Mitigating Data Scarcity for Satellite Reaction Wheel Fault Diagnosis with Wasserstein Generative Adversarial Networks

Data scarcity is an issue that has been a deterrent for researchers looking to employ data-driven fault diagnosis and prognosis methods in satellites and spacecraft. With the ongoing boom in artificial intelligence and machine learning, it has become difficult not to notice this field’s significance and potential in the fault diagnosis and prognosis area. Moreover, the surge in the use of small satellites has contributed to the need to devise effective and efficient health monitoring methods. However, not much operational data on satellites’ subsystems is usually available when employing diagnostic data-driven methods. In this study, we propose an approach based on the Wasserstein Generative Adversarial Network (WGAN)s and Long Short-Term Memory (LSTM) networks to 1) mitigate the data scarcity issue by generating diverse datasets from whatever scarce data that is available and 2) perform fault detection and identification on a single satellite reaction wheel. The proposed dual-one-dimensional-WGAN-LSTM model is tested using reaction wheel time-series data and demonstrated successful performance in diagnosing faults.