Learning from Temporal Aviation Data to Detect Anomaly Events

Recently, data analysis of aviation sensors is a hot topic because it is important to aviation safety. Many researchers propose temporal models (RNN, LSTM, etc.) to predict the accuracy performance degradation of sensors, due to the mechanical property, and filter out anomaly data of sensors. However, aerial sensor data is usually long-term data because of intensive sample rates and long runs. The traditional models usually lose the gradients during the training phase for long-sequence data. To reduce this problem, in this paper, Residual-Cascaded LSTM (RC-LSTM) is proposed. It can effectively extract the rich features from large inputs to model large-range independence. Besides, it is found that the volume of anomaly data is much less than that of normal samples. Thus, a distribution-aware shuffling scheme is designed to assist models to learn more reasonable representations for input data. A large number of experiments are performed on the simulated flight sensor dataset and two public datasets (earthquake, Lightning2). The proposed RC-LSTM is better than the most advanced method (RNN, LSTM, etc.), and its accuracy is 80%.