A new method of constructing fuzzy neural network is presented and Rough set theory is applied to this method. Since Rough set theory has strong numeric analyzing ability and fuzzy neural network has exact function approaching ability, their combination can produce a neural network model with good intelligibility and fast convergence. First, some rules are acquired from given data set by rough set theory. Then, these rules are applied to constructing neural cell numbers and relative parameters in fuzzy neural network. Finally the initial network is trained by BP arithmetic and the whole network design is finished. Also in this paper, an example of nonlinear function approaching is discussed and the feasibility of this method is proved.