Deep Learning techniques (Generative Adversarial Networks, Variational Autoencoders and Adversarial Variational Optimization) are applied to anomaly detection in spacecraft telemetry data as well as to infer, from telemetry data, internal parameter values of the spacecraft thermal and electric models. Two experimental setups are presented: On the one hand, using real telemetry data from the ISS, the thermal conductance and the resistance of a basic model have been estimated. On the other hand, a simulator (based on the HELM solver) has been used to generate synthetic telemetry data of a more realistic model, including individual or multiple variations in the internal parameters, to test the capacity of the Deep Learning models to infer these variations. The introduced combination of physical modeling and data analytics, particularly the inference of physical model parameters from telemetry data, is a promising key element for the design of future Digital Twin software applications.
