Abstract: Many actual systems can be represented by the Hammerstein-Wiener model, where a linear dynamic system is surrounded by two static nonlinearities at its input and output. An improved on-line two stage identification algorithm is proposed to identify the Hammerstein-Wiener model with process noise. Firstly, the bias compensation recursive least squares is adopted to identify the parameter vector containing the product of the original system parameters. The estimation bias is compensated by introducing a correction term in the recursive least squares estimate. Secondly, the singular value decomposition method based on the tensor product approach is adopted to separate each parameter value from the original system. The accuracy of parameter separation is improved by introducing the tensor product of two matrixes to approach the weight coefficient of the weighted least squares. Theoretical analysis and computer simulation validate the effectiveness of the proposed algorithm.