系统仿真学报 ›› 2021, Vol. 33 ›› Issue (9): 2191-2201.doi: 10.16182/j.issn1004731x.joss.20-0369

• 仿真建模理论与方法 • 上一篇    下一篇

基于变分自编码器的飓风轨迹异常检测方法

秦婉亭, 老松杨, 汤俊, 卢聪   

  1. 国防科技大学 系统工程学院,湖南 长沙 410073
  • 收稿日期:2020-06-17 修回日期:2020-08-23 出版日期:2021-09-18 发布日期:2021-09-17
  • 作者简介:秦婉亭(1995-),女,博士生,研究方向为轨迹数据挖掘。E-mail:qinwt01@sina.com
  • 基金资助:
    青年人才托举项目(17JCJQQQ048); 学校科研计划重点项目(ZK18-02-12); 湖湘青年人才(2018RS3079); 湖南省研究生创新项目(CX20190046)

Hurricane Trajectory Outlier Detection Method Based on Variational Auto-encode

Qin Wanting, Lao Songyang, Tang Jun, Lu Cong   

  1. College of Systems Engineering, National University of Defense Technology, Changsha 410073, China
  • Received:2020-06-17 Revised:2020-08-23 Online:2021-09-18 Published:2021-09-17

摘要: 飓风常会引起无法估计的人身和经济损失,飓风轨迹异常检测可以为灾情提供辅助信息或异常告警。从深度学习的角度出发,提出了基于变分自编码器的飓风轨迹异常检测方法(Variational Auto-Encoder Outlier Detection, VAEOD)利用滑动窗口将长度不等的轨迹序列变成等长的子轨迹序列作为变分自编码器的输入,通过变分自编码器训练轨迹重构模型,将重构的轨迹与输入的轨迹通过平行、垂直和角度距离进行对比找出异常的轨迹段。通过真实的飓风数据进行仿真实验发现,VAEOD方法比经典的轨迹异常值检测算法(Trajectory Outlier Detection Algorithm, TRAOD)检测的结果更合理,实用性更强。

关键词: 变分自编码器, 空间距离, 滑动窗口, 异常检测

Abstract: Hurricanes often cause incalculable human and economic losses, and the trajectory outlier detection can provide the auxiliary information or abnormal warning of the disaster. On deep learning, a method of hurricane trajectory outlier detection based on variable auto encoder (VAEOD) is proposed in this paper. The trajectory is divided into equal sequence sub trajectories based on the sliding window as the input of VAE. The trajectory reconstruction model is trained by the VAE. The parallel, vertical and angle distance of reconstructed trajectory and the input trajectory are compared to find out the outlier trajectory segments. The simulation experiment on real hurricane data shows that the VAEOD method is more rational and practical than the classical TRAOD method.

Key words: variational auto-encoder, space distance, sliding window, outlier detection

中图分类号: