基于图的生成对抗网络无人机数据异常检测  

Anomaly Detection of UAV Data in Graph based Generative Adversarial Networks

作　　者：徐嘉闻 周航[1] 汪玥 徐泽楷 XU Jia-wen;ZHOU Hang;WANG Yue;XU Ze-kai(Nanjing University of Aeronautics and Astronautics,Nanjing 210000,China)

机构地区：[1]南京航空航天大学,江苏南京210000

出　　处：《航空计算技术》2024年第5期84-88,93,共6页Aeronautical Computing Technique

基　　金：国家自然科学基金委员会中国民用航空局民航联合研究基金重点项目资助(U2233208)。

摘　　要：无人机异常状态检测是保障无人机飞行安全的重要途经之一,其中基于无人机飞行数据的异常检测是最常用的手段。因此,提出了一种基于图的生成对抗网络的异常检测算法(TGAN GAT)。以TCN为生成对抗网络的基础网络,解决了传统循环神经网络不能并行计算的问题,提高了训练速度,并引入图注意力机制以清晰地捕捉不同时间序列之间的关系。异常检测方法则依据由于重构误差与鉴别误差构成的异常分数。实验表明:此异常检测算法相对于MTAD GAT等算法,在召回率与F 1分数上相对于第二名分别提升了10.68%和7.02%。UAV anomaly state detection is one of the important ways to ensure the safety of UAV flight,among which anomaly detection based on UAV flight data is the most commonly used means.Therefore,this paper proposes a graph based generative adversarial network for anomaly detection algorithm(TGAN GAT).In this paper,TCN is used as the base network of generative adversarial network,which solves the problem that traditional recurrent neural networks cannot be computed in parallel and improves the training speed.And the graph attention mechanism is introduced to clearly capture the relationship between different time series.The anomaly detection method is then based on the anomaly score due to the reconstruction error and the discrimination error.Experiments show that this anomaly detection algorithm improves the recall and F 1 score by 10.68%and 7.02%relative to MTAD GAT compared to the second place.

关 键 词：多元时间序列 生成对抗网络 异常检测 图注意力机制 无监督学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] V328[自动化与计算机技术—控制科学与工程]