Deep Learning Based Fault Diagnosis Methods for Satellite Power System

Due to the harsh space environment, the Satellite Power System (SPS) may be affected greatly. As the heart of satellite, fault in SPS can lead to mission failure and huge economic losses. Therefore, it is necessary to design Fault Diagnosis (FD) algorithm to find and solve the fault as soon as possible. However, since the configuration of different satellites varies widely, it is unrealistic to design a universal FD algorithm based on traditional model-based methods. Data-driven FD is a possible way, for which it can learn and extract features from data without precise knowledge of system. This paper proposes a novel data-driven FD approach for SPS using Gramian angular field (GAF) algorithm and Deep Learning (DL) technique. Specifically, GAF encodes multiple Time Series (TS) related to fault into 2-D images while retains their original information. Then we construct a dataset and use Convolutional Neural Network (CNN) based DL models to capture normal and fault features. Furthermore, Deep Transfer Convolutional Neural Network (DTCNN) is used to improve the accuracy, stability and convergence time of DL models. The accuracy and F1-score of GAF-DTCNN are 97.8% and 0.944 for a geostationary satellite, which illustrates the potential and feasibility of our proposed FD scheme for SPS.