Weighted principal component analysis for Wiener system identification – Regularization and non-Gaussian excitations

Qinghua Zhang (1) , Vincent Laurain (2) , Jiandong Wang (3)

1 I4S - Statistical Inference for Structural Health Monitoring
2 CRAN - Centre de Recherche en Automatique de Nancy
3 Peking University [Beijing]

Finite impulse response (FIR) Wiener systems driven by Gaussian inputs can be efficiently identified by a well-known correlation-based method, except those involving even static nonlinearities. To overcome this deficiency, another method based on weighted principal component analysis (wPCA) has been recently proposed. Like the correlation-based method, the wPCA is designed to estimate the linear dynamic subsystem of a Wiener system without assuming any parametric form of the nonlinearity. To enlarge the applicability of this method, it is shown in this paper that high order FIR approximation of IIR Wiener systems can be efficiently estimated by controlling the variance of parameter estimates with regularization techniques. The case of non-Gaussian inputs is also studied by means of importance sampling.

Mots clés: block-oriented nonlinear system; Wiener system identification; principal component analysis

Domaines: Automatique

Qinghua Zhang, Vincent Laurain, Jiandong Wang. Weighted principal component analysis for Wiener system identification – Regularization and non-Gaussian excitations. 17th IFAC Symposium on System Identification, SYSID 2015, Oct 2015, Beijing, France. ⟨hal-01232183⟩