Fault Diagnosis of Satellites under Variable Conditions based on Domain Adaptive Adversarial Deep Neural Network

Abstract

Fault diagnosis of satellite attitude control system is an important task to ensure the safe and reliable operation of on-orbit satellites. At present, most fault diagnosis methods are to diagnose independent identically distributed(i.i.d) task objects. However, even if the same device works under different working conditions, the distribution domain of the collected data almost always changes. At the same time, the training of fault diagnosis model under full working conditions can increase the model complexity and training time, and there may unknown working conditions. In view of the above situation, this paper proposed a domain adaptive adversarial deep neural network based fault diagnosis method. By combining the feature extractor, label classifier and domain classifier with the convolutional neural network and gradient inversion layer (GRL), the effective label classification can be achieved while the resolution of different domains can be reduced. We achieved feature extraction of the classification learning task in the source domain and transfer of the classification task between the two domains. The effectiveness of the diagnosis model is verified in the ground simulation data of a certain satellite under different conditions.