Deep learning based multiple sensors monitoring and abnormal discovery for satellite power system

The power system is a vital sub-system for satellite operated successfully. With test and working environment limitations, the telemetry data from sensors and actuators is the only message to communicate with the ground about the status of satellites. In this term, an efficient and accuracy anomaly detection method for satellite power system could promote a powerful manner for identifying fault and trend that decrease safe margins. However, mostly anomaly detectors have to seriously rely on the prior expert knowledge and a nonlinear dimension reduction on telemetry data as the preliminary to reduce the computation scale and complexity. In this paper, a deep learning-based multiple sensors monitoring and abnormal discovery method for satellite power system is proposed to alleviate the limitations mentioned above. Firstly, an overview of the abnormal discovery method for satellite telemetry data is described. Then, a LSTMs-based prediction model and anomaly detection method for satellite power system are established. The data of multi sensors are monitored in one-time-step prediction model simultaneously, and are detected with an unsupervised method to alleviate the dependency of experts' knowledge. Finally, the experiments are performed with the telemetry data from a simulated satellite power system. With the experiments, the proposed method shows great performance on the anomaly detection in a different type of faults with a high precision rate.