基于自编码器和HMM的民机接地点远事件检测

已有的飞行品质监控（FOQA）标准仅由地速的积分距离判定接地点远事件（LTE）的发生，无法结合多个快速存取记录器（QAR）参数检测并解释发生该事件的原因。通过滑动窗口对包含多个参数的QAR样本进行分段，按照分段位置形成若干分段样本集。由长短时记忆网络（LSTM）自编码器得到QAR样本分段及分段内向量的特征表示，采用K-means分别对这些表示向量进行聚类处理，实现QAR样本及其分段的符号化。使用正常航班QAR样本集的符号序列建立隐马尔可夫模型（HMM），以检测包含接地点远事件的航班。由发生及未发生接地点远事件的QAR样本各个分段的符号序列构建HMM后，采用Viterbi算法确定接地点远事件在QAR样本分段内的具体位置。在真实QAR数据集上的实验结果表明，与其他多维时间序列异常检测方法相比，所提方法不仅能有效检测接地点远事件，而且可以获得多个QAR参数的异常值，辅助领域专家分析发生该事件的原因。 

Civil aircraft long touchdown exceedance detection based on autoencoder and HMM

The existing flight operation quality assurance (FOQA) standard only uses the integral distance of the ground speed to define the long touchdown exceedance (LTE), which cannot detect and explain the exceedance using multiple quick access recorder (QAR) parameters. The QAR samples with multiple parameters were segmented by the sliding window, and the segmented sample sets were generated according to the segmentation position. The representation of the QAR sample segmentation and the vector within each QAR segment was obtained by the long short-term memory (LSTM) networks autoencoder, and the representation vectors were clustered by K-means to realize the symbolization of the QAR samples and the QAR segments. The hidden Markov model (HMM) model was built by using the symbolic sequence of the QAR sample set of the normal flights, which was used to detect the flights with the LTE. The second HMM model was constructed from the symbolic sequences of the segment of the normal QAR samples and the QAR samples including the LTE. Then the Viterbi algorithm was used to determine the specific positon of the LTE in the QAR sample segment. Experimental results on real QAR data sets show that, compared with other multi-dimensional time series anomaly detection methods, the proposed method can not only effectively detect the LTE, but also obtain the outliers of multiple QAR parameters, which can assist domain experts to analyze the cause of the exceedance.