Neural Network Prediction Uncertainty for Spacecraft Housekeeping Analysis

Abstract

The booming space industry shows us, that with growing computing power and the large number of satellites, we as humans are not capable any more to keep track of all the data being produced. This does not only include mass produced commercial spacecraft but also tailored scientific missions by now. But as measurements with high data throughput are already sorted by advanced algorithms, the satellites health status described by it’s housekeeping data is still under strict surveillance by humans on a higher level. This shows that there is still room for improvement regarding spacecraft autonomy with methods from the newly emerging field of machine and deep learning. Therefore we created ways and generic methods for analysing specifically satellite housekeeping data. This meant to first understand the data, to regularize and normalize it. And secondly, to identify as well as generate useful features for a further automatic process. With this process, a regression analysis was made to predict future values and also give the variance to increase the credibility of the result. All this has been done with parts of the Rosetta housekeeping data as a case study