Spacecraft Time-Series Anomaly Detection Using Transfer Learning

Abstract

Anomaly detection in telemetry channels is a high priority for spacecraft, especially when considering the harsh environment of space and the magnitude of launch and operation costs. Traditional spacecraft anomaly detection methods are limited in scope and rely on domain experts to correctly determine abnormal behavior. However, with thousands of distinct telemetry channels being transmitted, the amount of data is difficult to monitor manually. Deep learning models can be used to learn the normal behavior of the telemetry channels and flag or label any deviations. The problem is that we have to train a unique model for each channel to ensure best performance. With the large number of channels to monitor, this may not always be possible. We propose using principles of transfer learning to quickly adapt a general pre-trained model to any specific telemetry channel, greatly reducing the number of unique models needed and reducing the training time for each model. We present the results of our approach on the NASA SMAP/MSL dataset to show that we can achieve performance comparable to state-of-the-art anomaly detection methods.