基于VAE-LSTM模型的无人机飞行数据异常检测

无人机飞行数据是反映其自身飞行安全的重要状态参数，通过对飞行数据进行异常检测，是提高无人机整体飞行安全性的关键举措。尽管基于数据驱动方法不需专家先验知识和精确的物理模型，但缺乏参数选择且检测网络结构模型单一，使得检测模型由于参数过多导致过拟合以及无法有效捕捉数据异常模式的问题。文中结合变分自编码器和长短期记忆网络的优势，提出了一种基于VAE-LSTM的无人机飞行数据异常检测模型方法。首先，引入肯德尔相关性分析方法用于选择相关依赖的飞行数据参数集；其次，将具有相关性的参数集对所设计的VAE-LSTM深度混合模型进行训练，学习不同数据特征之间的关系映射；最后，以无监督异常检测方式在真实多维无人机飞行数据进行验证。实验结果表明，VAE-LSTM的精密度、检测率、准确率、F1分数及误检率的各项平均性能指标分别达到95.24%、98.71%、98.8%、 96.82%、 1.31%,相比于KNN、OC-SVM、VAE、LSTM模型，整体上展现出较好异常检测性能。

Anomaly detection of UAV flight data based on VAE-LSTM modeling

UAV flight data is an important state parameter reflecting its own flight safety, and it is a key initiative to improve the overall flight safety of UAVs through abnormal detection of flight data. Although data-driven methods do not require expert a priori knowledge and accurate physical models, the lack of parameter selection and a single model for the detection network structure make the detection model overfitting due to too many parameters and failing to effectively capture data anomaly patterns. In this paper, a VAE-LSTM based UAV flight data anomaly detection modeling method is proposed by combining the advantages of Variational Auto-Encoders and Long Short-Term Memory networks. First, the Kendall correlation analysis method is introduced for selecting relevant dependent flight data parameter sets; Second, the parameter sets with correlation are trained on the designed VAE-LSTM deep hybrid model to learn the relational mapping between different data features; And lastly, the validation is performed with unsupervised anomaly detection in real multi-dimensional Unmanned Aerial Vehicle flight data. The experimental results show that the various average performance metrics of precision, detection rate, accuracy, F1 score and false detection rate of VAE-LSTM reach 95.24%, 98.71%, 98.8%, 96.82%, and 1.31%, respectively, and show overall better anomaly detection performance compared to KNN, OC-SVM, VAE, and LSTM models.