In order to solve the problem of many parameters and difficult to set in the active disturbance rejection control (ADRC), a radial basis function (RBF) neural network based on LM algorithm and online optimization of network structure is proposed. Using the idea of sliding window, the online input samples are put into a fixed-length queue, the LM-RBF network is applied to ADRC, the controller parameters are set online, and the permanent magnet synchronous motor is used as the object for simulation analysis in Matlab. The results show that compared with the RBF-based conventional active disturbance rejection controller, the improved LM-RBF enables the controller to have faster response speed and better anti-interference ability, which can effectively improve the stability of the controlled system and meet the performance requirements of the nonlinear time-varying system for the active disturbance controller.