交通运输系统工程与信息

交通运输系统工程与信息 ›› 2025, Vol. 25 ›› Issue (6): 360-372.DOI: 10.16097/j.cnki.1009-6744.2025.06.033

• 工程应用与案例分析 • 上一篇    

新闻事件驱动的洋区不明飞行活动架次预测

孟令航*1 ，张林海1 ，陈敏2   

  1. 1. 中国民航大学，空中交通管理学院,天津300300；2.民航海南空管分局，海南海口571126
  • 收稿日期:2025-08-07 修回日期:2025-09-28 接受日期:2025-10-16 出版日期:2025-12-25 发布日期:2025-12-24
  • 作者简介:孟令航(1977—)，男，河南桐柏人，副教授，博士。
  • 基金资助:
    国家重点研发计划项目(2022YFB4300904)。

News Event-driven Prediction of Unidentified Flight Activities in Oceanic Areas

MENG Linghang*1, ZHANG Linhai1, CHEN Min2   

  1. 1. College of Air Traffic Management, Civil Aviation University of China, Tianjin 300300, China; 2. Hainan Sub-bureau of Middle South Regional ATMB of CAAC, Haikou 571126, Hainan, China
  • Received:2025-08-07 Revised:2025-09-28 Accepted:2025-10-16 Online:2025-12-25 Published:2025-12-24
  • Supported by:
    National Key Research and Development Program of China (2022YFB4300904)。

PDF

摘要： 近年来，随着国际地缘政治形势的愈发复杂，洋区不明飞行活动日趋频繁，严重影响域内民航航班运行的安全和效率。本文构建一种基于注意力机制的卷积记忆网络预测架构，旨在挖掘GDELT(Global Database of Events, Language, and Tone)新闻事件与不明飞行活动架次及其滞后性的关联特征。首先，采用格兰杰因果检验和相关性分析筛选与飞行活动显著相关的新闻事件类型，并构建输入特征空间；接着，提出一种“卷积神经网络-长短期记忆网络-多头注意机制(CNN-LSTM-MHA)”混合架构，通过CNN提取事件数据与架次数据的局部时空关联特征，通过LSTM捕捉事件滞后影响，并引入多头注意力机制优化学习权重。利用2015—2024年三亚情报区不明飞行活动数据验证模型，结果表明：该预测模型在测试集上表现出较优的性能，平均绝对误差(MAE)、均方根误差(RMSE)和决定系数(R²)分别为0.6049、0.7642和0.8103；模型对正常与异常飞行活动样本均能保持较高预测精度，且测试性能与训练集接近，显示出良好的泛化能力与预测稳定性。

关键词: 航空运输, 不明飞行活动预测, 深度学习, 飞行活动架次, GDELT

Abstract: In recent years, with the increasingly complex of international geopolitics, the unidentified flight activities (UFAs) in ocean areas have a rising frequency. These activities severely impact the safety and operational efficiency of civil aviation flights flying over ocean areas. This study constructs an attention-based convolutional memory network prediction architecture to explore the relationship between GDELT (Global Database of Events, Language, and Tone) news events, the counts of UFAs, and their temporal lag effects. Initially, the Granger causality test and correlation analysis are employed to identify news event types that are significantly correlated with UFA counts, and to construct the input feature space of the proposed hybrid deep learning network. Subsequently, a hybrid deep learning framework combining with the Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), and Multi-Head Attention (MHA) mechanism is developed. The CNN-LSTM-MHA framework utilizes the CNN to extract local spatio-temporal correlation features between GDELT events and UFA counts, explores the LSTM to capture the lagged influence of GDELT events on UFA counts, and incorporates the MHA to optimize learning weights of LSTM outputs and UFA counts. Finally, the proposed framework was validated by using the UFA data in Sanya Flight Information Region (FIR) from 2015 to 2024. The results indicate that: The predictive model achieved excellent performance on the test set, with a mean absolute error (MAE) of 0.6049, root mean square error (RMSE) of 0.7642, and coefficient of determination (R²) of 0.8103; It maintained high predictive accuracy for both normal and anomalous samples, while closely matching training set performance, demonstrating strong generalization capability and reliable prediction stability.

Key words: air transportation, prediction of unidentified flight activities, deep learning, flight activity sortie count, Global Database of Events, Language, and Tone (GDELT)

中图分类号: