Abstract: In order to address difficulties that traditional extraction methods have in extracting effective features from massive and high-dimensional QAR data which lack sufficient labeled data, a QAR2Vec model was proposed with Transformer as its core. Firstly, QAR data were co-encoded with location and flight phase information as input of the model. Secondly, pre-training task of autoregressive prediction was constructed to learn deep features of massive QAR data in a way of self-supervision. Finally, scale of pre-trained QAR2Vec model was recorded and slightly adjusted on flight state prediction and landing abnormal weather recognition tasks, and performance of QAR2Vec model was evaluated by comparing it with two deep learning algorithms—CNN-LSTM and MTL-LSTM without pre-training steps. The results show that QAR2Vec can extract features from QAR data more effectively, with lower errors and higher accuracy on flight status prediction and landing abnormal weather recognition tasks

Key words: QAR2Vec, feature extraction, Transformer net, auto-regression, pre-training, deep learning