基于深度时域网络的飞行流数据风险辨识

Risk identification of streamed flight data based on deep temporal network

摘要:

与航后飞行品质监控技术相比，监测飞行实时流数据将有助于实现运行风险的在线辨识和预警。本文提出基于飞行流数据深度学习的风险辨识方法，并以飞机进近阶段典型风险开展实验验证。设计以时域卷积网络与门控循环单元组合的深度时域网络（DTN）结构，基于离线飞行数据集，通过DTN 网络实现长时记忆和局部时间依赖的异常特征提取；将DTN 网络简化为适用于在线风险辨识的DTN-R 网络，开展真实进近段飞行流数据的在线风险辨识实验，结果表明：DTN-R 网络能够准确识别超速、下滑道偏离、襟翼延迟等典型风险，通过不同的风险特征的学习，DTN 网络可以推广应用到其他类型的风险辨识；DTN-R 网络提供了风险置信率得分，为协助分析风险异常的诱因、保障飞行安全提供了有效手段。

Abstract:

Compared with post-flight operations quality assurance, it is helpful to identify and pre-warn operation risks by monitoring streamed flight data. A risks identification method based on streamed flight data deep learning was studied, and the proposed method was further verified in aircraft approaching flight. A deep temporal network (DTN) was constructed by combining temporal convolutional networks (TCN) with gated recurrent units (GRU). Firstly, the DTN was used to extract the long-term memory and local time-dependent features based on the offline flight dataset. Secondly, the DTN was simplified to the DTN-R network which is suitable for online anomaly monitoring. The anomaly detection tests based on real streamed flight data show that the DTN-R is able to identify typical anomalies including high speed exceedance (HSE), glide slope deviation (GSD) and late flap setting (LFS), thus achieving the online risk monitoring. By studying different risk features, the DTN can be extended to identify other in-flight risks. Providing the score of risk confidence, the DTN-R is helpful to explain the causes of abnormal risks. Furthermore, the proposed methodology provides an effective way of assuring flight safety.