Abstract: In order to break the limited condition that the output nonlinear blocks are reversible in existing Hammerstein-Wiener model parameter identification methods， a new parameter identification method of multivariate Hammerstein-Wiener model was proposed based on separable nonlinear least square algorithm. The model was comprised of multiple univariate Hammerstein submodels and one multivariate nonlinear block. First， two parameter sets were identified for output nonlinear block and Hammerstein submodels using Levenberg-Marquardt algorithm under the minimum output error criterion. Second， parameters of input nonlinear block and middle linear block were identified by singular value decomposition (SVD) of tensor product from Hammerstein submodels. Then， the identification convergence was theoretically analyzed. Finally， simulation results showed the effectiveness of the proposed method.

Key words: multivariate, nonlinear model, Hammerstein-Wiener model, separable nonlinear least square, SVD(singular value decomposition), convergence