A fault diagnosis approach for rotor systems based on Local Mean Decomposition (LMD) and AR model was proposed. Firstly, by using LMD method, the vibration signal of rotor systems was decomposed into a number of product function (PF) components, whose instantaneous frequencies had physical meaning, and then the AR model of each PF component was established. Furthermore, the model parameters and the variance of remnant were regarded as the fault feature and served as the input parameter of neural networks to identify the condition and fault pattern of a rotor system. The study results have shown that both EMD and LMD method can be applied to the rotor system fault diagnosis effectively. However, the latter has better decomposition results.