Flight maneuver evaluation method based on similarity metric via triplet network

Accurately evaluating the quality of flight training and providing practical feedback are essential for enhancing the skills of flight students. Flight maneuver evaluation using flight data offers a promising solution to this challenge. However, traditional evaluation methods have consistently faced significant challenges in terms of generalizability, primarily due to the scarcity of labeled flight data. We propose a novel flight maneuver evaluation framework based on a triplet network and utilizing contrastive learning. This framework employs a specially designed Flight Data Feature Extraction Network (FDFEN) and a customized triplet loss function to project flight maneuver data into embedding vectors. Through training, the network learns to place embedding vectors of maneuvers with the same score close together while pushing apart those with different scores. The score of a flight maneuver is then determined through nearest-neighbor analysis between its embedding vector and those of other maneuvers. In our experiments, accuracy (ACC) and mean absolute error (MAE) are used to assess model performance. Additionally, an adjusted AIC, incorporating penalty terms for neural network complexity, is employed to holistically evaluate the quality of deep learning models. Experimental results demonstrate that our method achieves an accuracy (ACC) at least 5.03% higher and reduces the mean absolute error (MAE) by 0.064 compared to baseline methods. Furthermore, it exhibits enhanced robustness across various types of flight maneuvers.