The fault diagnosis problem of rotor system was aimed at, on the basis of synthesizing the advantages of Rough Set (RS) theory, Genetic Algorithm (GA) and Neural Network (NN), a new RS-GA-NN compositive classifier was put forward. In the model, the RS was used to carry out selection of sample features; the NN was used to realize the mapping between features and fault type of sample; the GA was used to optimize the structure of NN model in order to make it to reach the best generalization.The rotor fault experimental rig was used to simulate unbalance, misalignment, rubbing and oil whirling faults, and 127 faults samples are obtained. Finally, the RS-GA-NN compositive classifier of multi-faults recognition was established, and the intelligent fault diagnosis experiment was finished, and a very satisfied result is obtained.