Applications of deep learning neural networks to satellite telemetry monitoring

Abstract

As part of the Automated Telemetry Health Monitoring System (ATHMoS) being developed at GSOC, we performed an investigation into potential applications of artificial neural networks to our existing health monitoring system. In the end, we have created an experimental module which uses several deep learning neural networks to augment our existing data analysis algorithms. The module accomplishes three things; automatic feature extraction, anomaly detection, and telemetry prediction. Automatic feature extraction is used to determine numerical values which represent a time window of a single parameters telemetry data, then, using these abstract numeric values we augment our existing so-called feature vectors used in our ATHMoS anomaly detection system to provide additional information in the anomaly detection. Additionally, we use the same type of neural network to perform anomaly detection on each telemetry parameter in order to take an ensemble machine learning approach to detecting anomalies in telemetry. By combining the results of the neural network with our existing (non-neural network) anomaly detection algorithms, we can provide a more robust classification of whether or not new data should be flagged as anomalous. Lastly, we’ve created a neural network which given the most recent orbit data, can predict the general behaviour of a telemetry parameter over the next four and a half hours. Thus, if the prediction we obtain is off from the usual nominal behaviour of the telemetry parameter, we flag it and label it as a potential future anomaly-thereby performing anomaly prediction. Adding these experimental neural network capabilities to work concomitantly with the already existing anomaly detection modules which ATHMoS is comprised of, has already shown to be beneficial by providing new insights into the data as well as offering a more robust approach to anomaly detection via ensemble machine learning. Here we present an overview of a year long study into applying neural networks to telemetry monitoring and discuss in detail the way in which the three applications described above are being added to the current anomaly detection system at GSOC.