PROCESS PARAMETER IDENTIFICATION: A TOTAL LEAST SQUARE APPROACH

Abstract

The Total Least Square (TLS) approach to the fitting problem [Golub, 1980] is extended to cope with parameter identification of dynamical systems with multiple inputs and outputs. This approach to the identification problem circumvents the underlying assumption in the classical Least Square that all errors (caused by measurement noise etc.) are confined to the ‘observation vector’. The system to be identified is described by polynomial matrices. The proposed algorithm can be formulated recursively for real-time identification. The larger computational burden of TLS identification inhibits its use within time-critical processes. Alternatively, for off-line identification a non-recursive algorithm is presented. The algorithm is based on a Singular Value Decomposition (SVD) so that numerical stability is ensured. In presence of noisy measurements or a low precision computer arithmetic the proposed algorithm performes favourably compared to standard Least Square techniques.