A multidimensional feature fusion model based on a long short-term memory (LSTM) autoencoder was proposed to address the limitations of traditional trajectory similarity measurement methods, which suffered from low computational efficiency and insufficient classification accuracy due to considering only inter-trajectory distance. Positional features (longitude, latitude, altitude), shape features (heading, curvature), and motion features (speed, acceleration) were extracted from the original flight trajectories and fused using min-max normalization. The high-dimensional trajectory data were encoded into a two-dimensional embedding space, where similarity was measured using Euclidean distance. Experiments were conducted on a dataset of 1 291 arrival ADS-B trajectories at Shanghai Pudong International Airport, comparing traditional methods and deep learning approaches (Transformer, graph convolutional network). The results demonstrate that the proposed model significantly outperforms the comparative methods in trajectory classification accuracy (94.7%) and Davies-Bouldin index (DBI) index (0.33), while reducing computational time from 30 minutes to 10 minutes. By integrating multidimensional dynamic features with an efficient dimensionality reduction mechanism, the model provides a novel solution for real-time trajectory analysis and anomaly detection in air traffic management systems.