Aming at the problem that the Levenberg-Marquardt(LM) algorithm can not online train RBF network and the problem in RBF network structure design methods, this paper presents an online self-adaptive RBF network structure design method based on the LM algorithm. The ideal of sliding window and online structure optimization are introduced in this algorithm, the introduction of sliding window enables the RBF network to be trained online by the LM algorithm, and makes the RBF network more robust to the changes of the learning parameters and is easy to converge. The online structure optimization can online self-adaptive adjust the structure of RBF network based on the information of training errors and hidden unites to track the time-varying systems, which helps to maintain a compact netowrk and satisfactory generation. Finally, the experiment results show the performance of the proposed algorithm.