Artificial neural networks in support of spacecraft thermal behaviour modelling

In this work, we investigate the benefits and drawbacks of using data-driven models such as artificial neural networks (ANN) in support of spacecraft behaviour modelling process. This approach has been applied to the ESA mission CLUSTER to recover the readings of a simulated failed thermal sensor. The virtual sensor can recover it with an average error of 1,68%. ANNs have been also applied to another ESA mission, ROSETTA. In this case, the objective was to forecast the reading of certain key thermal sensors as a function of Sun distance and attitude, obtaining an average error of 5,5/spl deg/C. This paper discusses the results so far gained. The conclusions include an assessment of the proposed technique and guidelines for cases where it could be beneficial.