Abstract: Introducing the functional-like network theory, nonlinear filter based on neural networks(NNNF) for state estimation of stochastic nonlinear systems is proposed by the using of data of on-line measurements or available off-line measurements data. The performance of biaslessness and minimal error variances of NNNF are proven. NNNF is applied to the state estimations of glutamic acid fermentation and erythromycin fermentation processes disturbed by noise, and the estimation by experimental errors. The estimated results and experimental online state estimation coincide very well. NNNF is insensitive to noise distributions and initial state estimation, NNNF can be used in on-line measurements of biomass, substrate, and product concentrations.