Data monitoring of spacecraft using mixture probabilistic principal component analysis and hidden Semi-Markov models

Recently there are some researches for anomaly detection of spacecraft assuming that a data distribution of spacecraft is constrained to some modes. Based on this assumption, they learn the distribution as a mixture of local models and efficiently detect an anomaly by monitoring the local data distribution of the each mode. In this paper, because a system of a spacecraft has a periodic mode transition pattern, the authors propose the new anomaly detection system which can detect not only anomalies of outliers but also mode transition patterns by combining two methods, one is Mixture Probabilistic Principal Component Analysis to learn a mixture of local linear models and the other is Hidden Semi-Markov Models to learn mode transition patterns between the modes. Then the experiments were conducted to demonstrate the effectiveness of our new approach.