In order to improve the disturbance rejection property of Permanent-Magnet Synchronous Motor servo systems, a novel adaptive composite control method for the speed regulation problem is proposed in this paper. The composite controller is composed of a proportional feedback controller and a feedforward compensation based on a Radial Basis Function Network Disturbance Observer (P+RBFNDOB). The RBFNDOB is designed to estimate the lumped disturbances, including external disturbances and internal disturbances caused by parameter variations, and the estimation value is used for feedforward compensation design. Different from traditional Linear DOB (LDOB), an RBFN is used to approximate the inverse model of the system instead of selecting the inverse of the nominal model as in LDOB. By using an on-line learning algorithm, the identified inverse model can track the variations of the real plant. Thus, the RBFNDOB can still observe the disturbances when the parameters of the system vary in a wide range, while conventional DOB may not be suitable in such situations. So the composite controller obtained is inherently robust against parameter variations and external disturbances. A nearest neighbour clustering algorithm combining crude regulation and fine regulation is introduced as the on-line learning method to simplify the network structure and accelerate the learning speed. Rigorous analysis is also given to show why the RBFNDOB can effectively suppress the lumped disturbances of a closed-loop system. Simulation comparisons with two other methods, the composite control method with proportional feedback plus feedforward compensation based on LDOB (P+LDOB) and the conventional PI control method, verify the effectiveness of the proposed method.