A Physics-informed Transfer Learning Approach for Anomaly Detection of Aerospace CMG with Limited Telemetry Data

Abstract

Aerospace control moment gyro (CMG) is a critical control actuator in on-orbit spacecraft. Detecting its anomalies is of vital importance to ensure the attitude safety of entire vehicle. However, most of existing anomaly detection methods cannot be applied to aerospace CMG due to the data limitation of telemetry signals, such as low frequency, limited sensor types and small quantity. To address this issue, a physics-informed transfer learning-based approach is proposed. Firstly, the relationship between available telemetry signals is established by an artificial neural network (ANN) model, with its structure inspired by physical mechanism of CMG. Then, degradation features are extracted from the ANN model, by which a fine-tuning-based transfer learning strategy is explored to construct the performance index used for anomaly detection. Finally, anomaly threshold based on above performance index is determined by kernel density estimation method. The effectiveness of proposed approach is verified through a real dataset collected from a CMG serviced in orbit.