航空学报 > 2026, Vol. 47 Issue (1): 332106-332106   doi: 10.7527/S1000-6893.2025.32106

融合表征转换与模式回归的航迹插补方法

陶冶(), 汤锦辉, 闫震, 周臣, 王冲   

  1. 空中交通管理系统全国重点实验室,北京 100083
  • 收稿日期:2025-04-11 修回日期:2025-05-15 接受日期:2025-07-24 出版日期:2025-09-11 发布日期:2025-08-28
  • 通讯作者: 陶冶 E-mail:taoyedlmu@163.com
  • 基金资助:
    青年创新基金(P05-24-02-032)

A trajectory imputation method integrating representation transformation and pattern regression

Ye TAO(), Jinhui TANG, Zhen YAN, Chen ZHOU, Chong WANG   

  1. State Key Laboratory of Air Traffic Management System,Beijing 100083,China
  • Received:2025-04-11 Revised:2025-05-15 Accepted:2025-07-24 Online:2025-09-11 Published:2025-08-28
  • Contact: Ye TAO E-mail:taoyedlmu@163.com
  • Supported by:
    Youth Innovation Fun(P05-24-02-032)

摘要:

空中交通管制监视系统采集的民航航迹数据存在完整性缺失的问题,不仅影响了空管监视的时空连续性,更制约了基于航迹运行框架下智能化技术的应用潜力。针对这一问题,提出一种融合表征转换与模式回归的航迹插补方法。本方法的技术实现路径包含3个核心构件:首先,构建基于时空四维航迹数据向RGB通道彩色图像表征形式转换的编码模型,通过将缺失航迹插补问题转换为图像修复任务,实现问题域的创新性转换;其次,设计具备多尺度感知能力的回归模型架构,并在各分辨率信息传递路径中分别引入多卷积融合模块与改进型混合注意力机制模块,有效提升复杂时空模式的表征能力;最终,采用提出的基于自适应权重分配的全局与局部特征融合训练策略,系统性地优化模型整体拟合性能。实验环节在真实航迹数据集上构建多维评估体系,结果表明:相较于传统统计插值方法与现有深度学习方法,提出方法针对90%缺失率以下的航迹数据插补精度有明显提升,同时消融实验证实各创新模块与训练策略的有效性。上述结果充分说明提出方法在解决航迹完整性不足问题时,展现出泛化性能优势与工程应用鲁棒性。

关键词: 空中交通管制, 深度学习, 飞行航迹插补, 表征转换, 模式回归, 多尺度回归模型

Abstract:

The civil aviation trajectory data collected through air traffic control surveillance systems often suffer from integrity loss, which compromises spatial-temporal continuity and limits the application potential of trajectory-based intelligent operational frameworks. To address this challenge, a novel trajectory imputation method is proposed, integrating representation transformation with pattern regression mechanisms. The technical implementation comprises three core components: first, a coding model is developed to transform four-dimensional spatial-temporal trajectories into RGB-channel color image representations. This approach reformulates the missing trajectory completion problem as an image inpainting task, thereby achieving an innovative domain transformation. Second, a multi-scale perception regression model architecture is designed, incorporating multi-kernel fusion modules and improved hybrid attention mechanisms across resolution-specific information pathways. Within information transmission path of each resolution, multi-kernel fusion modules and hybrid attention mechanisms are integrated to strengthen the model’s ability to represent complex spatial-temporal patterns. Finally, an adaptive weight allocation training strategy is employed to systematically optimize the fitting performance of the model. During experimental validation, a multi-dimensional evaluation framework is implemented using a real-world trajectory dataset. Results demonstrate that, compared to traditional statistical interpolation and existing deep learning methods, the proposed method achieves superior imputation accuracy for trajectories with missing rates below 90%. Ablation study further confirms the significant contributions of each module and training strategy to overall performance. These findings underscore the robust generalization capabilities and engineering applicability of method in addressing trajectory integrity challenges.

Key words: air traffic control, deep learning, flight trajectory imputation, representation transformation, pattern regression, symmetric multi-scale regression model

中图分类号: