Flight data outlier detection by constrained LSTM-autoencoder

Detecting outliers of flight data is an important research field for flight safety. Deep learning methods have achieved remarkable performance in the outlier detection tasks for time series data. The majority of previous deep-learning-based outlier detection methods for flight data focus on either learning descriptive features by matching the distribution of inliers with autoencoder-based models, or learning semantic features by mapping inliers into a hyper-sphere with kernel functions, while the information of the given class samples is insufficiently utilized. To address this issue, in this paper, we propose a novel multi-task-based model that can jointly learn descriptive and semantic features. The proposed model is based on an LSTM autoencoder to reconstruct the inputs, and we design a constraining layer to pull the learned semantic features together. By jointly training two branches of the model, the proposed method can learn to fit the distribution of inputs as well as map inliers into a tight hyper-sphere, thus making outliers and inliers more distinguishable. Experimental results on the real flight dataset demonstrate the effectiveness of the proposed method compared to previous algorithms.