基于密度峰值的终端区航迹聚类与异常识别

为有效解决高流量终端区内标准飞行模式、非标准飞行模式和异常飞行模式难以自动分离的问题，采用广泛记录的广播式自动相关监视(ADS-B)数据，构建了基于稳健深度自编码器(RDAE)和快速搜索并寻找密度峰值的聚类(CFSFDP)算法的航迹聚类模型; 使用RDAE降维提取终端区内航迹集的非线性特征，利用多种正则化手段约束内部低维流形，以重建更紧密的航迹并将其作为CFSFDP算法的输入，利用轮廓系数选取不同密度飞行模式的聚类中心，并调节边缘密度参数识别出异常航迹; 选取主成分分析(PCA)结合有噪声的空间密度聚类(DBSCAN)算法、动态时间规整(DTW)结合DBSCAN的2种常用航迹聚类模型作为对比项，分别在广州白云机场1 d的少量数据和45 d的大量数据上进行试验。分析结果表明:DTW与CFSFDP的结合模型在少量数据集上具有最优的航迹聚类性能，轮廓系数比对比项分别提升了62%和28%，且可以自动识别出遵循区域导航标准飞行模式的航班和特定环境下遵循管制偏好的非标准飞行模式的航班，识别异常航迹的精确度也分别提高了57%和10%;大量数据下，提出的RDAE结合CFSFDP模型的聚类性能比经典的PCA结合DBSCAN算法提升了13%，且具备可接受的时间复杂度。由此可见，建立的终端区飞行模式区分模型可为空域级交通流性能评估和航班级航迹预测与优化提供数据提取平台。 

Aircraft trajectory clustering in terminal area and anomaly recognition based on density peak

To effectively solve the problem that it is difficult to automatically separate the standard flight, non-standard flight and anomalous flight patterns in high-traffic terminal areas, an aircraft trajectory clustering model was established based on the robust deep auto-encoder (RDAE) and clustering by fast search and find of density peaks (CFSFDP) using the widely recorded automatic dependent surveillance-broadcast (ADS-B) data. The RDAE was designed to reduce the dimensionality and extract nonlinear features from the aircraft trajectory dataset of terminal areas, while various regularization methods were adopted to constrain the internal low-dimensional manifolds to reconstruct a denser aircraft trajectory, and the aircraft trajectory was input to the CFSFDP algorithm. The silhouette coefficient was used to select the cluster centers for flight patterns with different densities and recognize anomalous trajectories by adjusting the edge density parameter. Two widely used aircraft trajectory clustering models, namely principal component analysis (PCA) combined with density-based spatial clustering of applications with noise (DBSCAN) as well as dynamic time wrapping (DTW) combined with DBSCAN, were taken as comparisons. Experiments were conducted on a small data of 1 d and a large data of 45 d of Guangzhou Baiyun Airport. Analysis results demonstrate that the model combining DTW and CFSFDP provides the best aircraft trajectory clustering performance on the small data, and the silhouette coefficient is 62% and 28% higher than those of comparisons, respectively. The DTW/CFSFDP model can automatically recognize standard flights following the area navigation procedures and non-standard flights that reflect controllers' preferences in specific environments, and the accuracies for identifying anomalous aircraft trajectories also improve by 57% and 10%, respectively. For the large data, the clustering performance of the proposed RDAE/CFSFDP model improves by 13% compared to that of the classical PCA/DBSCAN algorithm. Further, the proposed model exhibits acceptable time complexity. In summary, the established flight pattern discrimination model for terminal areas can provide a data extraction platform for the airspace-level traffic flow performance evaluation and the flight-level aircraft trajectory prediction and optimization. 2 tabs, 10 figs, 31 refs.