Unsupervised anomaly detection in flight data using convolutional variational auto-encoder

Abstract

The modern National Airspace System (NAS) is an extremely safe system and the aviation industry has experienced a steady decrease in fatalities over the years. This is in part due the airlines, manufacturers, FAA, and research institutions all continually working to improve the safety of the operations. However, the current approach for identifying vulnerabilities in NAS operations leverages domain expertise using knowledge about how the system should behave within the expected tolerances to known safety margins. This approach works well when the system has a well-defined operating condition. However, the operations in the NAS can be highly complex with various nuances that render it difficult to assess risk based on pre-defined safety vulnerabilities. Moreover, state-of-the-art machine learning models that are developed for event detection in aerospace data usually rely on supervised learning. However, in many real-world problems, such as flight safety, creating labels for the data requires specialized expertise that is time consuming and therefore largely impractical. To address this challenge, we develop a Convolutional Variational Auto-Encoder (CVAE), an unsupervised deep generative model for anomaly detection in high-dimensional time-series data. Validating on Yahoo's benchmark data as well as a case study of identifying anomalies in commercial flights' take-offs, we show that CVAE outperforms both classic and deep learning-based approaches in precision and recall of detecting anomalies.