ICA based Noise Subtraction for Linear System Identification in Additive Noisy Output

This paper presents a new paradigm for linear system identification with additive noisy output and finds a powerful noise cancellation method. This method treats the model of system identification as an ICA problem with source signals received by several observed signals so that the estimate of noise can be obtained from the observed signals and then reduced from the noisy output by using an easy subtraction. This method does not rely on the statistic characteristics of the additive noise and can work well under low SNR conditions. Moreover, it settles the two ambiguities of the separated noise that are inherent in ICA by using some special characters of the mixing matrix. Synthetic data are applied to validate the effectiveness of the proposed method, and improved performance is obtained.