Anomaly Detection in Aircraft Performance Data



Nanduri, Syam Kiran Anvardh

Journal Title

Journal ISSN

Volume Title



Detecting anomalous behavior in aircraft both during and after the flight is very important and it is of high interest for aviation safety agencies as well as airliners to ensure safe, efficient and environmentally clean flight operations. In this thesis, we study the capability of artificial neural networks in learning to identify anomalous behaviors in archived multivariate time series aircraft data and also in online data streams. The data is collected from flying 500 approaches to San Francisco International Airport on Boeing 777-200 ER aircraft using the X-Plane aircraft simulation. We analyzed the performance of Deep Autoencoders, Recurrent Neural Networks with Long Short-Term Memory units and Recurrent Neural Networks with Gate Recurrent Neural Networks units on the archived timeseries data. Once trained, the Recurrent Neural Network based algorithms can be applied to either previously collected flight data for retrospective analysis (offline mode) or they can be deployed during the flight to detect the anomalies in real time and alert the crew members (online mode). The performance of these algorithms is compared against MKAD, a Support Vector Machine based algorithm developed at NASA. These algorithms detected the anomalous flight types which MKAD was able to detect and also other anomalous flight types which MKAD was not able to detect. Experiments were conducted using various parameters combinations for MKAD to see how the resolution of Symbolic Aggregate Approximation encoding and size of alphabet set impact its performance. Similarly, various architectures of autoencoders and recurrent neural networks were designed and their performance was evaluated in terms of precision, recall and F1 score. Experimental results show that recurrent neural networks outperformed all other models in overall performance.



Anomaly detection, Aircraft performance, Recurrent neural networks, Simulation data, Autoencoders, Machine learning