Abstract: A soft sensor modeling algorithm based on improved fuzzy neural network is presented in this paper. The normalized average output membership functions are defined as fuzzy basis functions for defuzzification calculation. In order to improve the convergence, some parameters are trained by Levenberg-Marquardt algorithm, and the others are trained by steepest descent method. Finally, a soft sensor model of melt index in polymer reaction based on the proposed method is presented, and the simulation results show that, in contrast to the traditional fuzzy neural network, the proposed method which is not sensitive to initial parameters and has good convergence performance and prediction precision, is suitable to practical applications.