This paper presents a self-organizing feature map (SOFM) based neural network approach to handle the multiple faults diagnosis of steam turbine-generator sets. The proposed method is trained by using the single fault sample and thus can diagnose the vibration faults of the rotating machinery according to the feature map of the trained network. The proposed approach has been tested on the practical vibration data records of units and compared with the existing methods. The test results confirm that the proposed model possesses superior diagnosis accuracy.