PLMR: An Efficient Pre-trained Model for Aircraft Maneuver Recognition

Recognizing aircraft maneuvers is crucial for pilot training and flight safety. Previous works mainly relied on expert knowledge and time window division, which requires a considerable amount of maneuver data and fails to capture the temporal dependencies of the maneuver trajectory. To address these problems, this paper proposes an efficient pre-trained model (PLMR) that incorporates pre-training techniques and deep learning models. PLMR adeptly captures the temporal dependencies of the maneuver trajectory and exhibits superior performance in downstream tasks with minimal data requirements and reduced computational demands. The task was divided into two stages: pre-training and fine-tuning. In the pre-training stage, a self-supervised pre-training task called masked trajectory modeling was introduced. This stage involved randomly masking a portion of the trajectory vectors of the input time series and training a temporal autoencoder to reconstruct the masked trajectory vectors, which enabled better temporal dependencies capture of the input sequence. In the fine-tuning stage, the temporal autoencoder that was pre-trained in the first stage was loaded into downstream tasks for end-to-end weight fine-tuning. The experimental results conducted on eight downstream datasets revealed that PLMR outperformed existing methods with fewer participation numbers and reduced computational resources.