Abstract: A new non-linear system identification approach is proposed, which combines wavelet theory and NARX(non-linear auto-regressive with exogeneous inputs) model properly. The approach utilizes efficient approximation power of wavelet (multi-dimensional wavelet) function to remove the complicated process of model structure determinetion. It constructs a rather general framework of identification without depending on a priori information of the system. The Recursive Least Square (RLS) algorithm can be used to estimate the parameters of the new identification model, which is feasible to realize on-line identification. The results of two simulation examples illustrate effectiveness of the new identification approach.