A Transformer-Based Deep Learning Approach to Anomaly Detection of High-Bandwidth Multivariate Time-Series Satellite Communications

Monitoring incoming spacecraft data to detect anomalies is essential for the continuation of a successful mission. Spacecraft Operations Engineers have traditionally relied on manual monitoring of satellite data, with limited statistical support for anomaly detection, focusing primarily on threshold-based monitoring of individual data items to identify and resolve any anomalous behavior; however, as the number of satellites launched increases exponentially every year with growing bandwidth per new satellite, the amount of data Operations Engineers need to monitor is increasing at an alarming rate. There is a pressing concern of monitoring all the incoming data without increasing operating costs. In this paper, we demonstrate that modern deep-learning anomaly detection methods can aid existing Spacecraft Operations Engineers. Previous research has explored the use of machine learning to monitor incoming data and alert Operations Engineers about potential anomalies. We build upon the literature by taking a multivariate time-series dataset (DSN spacecraft monitor data) and feeding it into a transformer-based deep learning algorithm. Whereas previous work uses complete data with a single data type, this dataset proves challenging due to various data types per time-step and missing data. We refine the data preprocessing and architecture of the TranAD algorithm to improve its performance on both the DSN dataset and existing public anomaly detection datasets. We demonstrate that the modified TranAD algorithm improves performance over MTAD-GAT, OmniAnomaly, and USAD on the DSN dataset. We also tested some public datasets, including SMAP, MSL, SWaT, SMD, and WADI, and demonstrated improved performance on SMAP, SWaT, and SMD. The improved performance promises the ability for Spacecraft Operations Engineers to instantaneously detect anomalies in the high-bandwidth regime of modern satellite communications.