Abstract: Satellite power system usually has a variety of working modes, and the observation data in different working modes have different statistical characteristics. Due to the fact that the actual observation data of satellite power system cannot provide the state identification priori information, the traditional anomaly detection methods cannot distinguish the different working modes of the satellite power system. Therefore, the traditional methods have great limitations. In order to solve the problem of anomaly detection without state identification, a data driven anomaly detection method is proposed for satellite power system based on Gaussian mixture model (GMM). As a data clustering method, GMM can mine the intrinsic characteristics of data in the lack of working state identification, and realize the clustering and recognition of multiple working modes. Then, indexes are given to evaluate the GMM method from three aspects: distinguishability, stability and information. These criteria can ensure that the cluster number is reasonable. Furthermore, in the anomaly detection stage, the trained GMM is used to construct the pattern recognition criteria. The distance information and F distribution are used to construct the detection threshold. And the detection effect is improved by increasing the window length of the testing data. Finally, numerical simulation and experimental verification are carried out for the solar array drive assembly (SADA) of satellite power system. The results of anomaly detection show that the proposed method can effectively realize anomaly detection in a variety of working modes, and has high precision and recall rate.

Key words: anomaly detection, satellite power system, gaussian mixture model, EM algorithm, lack of state identification, data driven