基于自注意力机制与高斯混合变分自编码器的飞行轨迹聚类方法研究  

Study on Flight Trajectory Clustering Method Based on Self-Attention Mechanism and Gaussian Mixture Variational Autoencode

作　　者：张召悦[1,2] 李莎 鲍水达 ZHANG Zhaoyue;LI Sha;BAO Shuida(College of Air Traffic Management,Civil Aviation University of China,Tianjin 300300,China;Institute of Science and Technology Innovation,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学空中交通管理学院,天津300300 [2]中国民航大学科技创新研究院,天津300300

出　　处：《河南科技大学学报(自然科学版)》2025年第1期25-33,M0003,M0004,共11页Journal of Henan University of Science And Technology:Natural Science

基　　金：国家重点研发计划项目(KJZ25420200012);中央高校基本科研项目(3122022105)。

摘　　要：为精确识别飞行轨迹的运行模式,提出了一种基于自注意力机制(SA)与高斯混合变分自编码器(GMVAE)的飞行轨迹聚类方法。SA-GMVAE是一种端到端的深度聚类方法,GMVAE利用变分推断估计每条轨迹的潜在分布,将输入的飞行轨迹数据映射到由多个高斯分布组成的潜在空间,同时依据轨迹分布特征进行聚类。考虑到GMVAE无法兼顾潜在特征的全局关键信息,将自注意力机制嵌入到编码器中,以便于在特征提取时能够捕获全局依赖关系并自动分配权重,突出关键特征,提升轨迹聚类效果。最后,以天津滨海国际机场的进场飞行轨迹数据集为例验证了模型的有效性,实验结果表明:SA-GMVAE相较于K-means、DBSCAN、DTW+HDBSCAN、AE+DP与AE+GMM 5种聚类方法,轮廓系数分别提高了27.6%、20.2%、18.2%、18.6%、15.7%;与未引入自注意力机制的GMVAE聚类模型相比,轮廓系数提高了9.5%,能够更准确地对飞行轨迹进行聚类。In order to accurately identify flight trajectory patterns,a flight trajectory clustering method based on Self-Attention mechanism(SA)and Gaussian Mixture Variational Autoencoder(GMVAE)is proposed.SA-GMVAE is an end-to-end deep clustering method.GMVAE uses variational inference to estimate the potential distribution of each trajectory,maps the input flight trajectory data to the potential space composed of multiple Gaussian distributions,and performs clustering according to the trajectory distribution characteristics.Considering that GMVAE cannot take into account the global key information of potential features,the Self-Attention mechanism is embedded in the encoder to capture global dependencies and automatically assign weights during feature extraction,so as to highlight key features and improve trajectory clustering effect.Finally,the approach flight trajectory data set of Tianjin Binhai International Airport is taken as an example to verify the effectiveness of the model.The experimental results show that:Compared with K-means,DBSCAN,DTW+HDBSCAN,AE+DP and AE+GMM,the contour coefficients of SA-GMVAE are increased by 27.6%,20.2%,18.2%,18.6%and 15.7%,respectively.Compared with the GMVAE clustering model without Self-Attention mechanism,the profile coefficient is increased by 9.5%,which can cluster the flight trajectory more accurately.

关 键 词：飞行轨迹 模式识别 变分自编码器 自注意力机制 

分 类 号：V355[航空宇航科学与技术—人机与环境工程]