Scholarlink

<p>Flight trajectory prediction is fundamental to modern air traffic management and air combat maneuver decision-making. However, traditional deep learning models under general training paradigms often struggle to capture long-term dependencies when processing extended sequences, leading to imprecise data representations and significant error accumulation in long-range predictions. To address these challenges, this paper proposes a Trajectory Representation Learning (TRL) model based on self-supervised learning. The model employs a two-stage learning strategy: a self-supervised pre-training phase that utilizes an innovative asynchronous masking module to learn intrinsic spatio-temporal dependencies and physical constraints, followed by a fine-tuning phase for specific prediction tasks. By integrating a multi-scale convolutional encoder, the model effectively extracts multi-level trajectory features while suppressing error divergence. The model is validated through experiments on a civil flight trajectory dataset. Experimental results on real-world civil aviation datasets demonstrate that TRL significantly outperforms existing baselines. Furthermore, TRL exhibits superior generalization and robustness in zero-shot and few-shot scenarios, effectively mitigating the risks of over-fitting and error accumulation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Flight Trajectory Representation Learning Model Based on Self-Supervised Learning

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Gaoyong Lu,
  • Xu Si,
  • Qingyue Hu,
  • Yang Ou,
  • Jing Liang

摘要

Flight trajectory prediction is fundamental to modern air traffic management and air combat maneuver decision-making. However, traditional deep learning models under general training paradigms often struggle to capture long-term dependencies when processing extended sequences, leading to imprecise data representations and significant error accumulation in long-range predictions. To address these challenges, this paper proposes a Trajectory Representation Learning (TRL) model based on self-supervised learning. The model employs a two-stage learning strategy: a self-supervised pre-training phase that utilizes an innovative asynchronous masking module to learn intrinsic spatio-temporal dependencies and physical constraints, followed by a fine-tuning phase for specific prediction tasks. By integrating a multi-scale convolutional encoder, the model effectively extracts multi-level trajectory features while suppressing error divergence. The model is validated through experiments on a civil flight trajectory dataset. Experimental results on real-world civil aviation datasets demonstrate that TRL significantly outperforms existing baselines. Furthermore, TRL exhibits superior generalization and robustness in zero-shot and few-shot scenarios, effectively mitigating the risks of over-fitting and error accumulation.