基于图神经网络的航空数据异常检测

Aviation Data Anomaly Detection Based on Graph Neural Networks

飞行品质监控(FOQA)数据记录了飞行状态的详细参数,对于评估飞行操作的质量和安全性至关重要。传统的“超限检测”算法通过与预先建立的阈值进行比较来识别异常行为。相比之下,深度学习方法能够更全面、灵活地分析FOQA数据,提高异常行为的检测精度。文章提出的TAGDNet是用于FOQA数据多类别异常检测的创新框架,包括时序卷积网络、图神经网络和分层图池化等关键组件。该框架首先通过时序卷积网络提取时序特征,然后通过引入图神经网络进行节点间信息传播,最后通过分层图池化获得异常检测结果。通过在公开可用的FOQA数据多类别异常检测数据集上进行大量实验证明,该方法相较于其他先进的方法表现更为优越。

Flight Operational Quality Assurance(FOQA)data records detailed parameters of flight status,which is crucial for evaluating the quality and safety of flight operations.Traditional“Exceedance Detection”algorithm identifies abnormal behavior by comparing it with predefined thresholds.In contrast,Deep Learning methods can comprehensively and flexibly analyze FOQA data,improving the accuracy of abnormal behavior detection.The TAGDNet proposed in the paper is an innovative framework for multi-class abnormal detection in FOQA data,including key components such as Temporal Convolutional Networks,Graph Neural Networks,and Hierarchical Graph Pooling.The framework extracts temporal features through Temporal Convolutional Networks firstly,then propagates information between nodes through introducing Graph Neural Networks and finally obtains abnormal detection results through Hierarchical Graph Pooling.Through extensive experiments on publicly available FOQA multi-class abnormal detection datasets,it has been demonstrated that this method outperforms other state-of-the-art methods.