a dataset of generated control instructions derived from the "Radiotelephony Communications for Air Traffic Services" standard. Under ROUGE evaluation criteria, the BRL model exhibited significant improvements across all datasets, validating its efficacy. Additionally, we conducted a comprehensive evaluation of the model's performance based on the intent of land-air communication instructions, revealing a marked enhancement in standardizing instructions across diverse types. Compared to the baseline model, the overall accuracy increased by 10.5%, effectively alleviating the generalization challenges of deep learning models across disparate datasets.