Abstract: Based on the intrinsic physical background of nonlinear system, a system identification model is derived from the inherent systematic characteristics, a priori knowledge and experiences. And then, the GFNN (generalized fuzzy neural network) is put forwand, the GFNN approximation theorem is proved. The structure-self-organization and parameter-self-learning algorithm is proposed, which can automatically and simultaneously deal with the process of the system structure identification and parameter self-learning under predefined precision, so that the novel on-line structure self-organization of GFNN is realized. Simulation shows the nonlinear approximation abilities of GFNN, especially for identification of slow time-varying plant. The GFNN is a successful integrated algorithm of fuzzy logic and neural network.