Abstract: QAR(quick access recorder) data retains a large number of flight parameters of the aircraft, which makes it possible to study the flight safety of the aircraft and ensure the flight quality. Aiming at the requirement of multi-dimensional QAR data security detection, an effective and interpretable multi-dimensional time series anomaly detection model based on convolution VAE and multi-head self-attention-LSTM is proposed. The new model is able to capture both spatial and temporal information between multiple time series data channels, and perform interpretable reconstruction of sequential patterns to detect all types of anomalies. In the experiment, compared with the existing methods such as iForest, LSTM, VAE, LSTM-VAE, and USAD on real QAR datasets, the results show that the F1-scores of this model in three different stages of anomaly detection reach 0.891 2, 0.942 4, and 0.953 7, respectively, which are superior to other models and can accurately detect anomalies in multidimensional QAR data.

Key words: anomaly detection, QAR data, multivariate time series, neural networks