Abstract:

Low efficiency of Quick Access Recorder （QAR） data analysis highlights the importance of feature extraction from QAR data. In response to issue of insufficient focus on temporal trend features in QAR data feature extraction， Piecewise Cubic Hermite Interpolating Polynomial （PCHIP） module and ordering relation algorithm （G1） weight assignment module were integrated to form the interpolation weighting， Piecewise Cubic Hermite Interpolating Polynomial — ordering relation algorithm （PG）， then a Convolutional AutoEncoder （CAE） was combined to construct PG-CAE model， and a flight pitch operation feature extraction method based on QAR data was proposed to support analyses such as flight-level anomaly detection. Firstly， the PCHIP was used to standardize data length. Secondly， the G1 weight assignment module was used to determine the weights according to causal temporal correlation between flight operations and flight attitudes， thereby quantifying temporal importance of flight pitch operation data. Thirdly， the CAE module was employed to extract features from the weighted data. Finally， the model validation was performed on the basis of pitch operation data from 406 flight segments of a certain airline's A319 aircraft. Experimental results indicate that PG-CAE model outperforms CAE model significantly by introducing PCHIP and G1 modules， so as to employ the reconstruction error to measure compliance of individual data samples with original data， which determines model acceptability， and use standard deviation to assess model's capability to extract overall trend features from dataset. Ultimately， CAE5 model with a 5- convolutional-pooling layers is identified as the optimal model structure， demonstrating a reconstruction error of 0.032 84 and standard deviations of （0.162 1， 0.280 5）. Furthermore， by combining K-means algorithm， a comparison of point clustering effect after PG-CAE feature extraction with curve clustering effect without feature extraction further demonstrates that PG-CAE model can extract line cluster data of temporal trend data as point cluster data of two-dimensional features， serving research such as flight-level anomaly detection based on QAR data.

Key words: flight operation, feature extraction, Quick Access Recorder (QAR), Convolutional AutoEncoder (CAE), ordering relation algorithm

摘要：

快速存取记录器（QAR）数据分析效率低导致对QAR数据进行特征提取至关重要。针对QAR数据特征提取对于时序趋势特征关注不足的问题，融合分段三次Hermite插值（PCHIP）模块和序关系分析法（G1）赋权模块形成模型插值赋权部分分段三次Hermite插值-序关系分析法（PG），然后结合卷积自编码器（CAE）构建PG-CAE模型，提出一种基于QAR数据的飞行俯仰操作特征提取方法，旨在为飞行级异常检测等分析提供支持。首先，利用PCHIP统一数据长度；其次，利用G1赋权模块根据飞行操作与飞行姿态的因果时序相关性确定权重，从而量化飞行俯仰操作数据的时序重要性；再次，使用CAE模块对赋权后的数据进行特征提取；最后，基于某航司A319机型406个航段的俯仰操作数据进行模型验证。实验结果表明：通过引入PCHIP与G1模块， PG-CAE模型的结果明显优于CAE模型，从而以重构误差来度量单一数据样本与原始数据的符合度，并将它作为模型是否可接受的底线标准，同时以标准差来度量模型对数据集整体趋势特征的提取能力，最终确定具有5重卷积池化层的CAE5模型为最优模型结构，它的重构误差为0.032 84、标准差为（0.162 1，0.280 5）。此外，结合K-means算法，对比PG-CAE特征提取后的点聚类效果与未经特征提取的曲线聚类效果，进一步证明PG-CAE模型可将时序趋势数据的线簇数据提取为二维特征的点簇数据，从而服务于基于QAR数据飞行级异常检测等研究。

关键词: 飞行操作, 特征提取, 快速存取记录器, 卷积自编码器, 序关系分析法