In view of the problem that fault diagnosis for gas turbine vibration generator set parameters is difficult to reflect the state of unit fault directly, a fusion of rough set and neural network for gas turbine generator set vibration fault diagnosis is presented. Rough sets theory is applied in reduction of the original features of the vibration signal characteristic value data to remove unnecessary attributes. An optimized neural network structure which is used to fault diagnosis of gas turbine generator set is established based on rough sets. The experimental results show that the method is effective and provides a new idea for gas turbine generator set vibration fault diagnosis.