It is significant to estimate the important process variables when process models are unknown. Therefore, the method of combining unscented Kalman filter(UKF) with neural network is used to solve the state estimation problems for a class of nonlinear systems whose models are unknown. The dynamic neural network is used to model for the nonlinear system, and the state and weights are updated at the same time by using UKF, which can achieve the purposes that the neural network approximate the real model, and the estimated values follow the real values. Two simulation examples are given to verify that the proposed approach gets good effects of estimation, and the greater the proportion of state in the output, the higher the estimation precision.