关键词: 航迹相似性, 航迹分类, 随机深度网络, 异常检测, 民用航空

Abstract:

In traditional trajectory similarity research， clustering algorithms have the problems of indistinct trajectory identification results， imprecise clustering， and larger time cost under the condition of high-dimensional data， especially under the condition of complex trajectory. To address these issues， we present a trajectory classification and anomaly detection model based on the improved stochastic depth network. Firstly， we optimize the attention mechanism Squeezed-and-Excitation （SE） module and Global Average Pool （GAP） module based on the ResNet model， and construct a trajectory classification network model. Secondly， in the data processing stage， the continuous spatiotemporal trajectory model is used to convert the discrete trajectory data into continuous data of the trajectory function of time for graph neural network processing. Then， nominal trajectory data is introduced into the training set to realize trajectory classification， with nominal trajectories as the reference. Finally， an improved twin neural network is developed based on the trajectory classification results， and is utilized for abnormal trajectory detection. Comprehensive experiments show that the proposed algorithm can efficiently complete the task of track classification according to nominal trajectories and detect abnormal trajectories accurately， compared with traditional track clustering algorithms.

Key words: trajectory similarity, trajectory classification, stochastic depth ResNet, anomaly detection, civil aviation