抄録
In this paper we consider the EM algorithm-based identification method for linear stochastic systems subject to observation outliers, where the observation noise contains large values with low probability. We derive a state space identification method by using the EM algorithm, which is initialized by two subspace identification methods: MOESP [10] and ORT [11]1. By using the median of residuals, outliers are detected and then deleted by a simple scheme in robust statistics. We compute the E- and M-steps in the EM algorithm accordingly, and then compute the corresponding Kalman filter and smoother. We show by numerical examples that the EM algorithm can monotonically improve the initial estimates obtained by subspace identification methods.
