Abstract
This paper investigates the problem of identifying errors-in-variables (EIV) models, where the both input and output measurements are corrupted by white noise, and addresses a new method to solve the problem. The identification problem of EIV models with unknown noise variances has been extensively studied and several methods have been proposed. To be further developed in terms of estimation accuracy, a generalized eigenvector method with no requirement of a priori knowledge about the noise variances is proposed by using the biased weighted least squares estimator. The proposed generalized eigenvalue problem can be derived by removing only the bias induced by the output noise, and thus the system parameter can be obtained as the generalized eigenvector without requiring the use of iterative identification procedure. Moreover, the bias compensation principle based algorithm, which is suitable for on-line implementation, is derived to solve the proposed generalized eigenvalue problem. The results of simulated examples indicate that the proposed approach provides good parameter estimates.
