In order to solve the coupling and disturbance problems of Model750 control moment gyroscope (CMG), a compound decoupling control method for CMG based on radial basis function (RBF) neural network inverse system and linear extended state observer (LESO) is proposed. An inverse system is constructed by using the nonlinear approximation capability of the neural network and is connected in series with the original system, so that the original system is decoupled into two equivalent pseudo-linear subsystems; The linear extended state observer (LESO) is used to estimate the residual coupling and disturbance terms of the equivalent system to compensate them, and the closed-loop is formed with the proportion differentiation (PD) controller to improve the dynamic control performance of the system. The proposed control method is compared with the PID-RBF inverse control method by simulation, and the results show that the method can effectively decouple the Model750 system, and has better dynamic control performance and robustness.