Stochastic subspace identification of linear systems with observation outliers

This paper considers a problem of identifying stochastic linear systems subject to observation outliers, where the observation noise contains large values with a low probability. A stochastic subspace identification method for the problem is developed based on a block LQ decomposition, introducing a weighting matrix to delete outputs which are suspected as outliers. The weighting matrix is generated automatically, and is incorporated in the block LQ decomposition to get improved estimates of the forward innovation representation. A numerical simulation result is included to show effectiveness of the proposed method.