Satellite battery degradation prognostic based on data analytics and big data infrastructure

摘要

Communication satellites are complex systems remotely operated. As it is impossible to request any measure on the system that was not planned at design phase, it is assembled with hundreds of sensors. All those sensors capture many physical variables at a sampling rate of roughly one point per minute. Representing million points per sensor over years, this is the only information satellite operators have to diagnose the behavior of their satellite. Satellite monitoring is today supported by expert definition of processing. It requires experienced understanding of physical or chemical behind a subsystem, and this knowledge is making use of available parameters. Being the most accurate method to monitor a subsystem, it is hardly applicable to a constellation of hundreds of satellites. Lessons learnt acquired from satellite operators and airlines lead as follows: while similar or even identical product at beginning of life, due to their operating conditions, each element of a fleet or a constellation rapidly behaves differently from the others. While still comparable, they are no longer identical. Expert approach not being scalable to a whole fleet or constellation at reasonable cost and in a reasonable planning, we investigate data analytics approach to support experts. All battery data over all Airbus E3000 platforms represent about 300 years of observation. Benefiting from repetition of operating conditions due to the orbit and sun/shadow alternation, data was split in time windows. It is then demonstrated that Out Of Limits (OOL) monitoring can't capture all variability in data. Hence a functional decomposition is proposed to capture all signals dynamic and concentrate it into a smaller set of coefficients. Benefiting from this reduced dataset, a surrogate modeling approach is then applied to capture the correlation between battery data and battery age. This modeling technique accuracy is evaluated onto a test set and demonstrates good prognostic accuracy. It can be now an additional tool provided to experts to support decision on slowing or accelerating degradation of a given satellite battery. This modeling does not rely on any strong assumption related to subsystem knowledge. It can be applied to other subsystems. But due to the data volume accessible and leveraged through this method, we investigate distributed storage and computation techniques on a Hadoop/Spark cluster. After a benchmark of storage techniques to store our time series, data analytics calculation is performed and measured. We present here final choice of techniques and time performance of overall process to support analytics on sensor data in a big data infrastructure.