Single scaling wavelet frame theory is used to build up an initial multidimensional wavelet network. Then; Recursive Orthogonal Least Squares(ROLS) is introduced to recursively orthogonalize the wavelets instead of batch Orthogonal Least Squares. After orthogonalization; it is feasible to calculate the contribution of each wavelet to the output. Two approaches of selecting wavelet regressors; forward selection method and backward elimination method; are proposed to establish an effective and concise wavelet network. Finally; numerical experience on nonlinear system identification is presented for illustration.